Assessing Phylogenetic Network Accuracy: A Comprehensive Guide for Introgression Characterization in Biomedical Research

Abstract

This article provides a comprehensive examination of the accuracy and application of phylogenetic networks for characterizing introgression in evolutionary genomics. As genomic datasets expand across diverse taxa, accurately distinguishing true introgression from other sources of genealogical discordance like incomplete lineage sorting has become crucial for biomedical and drug development research. We explore foundational concepts, current methodological approaches including summary statistics, probabilistic modeling, and machine learning techniques, while addressing significant scalability challenges and optimization strategies. The review synthesizes validation frameworks and comparative performance analyses, offering researchers practical guidance for selecting appropriate methods and interpreting results with confidence in studies of disease evolution, host-pathogen interactions, and comparative genomics.

Understanding Gene Tree Discordance: The Biological Basis for Phylogenetic Networks

In the field of phylogenomics, the genomic landscapes of closely related species are often characterized by conflicting genealogical histories across different loci. Two major processes responsible for these incongruences are introgression, the transfer of genetic material between species through hybridization, and incomplete lineage sorting (ILS), the failure of ancestral genetic polymorphisms to coalesce (reach a common ancestor) within the time span of successive speciation events [1] [2]. Distinguishing between these processes remains a critical task in evolutionary studies, as both result in discordance between gene trees and the species tree [1]. The accurate characterization of these processes is fundamental to constructing reliable phylogenetic networks and understanding the forces that shape genomic evolution. This guide provides a structured comparison of these two phenomena, summarizing key diagnostic features, experimental methods, and analytical tools used by researchers to disentangle their complex signals.

Conceptual and Historical Framework

Introgression, a form of reticulate evolution, requires the successful hybridization between species followed by backcrossing, leading to the incorporation of alien alleles into a new gene pool. This process creates a non-bifurcating relationship among species and can introduce adaptively important variants [1] [3]. Incomplete lineage sorting, by contrast, is a canonical feature of the multispecies coalescent model. ILS occurs when the time between successive speciation events is sufficiently short that genetic lineages from an ancestral population do not have enough time to coalesce, causing some ancestral polymorphisms to persist and be sorted randomly into descendant lineages [2]. This is particularly common during rapid radiations, where short internodal branches increase the probability of ILS [2].

The following table summarizes the core conceptual differences between these two processes.

Table 1: Key Conceptual Differences Between Introgression and Incomplete Lineage Sorting

Feature	Introgression	Incomplete Lineage Sorting (ILS)
Underlying Process	Hybridization and gene flow between species [1].	Random sorting of ancestral polymorphisms due to short internode times [2].
Evolutionary Relationship	Creates non-bifurcating, reticulate relationships [4].	Occurs within a bifurcating species tree model.
Typical Genomic Signature	Localized, "island-like" patterns of elevated similarity between specific species [5].	Genome-wide, stochastic discordance across loci [2].
Dependence on Gene Flow	Requires interspecific gene flow.	Can occur in the complete absence of gene flow.
Impact on Phylogenetic Inference	Can mislead species tree inference if unrecognized, even with low levels of gene flow under high ILS [4].	Causes difficulties in phylogenetic reconstruction, but methods exist to account for it (e.g., coalescent-based species tree inference) [2].

Diagnostic Signatures and Analytical Methods

The discrimination between introgression and ILS relies on detecting their distinct genomic footprints. A signature characteristic of introgression is the asymmetric distribution of sequence similarity. Introgressed regions display exceptionally high similarity between the specific donor and recipient species, a signal that is localized and can be detected using statistics sensitive to recent coalescence events [5]. In contrast, the discordance caused by ILS is typically more symmetric and stochastic across the genome, lacking a consistent directional signal toward one sister species [2].

Powerful methods have been developed to detect the signature of introgression. These include summary statistics such as dXY (the average number of sequence differences between two species), dmin (the minimum sequence distance between any pair of haplotypes from two taxa), and related metrics like Gmin (dmin/dXY) and RNDmin, which are normalized to be robust to variation in mutation rates [5]. The ABBA-BABA test (and related D-statistic) is another widely used method that leverages a four-taxon structure to test for asymmetrical patterns of allele sharing indicative of introgression [2] [3].

For a more model-based approach, coalescent genealogy samplers provide a statistical framework to estimate parameters such as population sizes, divergence times, and migration rates, allowing for a direct test of the introgression hypothesis [1]. Furthermore, supervised machine learning is an emerging and powerful complement to traditional phylogenetic methods. Models can be trained on features derived from phylogenomic datasets to accurately classify whether the underlying history is best explained by speciation with ILS or by introgression [4]. An even more advanced approach uses convolutional neural networks (CNNs) to learn complex patterns directly from genotype matrices, achieving high precision in identifying regions of adaptive introgression [3].

Table 2: Key Methods for Disentangling Introgression and ILS

Method Category	Examples	Key Function	Key Advantages
Summary Statistics	dmin, Gmin, RNDmin [5], D-statistic [2]	Detect elevated genetic similarity between specific species.	Simple, fast to compute; Gmin/RNDmin are robust to mutation rate variation [5].
Population Genetic Inference	Coalescent-based samplers [1]	Jointly estimate population history parameters (divergence time, migration rate).	Model-based; can provide direct estimates of gene flow.
Phylogenetic Networks	Tree-child networks (e.g., inferred by ALTS) [6], PhyloNet [4]	Reconstruct evolutionary histories that include reticulate events (hybridization/introgression).	Explicitly models non-tree-like evolution.
Machine Learning	Supervised learning [4], Convolutional Neural Networks (CNNs) [3]	Classify genomic regions based on patterns from simulated data.	High accuracy; can integrate complex, multi-feature information without a defined analytical model [4] [3].

The diagram below illustrates the logical workflow for distinguishing between introgression and ILS in a genomic analysis pipeline.

Experimental Protocols and Workflows

Protocol: Detecting Introgression using theRNDminStatistic

The RNDmin statistic is a robust method for detecting introgressed regions between sister species, designed to be insensitive to variation in mutation rates [5]. The following workflow details its application:

• Data Collection and Preparation: Obtain phased haplotype data from population genomic samples of two sister species (Populations X and Y) and an outgroup species (O). The outgroup should not have experienced introgression with the target species [5].
• Calculate Sequence Distances: For a given genomic window or locus:
  • Compute dmin, defined as the minimum pairwise sequence distance between any haplotype from species X and any haplotype from species Y [5].
  • Compute dXY, the average pairwise sequence distance between all haplotypes in X and all haplotypes in Y [5].
  • Compute dXO and dYO, the average distances from each sister species to the outgroup. Then calculate d_out = (dXO + dYO)/2 [5].
• Compute RNDmin: Calculate the statistic as RNDmin = dmin / d_out [5].
• Generate a Null Distribution: Simulate the expected distribution of RNDmin under a model of no migration using coalescent simulations. This null model must incorporate the specific demographic history and variation in neutral mutation rates [5].
• Statistical Testing: Identify candidate introgressed regions by comparing the observed RNDmin value to the simulated null distribution. Significantly low values of RNDmin (in the lower tail of the distribution, below a specified P-value threshold) provide evidence for introgression, as they indicate regions with exceptionally high similarity between species that cannot be explained by shared ancestry alone [5].

Protocol: A Machine Learning Workflow for Classification

Supervised machine learning offers a powerful, multi-faceted approach to distinguish between speciation with ILS and histories involving introgression [4] [3].

• Training Data Simulation: Use a forwards-in-time simulator (e.g., SLiM within the stdpopsim framework) to generate a large number of genomic datasets under two distinct scenarios: (a) a pure speciation model with ILS and no gene flow, and (b) a model that includes periods of introgression between diverging lineages [3].
• Feature Extraction: From each simulated dataset, calculate a suite of summary statistics and features that are informative of demographic history. These can include:
  • Site Frequency Spectrum (SFS) summaries.
  • Statistics of linkage disequilibrium (LD).
  • Pairwise genetic distances (dXY).
  • Measures of tree topology and branch length (e.g., gene tree node heights) [4].
  • For CNN-based approaches, input is often a genotype matrix from multiple populations, preserving spatial information of variants [3].
• Model Training and Validation: Train a classifier (e.g., a supervised learning model such as a random forest, or a Convolutional Neural Network) using the simulated features as input and the known evolutionary scenario (e.g., "ILS" or "Introgression") as the label. Validate model accuracy on a held-out portion of the simulated data [4] [3].
• Application to Empirical Data: Apply the trained and validated model to empirical genomic data from the species of interest. The model will output a probability or classification for the observed data, indicating the most likely underlying evolutionary history [4].

The Scientist's Toolkit: Essential Research Reagents and Solutions

This section catalogs key methodological "reagents" — computational tools and analytical frameworks — essential for research in this field.

Table 3: Key Research Reagent Solutions for Phylogenomic Conflict Analysis

Research Reagent	Category	Primary Function
RNDmin/Gmin [5]	Summary Statistic	A mutation-rate-robust metric to detect introgressed loci based on minimum sequence divergence.
D-statistic (ABBA-BABA) [2]	Summary Statistic	Tests for asymmetry in allele sharing patterns among four taxa to signal introgression.
Coalescent Samplers [1]	Probabilistic Model	Infers population parameters (divergence time, migration rates) using a model-based framework.
PhyloNet [4]	Phylogenetic Network	Infers phylogenetic networks and estimates the contribution of introgression to genomic data.
ALTS [6]	Phylogenetic Network	Infers the minimum tree-child network that displays a set of input gene trees.
SLiM + stdpopsim [3]	Simulation Framework	Forward-time simulator for generating genomic data under complex evolutionary models (e.g., with selection and introgression).
Convolutional Neural Networks (CNNs) [3]	Machine Learning	Classifies genomic windows as evolving neutrally or under adaptive introgression from genotype matrices.

Introgression and incomplete lineage sorting are distinct evolutionary forces that leave complex and often confounding signatures in genomic data. Introgression acts as a structured, directional force that can transfer adaptive traits across species boundaries, while ILS is a stochastic outcome of the sorting process in rapidly diverging lineages. Disentangling them requires a multi-pronged approach, leveraging both traditional summary statistics and model-based methods, as well as the emerging power of machine learning. The accurate characterization of these processes is not merely an academic exercise; it is fundamental to reconstructing the true history of life, which is often reticulate rather than strictly tree-like, and for identifying the genetic basis of adaptive evolution.

In the field of phylogenomics, gene tree heterogeneity presents a fundamental challenge to accurately reconstructing evolutionary histories. This phenomenon, where different genomic regions tell conflicting stories about species relationships, complicates the characterization of introgression using phylogenetic networks. Biological processes including incomplete lineage sorting (ILS), gene flow, and hybridization create patterns of discordance that can be difficult to distinguish from analytical artifacts. Genomic data reveals that a species' evolutionary history is not always best represented by a single bifurcating tree, but rather by a complex network of relationships where the genome functions as a mosaic of different evolutionary histories [7]. Understanding the relative contributions of these biological sources is crucial for developing more accurate phylogenetic networks, particularly for introgression characterization research where distinguishing true historical gene flow from other sources of discordance is paramount.

The emerging consensus suggests that recombination rate variation across genomes plays a critical role in structuring this phylogenetic discordance. Regions with high recombination rates experience more frequent introgression because genetic material can be more effectively unlinked from negative epistatic interactions in hybrid backgrounds. Conversely, genomic regions with low recombination rates tend to better preserve the true species history [7]. This review systematically compares the biological factors contributing to gene tree heterogeneity, providing experimental data and methodological frameworks essential for researchers aiming to improve the accuracy of phylogenetic networks in characterizing introgression.

Relative Contributions of Different Biological Factors

A comprehensive decomposition analysis conducted on Fagaceae species quantified the relative contributions of different factors to gene tree variation. The results, drawn from 2124 nuclear loci across 90 species, provide crucial insight into the primary drivers of phylogenetic discordance [8].

Table 1: Relative Contributions to Gene Tree Discordance in Fagaceae

Biological Source	Contribution to Variation	Key Characteristics
Gene Tree Estimation Error (GTEE)	21.19%	Analytical artifact arising from limited phylogenetic signal, particularly problematic with short sequence alignments and high rate heterogeneity
Incomplete Lineage Sorting (ILS)	9.84%	Results from retention of ancestral polymorphisms during rapid speciation events; creates random discordance patterns
Gene Flow/Hybridization	7.76%	Creates structured discordance patterns; often shows relationship with recombination landscape

The same study revealed that approximately 58.1-59.5% of genes exhibited consistent phylogenetic signals ("consistent genes"), while 40.5-41.9% displayed conflicting signals ("inconsistent genes") [8]. Consistent genes showed stronger phylogenetic signals and were more likely to recover the species tree topology, though interestingly, consistent and inconsistent genes did not significantly differ in terms of sequence- and tree-based characteristics. This finding suggests that identifying problematic genes based on inherent sequence properties alone remains challenging.

Impact on Phylogenetic Inference Methods

The biological sources of heterogeneity differently impact various phylogenetic inference approaches:

• Concatenation-based methods: Assume a shared evolutionary history across all genes, making them vulnerable to inaccuracies when substantial ILS or gene flow is present [8]
• Coalescent-based methods: Explicitly account for ILS but may be confounded by extensive gene flow, particularly when GTEE is present [8]
• Quartet-based approaches: Offer computational efficiency and strong theoretical foundation but struggle with model misspecification when the true evolutionary history is more complex than assumed [9]

Table 2: Methodological Performance Across Heterogeneity Types

Phylogenetic Method	Performance with ILS	Performance with Gene Flow	Limitations
Concatenation	Poor with high ILS	Poor with extensive gene flow	Assumes shared evolutionary history
Coalescent (Summary)	Excellent	Moderate	Sensitive to GTEE
Quartet-based	Good	Moderate	Struggles with complex networks
PsiPartition	Good	Good	Automated partitioning reduces error [10]

Recent computational advances like PsiPartition offer promising approaches for handling site heterogeneity by dividing DNA data into evolutionary rate categories using advanced algorithms and Bayesian optimization. This tool automatically identifies the optimal number of partitions, saving time and reducing errors common in traditional methods [10].

Experimental Evidence and Biological Mechanisms

Cytoplasmic-Nuclear Discordance in Fagaceae

Empirical evidence from Fagaceae research demonstrates how biological processes create recognizable patterns of discordance. Phylogenetic analyses of chloroplast DNA (cpDNA) and mitochondrial DNA (mtDNA) divided Fagaceae species into New World and Old World clades, a pattern that sharply contrasted with phylogenetic relationships inferred from nuclear genome data [8]. These cytoplasmic-nuclear discordances strongly suggest ancient interspecific hybridization, where the cytoplasmic genomes (typically maternally inherited) captured a different evolutionary history than the nuclear genome.

This research employed detailed methodological protocols to generate robust evidence:

• Mitochondrial genome assembly: Used GetOrganelle v1.7.1 with depth <25× filtering to eliminate nuclear contamination [8]
• SNP calling: Implemented GATK "HaplotypeCaller" with quality filtering (base quality ≥30, mapping quality ≥30) [8]
• Phylogenetic analysis: Combined Maximum Likelihood (IQ-TREE) and Bayesian inference (MrBayes) for robustness [8]

The experimental workflow exemplifies comprehensive approaches needed to distinguish biological heterogeneity from analytical artifacts.

Recombination Rate Variation as a Predictor

The recombination landscape has emerged as a reliable predictor of genomic regions that best represent the species tree. Research across diverse eukaryotic taxa demonstrates that:

• Low-recombination regions more accurately preserve species history because introgressed ancestry is effectively purged by selection due to linkage with negatively selected variants [7]
• High-recombination regions exhibit more frequent introgression because foreign alleles can be unlinked from negative epistatic interactions in hybrid backgrounds [7]
• Sex chromosomes (X/Z chromosomes) in heteromorphic systems consistently show enrichment for the species tree, likely due to their lower effective population sizes and reduced recombination rates [7]

This recombination-based heterogeneity creates a genomic mosaic where different chromosomal regions reflect different evolutionary histories, complicating species tree inference but providing valuable information about historical introgression events.

Methodological Framework for Characterizing Heterogeneity

Analytical Workflow for Discordance Investigation

A robust methodological framework is essential for accurately characterizing biological sources of gene tree heterogeneity. The following workflow synthesizes best practices from recent studies:

Research Reagent Solutions for Phylogenomic Studies

Table 3: Essential Research Reagents and Computational Tools

Tool/Reagent	Function	Application in Heterogeneity Research
GetOrganelle v1.7.1	Organelle genome assembly	Assembling mitochondrial and chloroplast genomes for cytoplasmic discordance analysis [8]
GATK "HaplotypeCaller"	SNP calling from aligned reads	Identifying reliable genetic variants while filtering low-quality data [8]
PsiPartition	Evolutionary rate partitioning	Automatically grouping DNA data into evolutionary rate categories to account for site heterogeneity [10]
IQ-TREE v2.3.6	Maximum likelihood phylogenetic inference	Estimating gene trees with robust statistical support [8]
MrBayes v3.2.6	Bayesian phylogenetic inference	Alternative tree inference using Bayesian Markov Chain Monte Carlo methods [8]
BEAST2	Bayesian molecular dating	Estimating divergence times with relaxed clock models [11]
ASTRAL	Coalescent-based species tree estimation	Handling incomplete lineage sorting in species tree reconstruction [9]

Implications for Introgression Characterization Research

For researchers focused on accuracy of phylogenetic networks for introgression characterization, understanding biological heterogeneity sources has profound implications:

• Discordance Pattern Recognition: Gene tree heterogeneity resulting from introgression shows distinct patterns from ILS-induced heterogeneity. Introgression creates structured phylogenetic discordance that correlates with recombination landscape, while ILS creates more random discordance patterns [7]
• Data Filtering Strategies: Removing 40.5-41.9% of "inconsistent genes" (those displaying conflicting phylogenetic signals) can significantly reduce concatenation- and coalescent-based approach inconsistencies [8]
• Genome Assembly Quality: Chromosome-level genome assemblies are crucial for understanding the interplay between chromosome evolution, recombination landscape, and phylogenetic signal [7]
• Temporal Framework Integration: Molecular dating of single gene trees faces significant uncertainty due to substitution rate variation, complicating the temporal placement of introgression events [11]

The integration of recombination rate evolution and phylogenetic variation represents the future of accurate introgression characterization, moving beyond the assumption of a single bifurcating tree toward network-based models that accommodate the mosaic nature of genomic ancestry [7]. Methods that jointly model duplication, loss, introgression, and coalescence offer promising frameworks for detecting introgression presence and determining the number of unique introgression events in a species tree [9].

Biological sources of gene tree heterogeneity, particularly incomplete lineage sorting (9.84%) and gene flow (7.76%), present significant challenges but also opportunities for refining phylogenetic networks in introgression research. The quantitative decomposition of these factors enables more targeted analytical approaches, while recognizing recombination rate variation as a predictor of phylogenetic signal location provides a roadmap for selecting genomic regions most likely to preserve species history. For researchers characterizing introgression, the strategic exclusion of inconsistent genes, careful attention to recombination landscapes, and utilization of emerging computational tools like PsiPartition [10] and unified models of introgression and coalescence [9] will significantly enhance accuracy. Future directions should prioritize the development of recombination-aware phylogenomic methods and the collection of chromosome-scale genomes to fully leverage the predictable patterns of heterogeneity revealed by recent studies.

The Multispecies Coalescent (MSC) model represents a fundamental extension of the single-population coalescent to multiple species, integrating both the phylogenetic process of species divergences and the population genetic process of coalescence [12]. This mathematical framework has emerged as a powerful approach for addressing complex evolutionary questions using genomic sequence data from multiple species. By modeling how gene lineages coalesce within a species tree, the MSC provides the statistical foundation for understanding genealogical discordance—the phenomenon where gene trees differ from each other and from the species tree [12] [13]. This discordance arises naturally from population genetic processes such as incomplete lineage sorting (ILS), which occurs when ancestral polymorphisms persist through multiple speciation events and are randomly fixed in descendant lineages [14].

The MSC model has revolutionized phylogenomics by shifting the perspective on gene tree heterogeneity from being considered a "problem" to being recognized as a valuable source of information about evolutionary parameters such as ancestral population sizes and rates of cross-species gene flow [12]. When extended to phylogenetic networks through the Network Multispecies Coalescent (NMSC), this framework can simultaneously account for both ILS and reticulate evolutionary processes such as hybridization and introgression, providing a more comprehensive model for inferring evolutionary histories [14]. This integrated approach is particularly valuable for characterizing introgression, as it allows researchers to distinguish between signals of deep coalescence and those resulting from historical gene flow.

MSC Theory and Extension to Phylogenetic Networks

Theoretical Foundations of the Multispecies Coalescent

The MSC model builds upon the standard coalescent theory, which describes the genealogical history of a sample of DNA sequences taken from a population as a stochastic process tracing lineage joining backwards in time [12]. The key innovation of the MSC is its placement of this coalescent process within the context of a species phylogeny, requiring two sets of parameters: species divergence times (τ) and population size parameters (θ) for both extant and ancestral species [12]. In this model, coalescent events occur independently in different populations with rates determined by population sizes, and when lineages reach speciation events backward in time, the coalescent process is reset to account for changes in population size and the addition of lineages from sibling species.

A crucial feature of the MSC model is that gene trees are embedded within species trees, meaning the divergence time between sequences from two species must be greater than the species divergence time [12]. This intrinsic constraint creates computational challenges but also provides the statistical power for estimating evolutionary parameters. The MSC gives rise to two important probability distributions: the marginal probabilities of gene tree topologies and the joint distribution of gene tree topologies and coalescent times, both of which are utilized in different inference methods [12].

Expansion to the Network Multispecies Coalescent

The Network Multispecies Coalescent (NMSC) extends the MSC framework to accommodate reticulate evolution by incorporating hybridization nodes that allow two incoming branches from different parental species [14]. In this model, each reticulation event is parameterized by an inheritance probability (γ) representing the proportion of genetic material that the hybrid lineage derives from each parent [14]. This critical parameter distinguishes between symmetrical hybridization (γ ≈ 0.5), where both parents contribute roughly equally, and asymmetrical introgression (γ close to 0 or 1), where one parent contributes disproportionately more genetic material.

The NMSC provides a biologically intuitive approach to modeling evolutionary processes that cannot be adequately represented by strictly bifurcating trees. Unlike implicit phylogenetic networks that merely summarize discordance without biological interpretation, explicit networks under the NMSC directly link evolutionary processes to patterns in genetic data, enabling meaningful biological conclusions about historical reticulation events [14]. This makes the NMSC particularly valuable for introgression characterization, as it can distinguish gene flow signals from those produced by ILS alone and can localize introgression events in evolutionary history.

Table: Key Parameters in MSC and Network MSC Models

Parameter	Symbol	Interpretation	Role in Inference
Population size	θ	Measure of genetic diversity; θ = 4Neμ	Determines coalescence rate within populations
Divergence time	τ	Time of species splitting events	Provides temporal framework for gene tree embedding
Inheritance probability	γ	Proportion of genetic material from each parent in hybridization	Quantifies directionality and strength of introgression
Coalescent times	ti	Waiting times until lineage coalescence	Provides information about population sizes and divergence times

Comparative Analysis of Network Inference Methods

Methodological Approaches and Their Theoretical Bases

Various computational methods have been developed to infer phylogenetic networks under the NMSC framework, each with distinct theoretical foundations and statistical approaches. PhyNEST represents a novel composite likelihood method that estimates binary, level-1 phylogenetic networks directly from sequence data without requiring gene tree summarization as an intermediate step [15]. This approach uses site pattern frequencies across the genome as the basis for inference and implements both hill climbing and simulated annealing algorithms to search network space. Unlike earlier methods, PhyNEST maintains computational tractability while using full genomic data, assuming coalescent-independent sites evolving under the Jukes-Cantor substitution model with constant effective population size [15].

Alternative approaches include Bayesian species delimitation methods such as those implemented in BPP and BEAST2 packages (e.g., DISSECT, STACEY), which use the MSC to test species boundaries by determining whether sequence assignments to species correspond to distinct evolutionary lineages [16]. These methods employ different search strategies, including reversible-jump Markov Chain Monte Carlo (MCMC) in BPP and birth-death collapse models in DISSECT/STACEY, where samples with divergence times below a threshold (ϵ) are considered conspecific [16]. More recent developments like the Yule-skyline collapse model in the SPEEDEMON package allow the speciation rate to vary through time as a smooth piecewise function, increasing biological realism in species delimitation [16].

Table: Comparison of Network Inference Methods Based on MSC Framework

Method	Statistical Approach	Data Input	Key Assumptions	Computational Scalability
PhyNEST	Composite likelihood	Sequence alignments	Level-1 networks, constant population size	Suitable for genome-scale data [15]
BPP	Bayesian (reversible-jump MCMC)	Sequence alignments	User-specified guide tree, neutral evolution	Moderate; limited model flexibility [16]
DISSECT/STACEY	Bayesian (birth-death collapse)	Sequence alignments or SNPs	Threshold-based species assignment	Improved efficiency with multithreading [16]
SNAPP	Bayesian (MCMC)	SNP data	Neutral evolution, no recombination within loci	Efficient for SNP data [16]
StarBeast3	Bayesian (MCMC)	Multilocus sequences	Strict or relaxed molecular clock	High efficiency with parallelization [16]

Performance Benchmarks and Accuracy Assessments

Recent benchmarking studies provide quantitative comparisons of method performance under various evolutionary scenarios. StarBeast3 demonstrates significant efficiency improvements over earlier implementations when run in multithreaded mode, producing 1.3 to 9.5 times more effective samples per hour depending on the parameter and dataset [16]. This enhanced performance is attributed to parallelized gene tree inference and highly efficient relaxed clock proposals, enabling more rapid convergence of phylogenetic parameters in Bayesian MCMC analyses.

In simulation studies, PhyNEST has shown superior accuracy compared to existing composite likelihood methods like SNaQ and PhyloNet, particularly in scenarios with known hybridization events [15]. The method has proven robust to certain forms of model misspecification, such as analyzing data with a simpler nucleotide substitution model than the true generating model. These validation experiments demonstrate that MSC-based network inference methods can accurately recover known parameters and species assignments when model assumptions are reasonably met.

For species delimitation, validation studies using the Yule-skyline collapse model in both SNAPPER (for SNP data) and StarBeast3 (for sequence data) have demonstrated well-calibrated performance, with true parameter values falling within the 95% highest posterior density intervals in approximately 95% of simulations [16]. These methods also accurately estimate cluster support probabilities across the full range of possible values, providing reliable measures of uncertainty in species boundary hypotheses.

Experimental Protocols for Method Validation

Simulation-Based Validation Framework

Robust evaluation of MSC-based network inference methods relies on comprehensive simulation studies that quantify performance under known evolutionary scenarios. A standard validation protocol involves: (1) sampling species trees and associated gene trees from the prior distribution; (2) simulating sequence alignments or SNP datasets under the generated trees; (3) performing Bayesian MCMC inference on the simulated datasets; and (4) comparing estimated parameters to their true values to calculate coverage probabilities [16]. Well-calibrated methods should show approximately 95% coverage, where true parameter values fall within the 95% highest posterior density intervals in 90-99% of simulations.

For assessing species delimitation accuracy, cluster posterior supports are discretized into evenly-spaced bins (e.g., 20 bins from 0-100%), and for each bin, researchers count how frequently clusters with that support level correspond to true species boundaries in the simulation [16]. This approach validates that posterior probabilities provide accurate measures of statistical confidence, with clusters having 50-55% posterior support truly existing 50-55% of the time. Sensitivity analyses examining robustness to key parameters like the collapse threshold (ϵ) are also essential components of thorough method validation [16].

Empirical Validation with Biological Datasets

Empirical validation applies MSC-based network methods to organisms with well-established evolutionary histories or distinctive hybridization patterns. For example, researchers have applied these methods to Heliconius butterflies, known for extensive hybrid speciation, and Papionini primates, characterized by widespread introgression [15]. These biological test cases provide critical assessments of method performance using real genomic data where certain reticulation events have been previously documented through multiple lines of evidence.

Another important validation approach involves congruence testing across methods, where results from MSC-based network inference are compared to those from other phylogenetic approaches, such as D-statistics or demographic modeling [14]. Discrepancies between different methods—such as when phylogenetic networks detect fewer reticulation events than suggested by hybridization tests—highlight limitations of current approaches and areas needing methodological refinement [14]. Such comparative analyses on empirical datasets help establish the biological relevance and practical utility of MSC-based network inference.

Research Reagent Solutions for MSC-Based Inference

Table: Essential Computational Tools for MSC-Based Network Inference

Tool/Software	Primary Function	Data Requirements	Implementation
PhyNEST	Phylogenetic network estimation	Sequence alignments	Julia package [15]
BEAST2 with SPEEDEMON	Bayesian species delimitation	Sequence alignments or SNPs	BEAST2 package [16]
StarBeast3	Multispecies coalescent inference	Multilocus sequences	BEAST2 package [16]
SNAPPER	Species delimitation with SNPs	SNP data	BEAST2 package [16]
BPP	Species tree estimation and delimitation	Sequence alignments	Standalone program [16]

Signaling Pathways and Method Workflows

Future Perspectives and Research Directions

The field of MSC-based network inference continues to evolve rapidly, with several promising research directions emerging. Computational scalability remains a significant challenge, particularly for analyzing genome-scale datasets with complex evolutionary histories involving multiple reticulation events [15] [9]. Future methodological developments will likely focus on more efficient algorithms for exploring network space and approximating the likelihood function without sacrificing statistical accuracy.

Another important frontier involves integrating additional evolutionary processes into the MSC framework, such as gene duplication and loss, recombination, and selection [9]. Current research is already extending the MSC to model genealogical relationships among loci related by duplication events and to calculate gene tree probabilities when introgression is acting [9]. These developments will enhance the biological realism of MSC-based models and expand their applicability to diverse evolutionary scenarios.

As phylogenetic networks gain wider adoption in evolutionary biology and biodiversity research, they are poised to influence conservation biology by providing insights into historical connectivity between species and populations [14]. This is particularly relevant for groups of conservation concern that lack reference genome resources and explicit hypotheses from prior investigation. The emerging probabilistic framework for inferring historical reticulation events will enable more informed conservation decisions that account for complex evolutionary histories.

In the field of evolutionary biology, accurately reconstructing the history of trait evolution is fundamental to understanding diversification, adaptation, and the very drivers of speciation. Phylogenetic analyses traditionally assume that traits evolve along the species tree. However, the pervasive presence of gene tree discordance—where gene histories differ from the species tree—can severely challenge this assumption, leading to systematic errors in interpreting trait evolution [17]. Within this context, two distinct phenomena, hemiplasy and homoplasy, produce nearly identical patterns of trait incongruence but have profoundly different evolutionary implications.

Homoplasy represents true convergent evolution, where the same trait evolves independently multiple times via separate mutational events. Hemiplasy, in contrast, occurs when a single trait transition happens on a discordant gene tree, making it appear incongruent with the species tree despite a single origin [17] [18]. Distinguishing between these processes is not merely academic; it fundamentally affects inferences about the number, timing, and direction of trait transitions, and ultimately, our understanding of whether natural selection has repeatedly favored the same solution. This guide provides a structured comparison of hemiplasy and homoplasy, focusing on their implications for analyzing trait evolution within phylogenetic networks, particularly when characterizing introgression.

Defining the Concepts and Their Evolutionary Basis

Homoplasy: Independent Evolution of Similar Traits

Homoplasy, encompassing both convergence and parallelism, arises when similar phenotypic traits evolve independently in distinct lineages through different genetic mutations or developmental pathways. This process implies that natural selection has repeatedly arrived at the same adaptive solution in separate lineages facing similar environmental pressures. The inference of homoplasy relies on the assumption that the species tree accurately represents the history of all traits, an assumption now known to be frequently violated due to widespread gene tree discordance [17].

Hemiplasy: A Single Trait Transition on a Discordant Gene Tree

Hemiplasy occurs when a mutation arises on a branch of a gene tree that is discordant from the species tree. This single evolutionary event can create a distribution of character states among species that appears to require multiple independent origins when mapped onto the species tree, thus masquerading as homoplasy [17] [18]. The probability of hemiplasy is directly tied to the probability of gene tree discordance, which has two primary biological causes: Incomplete Lineage Sorting (ILS) and introgression.

Table 1: Fundamental Concepts in Trait Evolution Analysis

Concept	Definition	Evolutionary Mechanism	Key Implication
Homoplasy	Independent evolution of similar traits in different lineages	Convergent evolution via multiple independent mutations	Suggests strong, repeated selective pressure
Hemiplasy	Incongruence from a single trait transition on a discordant gene tree	Single mutation subject to ILS and/or introgression	Can mimic convergence without repeated selection
Incomplete Lineage Sorting (ILS)	Failure of gene lineages to coalesce before subsequent speciation	Deep coalescence; common with short internal branches/small populations	Causes discordance even without gene flow
Introgression	Transfer of genetic material between species through hybridization	Gene flow following hybridization events	Creates discordance with predictable topological patterns

Quantitative Comparison: Probabilities and Influencing Factors

The probability of hemiplasy versus homoplasy is influenced by distinct, quantifiable parameters. Guerrero and Hahn (2018) developed a model showing that the key factors are the internal branch length of the species tree and the mutation rate [17]. Short internal branches increase the likelihood of discordance due to ILS, thereby elevating the hemiplasy risk. Conversely, a low mutation rate reduces the probability of the multiple independent transitions required for homoplasy, making hemiplasy a more likely explanation for observed incongruences [17].

Introgression further modifies these probabilities. Recent and frequent introgression makes hemiplasy more likely than under ILS alone. Methods that account only for ILS will therefore be conservative, potentially underestimating the true risk of hemiplasy in systems with historical gene flow [17].

Table 2: Factors Influencing the Probability of Hemiplasy vs. Homoplasy

Factor	Effect on Hemiplasy Probability	Effect on Homoplasy Probability	Practical Implication for Inference
Short Internal Branches	Increases	No direct effect	Short branches elevate discordance risk, favoring hemiplasy解释
Low Mutation Rate	Increases	Decreases	Low rate makes multiple independent mutations unlikely
High Population Size	Increases	No direct effect	Increases ILS, thereby increasing discordance
Introgression	Increases	No direct effect	Makes hemiplasy more likely than ILS alone; must be modeled
Recent Introgression	Strongly increases	No direct effect	Recent gene flow dramatically elevates hemiplasy risk

Methodologies for Dissecting Hemiplasy and Homoplasy

The HeIST Tool: A Coalescent Simulation Approach

For complex phylogenies with more than three taxa, explicit mathematical solutions for hemiplasy probabilities become infeasible. HeIST (Hemiplasy Inference Simulation Tool) addresses this by using coalescent simulations within a user-specified phylogenetic network that incorporates both ILS and introgression [17]. The workflow involves:

• Input Specification: The user provides a species tree or network with branch lengths (in coalescent units), population sizes, and the location, direction, and timing of introgression events.
• Coalescent Simulation: HeIST simulates a large number of gene trees within the given network model.
• Trait Mapping: For each simulated gene tree, the tool models the evolution of a binary trait, assessing the number and location of mutations required to produce the observed trait pattern.
• Statistical Inference: The output provides an estimate of the probability that the observed trait incongruence is due to hemiplasy (a single transition) versus homoplasy (multiple transitions) [17].

Empirical Protocol from a Phylotranscriptomic Study

A study on Allium subgenus Cyathophora provides a clear experimental protocol for assessing these phenomena [18]:

• Data Collection: Generate large-scale molecular datasets, such as transcriptomes (as used in Allium) or hybrid capture data (Hyb-Seq), from the studied taxa. Include whole chloroplast genomes to compare organellar and nuclear histories.
• Phylogenetic Inference: Reconstruct the species tree using both concatenation and coalescence-based methods on single-copy genes (SCGs) to establish a robust primary hypothesis.
• Quantify Gene Tree Discordance: Calculate the proportion of SCGs whose topologies conflict with the established species tree. In Allium, 27%-38.9% of genes were discordant [18].
• Determine the Cause of Discordance: Use coalescent simulations to test whether the observed distribution of gene trees is consistent with ILS alone. Alternative explanations, particularly introgression, should be evaluated using phylogenetic network methods (e.g., PhyloNet) and tests like D-statistics.
• Calculate Hemiplasy Risk: Apply models (like those in HeIST or analytical formulas) to compute the hemiplasy risk factor, evaluating whether hemiplasy is a sufficient explanation for trait incongruence without invoking multiple independent origins [18].

Case Studies in Empirical Research

Hemiplasy Driven by ILS inAllium

A phylotranscriptomic study of Allium subgenus Cyathophora found high gene tree discordance (27%-38.9%) but determined through coalescent simulations that ILS was the primary driver, with no significant role for introgression. The study concluded that hemiplasy was the most likely explanation for the observed trait transitions and an anomalous chloroplast DNA tree, rather than multiple independent homoplastic mutations [18]. This demonstrates that even in the absence of introgression, failure to account for ILS can lead to overestimation of convergent evolution.

The Complex Role of Introgression inPicrisEvolution

A study on diploid Picris species in the Mediterranean Basin revealed that historical introgression played a major role in the genus's diversification. Phylogenetic network analyses identified two major introgression events. However, in one critical case, introgression was found to precede shifts in life strategy and fruit morphology, ruling out the direct transfer of these traits via adaptive introgression. This shows that while introgression can be a key driver of diversification, it does not always cause trait transitions through hemiplasy; its role must be tested on a case-by-case basis [19].

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Key Research Reagents and Computational Tools for Analysis

Tool/Reagent	Function/Application	Utility in Discrimination
HeIST (Hemiplasy Inference Simulation Tool)	Coalescent simulation in species networks	Estimates most likely number of trait transitions, accounting for ILS & introgression [17]
Phylogenomic Datasets (Hyb-Seq, RNA-seq)	Generating genome-wide single-copy nuclear genes	Provides data for robust species tree inference and quantification of GTD [18] [19]
PhyloNet/Similar Software	Inference and analysis of phylogenetic networks	Detects and models historical introgression events [19]
Coalescent Simulators (e.g., ms)	Simulating gene trees under ILS and introgression	Generates null distributions of GTD to test its primary cause [18]
D-Statistics (ABBA-BABA)	Testing for gene flow against a null of ILS	Provides a statistical test for introgression between taxa [17]

Hemiplasy and homoplasy are fundamentally different evolutionary processes that produce deceptively similar patterns. Distinguishing between them requires moving beyond simple trait mapping on a species tree to a more sophisticated framework that explicitly accounts for pervasive gene tree discordance. As the case studies show, the relative contributions of ILS and introgression to discordance are system-specific, and this directly impacts the probability of hemiplasy. Accurate inference therefore depends on the use of genomic-scale data, phylogenetic networks, and specialized tools like HeIST. For researchers and drug development professionals working with trait evolution, incorporating these concepts and methods is no longer optional but essential for generating biologically accurate conclusions about the number, timing, and selective basis of phenotypic transitions.

Phylogenetic networks are crucial for modeling complex evolutionary histories involving reticulate events such as introgression, hybridization, and horizontal gene transfer. Accurately reconstructing these networks from molecular data is fundamental for introgression characterization research, with significant implications for understanding drug target evolution and pathogen diversity. Rooted triplets (three-leaf rooted trees) and quartets (four-leaf unrooted trees) serve as fundamental building blocks for many phylogenetic inference methods. The minimum sampling requirements—the type and amount of data needed for reliable inference—differ substantially between these approaches due to their distinct statistical properties under various evolutionary models. This guide provides an objective comparison of the performance, data requirements, and applicability of triplet versus quartet-based methods for phylogenetic network reconstruction, with particular emphasis on characterizing introgression landscapes in evolutionary genomics.

Theoretical Foundations: Triplets vs. Quartets

Basic Definitions and Properties

Rooted triplets are rooted, binary phylogenetic trees with three leaves, representing the simplest possible resolved evolutionary relationships among three taxa. The three possible triplets on taxon set {A,B,C} are denoted as tA = A|BC, tB = B|AC, and tC = C|AB, where the notation X|YZ indicates that taxa Y and Z share a more recent common ancestor with each other than with X [20].

Quartets are unrooted, binary phylogenetic trees with four leaves, representing unrooted evolutionary relationships among four taxa. The three possible quartets on taxon set {A,B,C,D} are denoted as q1 = AB|CD, q2 = AC|BD, and q3 = AD|BC, where AB|CD indicates that taxa A and B form a clade separate from taxa C and D [20] [21].

Theoretical work has established that quartet-based methods offer important statistical advantages under many evolutionary models. Specifically, under the Infinite Sites plus Unbiased Error and Missingness (IS+UEM) model—a popular framework for tumor phylogenetics—there are no anomalous quartets, meaning the most probable quartet topology matches the true unrooted model tree topology. This property does not extend to triplets, which can be anomalous under the same model [20].

Consistency and Statistical Guarantees

Consistency is a crucial property for phylogenetic inference methods, ensuring that as more data (e.g., longer sequences or more loci) becomes available, the estimated tree or network converges to the true evolutionary history. Quartet-based methods have been proven statistically consistent under various models, including the multi-species coalescent (MSC) and IS+UEM models [20] [21].

Table 1: Theoretical Properties of Triplet vs. Quartet Approaches

Property	Rooted Triplets	Quartets
Anomaly Zone	Exists under IS+UEM model [20]	No anomalies under IS+UEM model [20]
Data Requirements	Lower theoretical minimum taxa	Requires minimum of 4 taxa
Statistical Consistency	Limited under certain models [20]	Proven under MSC and IS+UEM models [20] [21]
Resolution Power	Limited for deep evolutionary relationships	Strong for resolving conflicting signals [21]
Computational Complexity	Generally lower	Higher but more informative

The diagram below illustrates the fundamental structural differences between triplets and quartets and their relationship to full phylogenetic networks:

Figure 1: Phylogenetic inference workflow showing triplet and quartet integration paths

Methodological Comparison

Quartet-Based Method Implementations

Multiple software implementations have been developed for quartet-based phylogenetic inference, each with distinct approaches to handling quartet information:

QuartetSuite encompasses three primary methods: QuartetS (minimum method), QuartetA (average method), and QuartetM (maximum method). These methods function by iteratively decomposing all triplet and quartet weights into simple components based on full splits, differing primarily in how they handle multiple possible weights for a split. QuartetS takes the minimum value, QuartetA computes the average, and QuartetM selects the maximum value when multiple weighting scenarios exist [21].

ASTRAL is a leading method for species tree estimation based on quartet frequencies, widely regarded for its statistical consistency under the multi-species coalescent model. It operates by seeking the tree that shares the maximum number of quartets with the input gene trees [20].

Other quartet methods include QNet, SuperQ, and QuartetNet, each with specific consistency guarantees on different types of split systems (circular, weakly compatible, or 2-weakly compatible) [21].

Triplet-Based Method Implementations

While less emphasized in the search results, triplet-based approaches typically involve assembling larger phylogenetic structures from rooted three-taxon relationships. These methods often face limitations due to the potential for anomalous triplets under models like IS+UEM, where the most probable triplet topology may not match the true rooted model tree topology [20].

Performance Comparison: Experimental Data

Simulation Studies

Comprehensive simulation studies have evaluated the performance of triplet and quartet-based methods under controlled conditions with known evolutionary histories:

Table 2: Performance on Simulated Tree Data (100 replicates) [21]

Method	True Splits Reconstructed	False Positive Splits	Trivial Split Weight Accuracy
QuartetS	100%	None	Moderate (RMSE: N/A)
QuartetA	100%	None	High (RMSE: 0.016)
QuartetM	100%	None	Low
Quartet-Net	100%	Few with low weights	Low
Neighbor-Net	100%	10+ with bootstrap 15-40	Low
Neighbor-Joining	100%	None	Low

Table 3: Performance on Simulated Network Data with 3 Reticulate Events [21]

Method	True Splits Reconstructed	False Negative Splits	Non-Trivial Split Weight Accuracy
QuartetS	100%	None	High (RMSE: 0.054)
QuartetA	100%	None	Moderate (RMSE: 0.124 for trivial splits)
QuartetM	100%	None	Moderate
Quartet-Net	100%	None	Moderate
Neighbor-Net	<50%	Multiple major splits	Low
Neighbor-Joining	<50%	Multiple major splits	Low

Experimental protocols for these simulations typically involved:

• Data Generation: Using software like Dawg [21] to generate DNA sequences under evolutionary models (e.g., GTR+Gamma+I) with specified parameters such as substitution rate (0.01) and sequence length (10,000 bp for trees, 80,000 bp for networks).
• Multiple Replicates: Conducting 100 independent runs to account for stochastic variation.
• Evaluation Metrics: Assessing accuracy based on recovery of true splits, absence of false positives, and root mean square error (RMSE) between estimated and true split weights.

Real-World Biological Datasets

Bacterial Dataset Analysis

A study of 36 bacterial species using seven concatenated genes—where few reticulate events are expected—demonstrated that QuartetA most accurately reconstructed the known evolutionary relationships with minimal false positives, making it ideal for primarily tree-like phylogenies [21].

Influenza H7N9 Dataset Analysis

Analysis of 22 influenza A viruses related to H7N9 emergence pathways revealed that quartet-based methods successfully identified reassortment events and evolutionary relationships that triplet-based approaches and distance methods missed, providing critical insights into the origins of this public health threat [21].

The following diagram illustrates a typical experimental workflow for comparing phylogenetic methods:

Figure 2: Experimental workflow for phylogenetic method comparison

Minimum Sampling Requirements and Data Considerations

Taxon Sampling Requirements

The minimum taxon sampling requirements differ fundamentally between triplet and quartet approaches:

• Triplet-based methods technically require only three taxa for basic operations but need extensive taxon sampling across the phylogeny for reliable network inference.
• Quartet-based methods require at least four taxa for each quartet but demonstrate more reliable performance with moderate sampling, as each quartet provides more phylogenetic information than triplets.

For introgression characterization, dense sampling across putative hybrid zones and parental populations is essential regardless of methodological approach.

Sequence Data Requirements

The amount and quality of sequence data significantly impact method performance:

• Quartet methods generally require less data than triplet methods to achieve comparable accuracy due to their stronger statistical properties.
• Simulation studies indicate that accurate reconstruction of networks with reticulate events requires longer alignments (e.g., 80,000 bp for networks with three reticulations) compared to tree-like phylogenies (e.g., 10,000 bp) [21].
• Data type considerations: Quartet methods have been successfully applied to diverse data types including nucleotides, amino acids, and morphological characters [22].

The Scientist's Toolkit: Essential Research Reagents

Table 4: Key Research Reagents for Triplet and Quartet-Based Phylogenetics

Reagent/Software	Type	Function	Application Context
QuartetSuite	Software package	Implements QuartetS, QuartetA, QuartetM methods	Phylogenetic network reconstruction from sequence data [21]
ASTRAL	Software package	Species tree estimation from quartet frequencies	Coalescent-based species tree inference [20]
ALTS	Software program	Infers tree-child networks by aligning lineage taxon strings	Phylogenetic network inference from gene trees [6]
Dawg	Sequence simulator	Generates evolved DNA sequences under specified models	Method validation and benchmarking [21]
Multiple Sequence Alignment	Data preparation	Aligns homologous sequences for phylogenetic analysis	Essential preprocessing step for all methods

Implications for Introgression Characterization Research

Accurate characterization of introgression—the transfer of genetic material between species or populations—requires methods that can reliably detect and represent reticulate evolutionary events. Quartet-based approaches offer significant advantages for this research domain due to their ability to:

• Handle conflicting signals from different genomic regions resulting from introgression events [21]
• Maintain statistical consistency even when the underlying evolutionary history includes reticulations [20]
• Accurately represent complex evolutionary scenarios involving multiple introgression events [21]

Recent methodological advances have enabled the detailed study of genomic landscapes of introgression across diverse evolutionary scenarios, including adaptive and ghost introgression, with quartet-based methods playing an increasingly important role in these analyses [23].

Quartet-based phylogenetic methods demonstrate superior performance compared to triplet-based approaches for most introgression characterization applications, particularly under models involving reticulate evolution. The theoretical absence of anomalous quartets under commonly used evolutionary models, combined with empirical evidence from both simulated and biological datasets, establishes quartet methods as the preferred choice for accurate network reconstruction. While triplet methods may offer computational advantages for some applications, their susceptibility to anomalous topologies and lower accuracy in recovering true splits limits their utility for complex evolutionary analyses. For researchers investigating introgression in drug development contexts—where accurate evolutionary reconstruction can identify transferred genetic elements relevant to disease or treatment response—quartet-based approaches provide more reliable inference of evolutionary relationships.

Methodological Approaches for Introgression Detection: From Summary Statistics to Network Inference

The D-statistic, commonly known as the ABBA-BABA test, is a cornerstone method in evolutionary genomics for detecting gene flow between closely related populations or species. Developed initially to test for hybridization between Neanderthals and modern humans, this method has since been applied across a broad range of taxa, from bacteria to plants and animals [24] [25]. The test operates on a simple but powerful principle: it detects statistical deviations from a strict bifurcating tree model by comparing patterns of shared genetic variation, specifically targeting excess allele sharing between non-sister taxa that signals introgression [24] [26].

In the context of phylogenetic network accuracy research, the D-statistic provides a critical tool for characterizing reticulate evolutionary events. Unlike methods that assume a purely tree-like history, the D-statistic formally tests for gene flow that creates phylogenetic incongruences, treating these discordances not as noise but as meaningful biological signals [26]. This approach has transformed our understanding of species boundaries, revealing that introgression is far more common than previously recognized across the tree of life [27] [28].

Fundamental Principles and Statistical Framework

Core Conceptual Framework

The D-statistic is designed for a four-taxon system (quartet) with an established phylogeny: (((P1, P2), P3), O), where O is an outgroup used to determine ancestral (A) and derived (B) alleles [24] [29]. The method examines biallelic single nucleotide polymorphisms (SNPs) and focuses on two specific site patterns:

• ABBA: Sites where P2 and P3 share the derived allele (B), while P1 has the ancestral allele (A)
• BABA: Sites where P1 and P3 share the derived allele (B), while P2 has the ancestral allele (A)

Under the null hypothesis of no gene flow, with incomplete lineage sorting (ILS) as the only source of genealogical discordance, ABBA and BABA patterns are expected to occur with equal frequency. A significant imbalance between these patterns indicates introgression—excess ABBA suggests gene flow between P2 and P3, while excess BABA suggests gene flow between P1 and P3 [24] [25].

Calculation and Interpretation

The D-statistic is calculated as:

D = (NABBA - NBABA) / (NABBA + NBABA)

where NABBA and NBABA represent the counts of each site pattern in the analyzed dataset [27]. Statistical significance is typically assessed using a Z-score based on block jackknifing, with |Z| > 3 considered significant evidence of introgression [25].

The value of D ranges from -1 to 1, with magnitude reflecting the strength of the introgression signal. However, D is not a direct measure of the proportion of introgressed genome, as its value is influenced by various factors including population sizes, divergence times, and the timing of gene flow [26].

Software Implementation Comparison

Multiple software packages implement the D-statistic and related methods, each with different capabilities, input requirements, and computational efficiencies. The table below provides a comparative overview of major tools:

Table 1: Comparison of Software Packages for D-statistic and Related Analyses

Software	VCF Input Support	Genome-wide D	f4-ratio	f-branch	Sliding Window Analyses	Specialized Statistics
Dsuite	Yes	Yes	Yes	Yes	Yes (fd, fdM, df)	fdM, f-branch
ADMIXTOOLS	Limited	Yes	Yes	No	No	D, f4-ratio
ANGSD	Yes	Yes	No	No	No	D
Comp-D	No	Yes	No	No	No	D
HyDe	Limited	Yes	No	No	No	Hybridization detection
PopGenome	Yes	Yes	No	No	Yes (D, fd, df)	D, fd, df

Dsuite emerges as a particularly comprehensive implementation, combining support for standard VCF format input with computational efficiency that enables analyses across hundreds of populations [24]. It uniquely implements several statistics not available in other packages, including the f-branch metric for assigning gene flow evidence to specific phylogenetic branches and fdM for window-based analyses [24]. This makes Dsuite especially valuable for large-scale genomic studies where computational practicality is a concern.

Methodological Protocols

Standard D-statistic Analysis Workflow

Diagram: D-statistic Analysis Workflow

The standard workflow begins with data collection and preparation, typically involving whole-genome sequencing data stored in VCF format. The researcher must define the phylogenetic relationships of the study system, selecting appropriate populations for the P1, P2, P3, and O roles in the quartet [24] [25].

Next, site pattern counting is performed across the genome, tallying occurrences of ABBA and BABA patterns. For reliable results, this should be based on a substantial number of informative sites—typically whole genomes or thousands of loci are required to achieve sufficient statistical power [29].

The D-statistic calculation follows, computing the normalized difference between ABBA and BABA counts. Finally, significance testing assesses whether the observed D-value significantly deviates from zero, typically using a block jackknife procedure to account for linked sites and generate a Z-score [25].

Advanced Methodological Extensions

D Frequency Spectrum (DFS)

The standard D-statistic averages signal across all allele frequencies, potentially obscuring important biological information. The D Frequency Spectrum (DFS) extension partitions the signal according to the frequency of derived alleles in P1 and P2, providing insights into the timing and history of introgression [29].

Diagram: D Frequency Spectrum (DFS) Concept

Recent gene flow typically produces a strong DFS peak among low-frequency derived alleles, while ancient introgression shows more dispersed signals across frequency bins as introgressed alleles have had time to drift to higher frequencies [29]. This distinction helps discriminate true introgression from artifacts caused by ancestral population structure, which tends to produce signals biased toward higher frequency bins [29].

Related Statistical Framework

Beyond the basic D-statistic, several related statistics provide additional insights:

• f4-ratio: Estimates the proportion of admixture in a population [24]
• f-branch: Assigns gene flow evidence to specific branches of a phylogeny, useful for interpreting results across many populations [24]
• fd and fdM: Window-based statistics designed to identify specific introgressed loci [24]

These statistics can be implemented separately or as part of integrated toolkits like Dsuite, which calculates them efficiently across all combinations of populations in large datasets [24].

Performance and Sensitivity Analysis

Factors Influencing Detection Accuracy

The performance of the D-statistic depends on several biological and methodological factors. The table below summarizes key sensitivity considerations based on empirical and simulation studies:

Table 2: Sensitivity Analysis of D-statistic Performance

Factor	Impact on D-statistic	Optimal Conditions	Potential Pitfalls
Population Size	High sensitivity; larger populations increase ILS, diluting signal	Smaller populations relative to divergence time	High false negatives with large populations
Divergence Time	Robust across wide range of genetic distances	Recent to moderate divergence (0.3-5% sequence divergence)	Saturation effects at high divergence
Gene Flow Timing	Strongly affects magnitude and direction of D	Detectable for events occurring after P1-P2 split	Very ancient gene flow may be missed
Rate Variation	High false positive rate with lineage-specific rate variation	Molecular clock assumption holds	>17% rate difference causes 35% FPR; >33% causes 100% FPR [27]
Outgroup Distance	Moderate impact; more distant outgroups increase multiple hits	Appropriately distant to polarize alleles	Very distant outgroups exacerbate rate variation artifacts [27]
Genomic Scale	Critical for statistical power; more loci reduce variance	Whole genomes or 1000s of independent loci	High variance with few loci; linkage effects

The D-statistic shows particular sensitivity to population size, as larger populations generate more incomplete lineage sorting, which can dilute the signal of introgression [26]. Perhaps most importantly, recent research has revealed that the method is highly sensitive to violations of the molecular clock assumption, with even moderate rate variation (17% difference) between sister lineages inflating false positive rates to 35%, and stronger rate variation (33% difference) causing 100% false positives in shallow phylogenies [27].

Comparison with Alternative Methods

Several alternative methods exist for detecting introgression, each with different strengths and limitations:

• HyDe: Similar site-pattern method designed specifically for detecting hybrid speciation [27]
• D₃: A three-sample test that uses genetic distances instead of an outgroup [25] [30]
• D_FOIL: Extension for five taxa providing more detailed directionality of introgression [30]
• D_GEN: Generalization of D-statistic principles to arbitrary numbers of taxa and complex introgression scenarios [30]

These methods complement the D-statistic, with choice depending on specific research questions, sampling design, and available data.

Research Reagent Solutions

Table 3: Essential Research Reagents and Computational Tools for D-statistic Analyses

Tool/Resource	Type	Primary Function	Key Features
Dsuite	Software Package	Comprehensive D-statistic analyses	VCF support; f-branch; fdM; efficient for large datasets [24]
ADMIXTOOLS	Software Package	Population admixture inference	Implements D, f4-ratio; established community use [24]
VCF Files	Data Format	Standardized variant calling output	Interoperability between variant callers and analysis tools [24]
Whole Genome Sequences	Data Type	Primary input data	Maximum statistical power for detection [29]
Reference Genome	Data Resource	Genomic coordinate system	Alignment and variant calling reference
msprime/slim	Simulation Tools	Demographic model testing	Validate interpretations under known parameters [29]

Biological Applications and Case Studies

The D-statistic has been successfully applied across diverse biological systems, providing insights into evolutionary history and species boundaries:

In hominin evolution, the method first revealed Neanderthal introgression into modern human populations outside Africa [24] [26]. In Lissotriton newts, Dsuite analyses revealed extensive introgression that had complicated previous phylogenetic estimates, particularly affecting the placement of L. montandoni within the L. vulgaris complex [31].

Even in bacterial systems, where introgression detection presents unique challenges, modified D-statistic approaches have quantified core genome introgression levels averaging 2% across 50 major lineages, reaching up to 14% in Escherichia-Shigella [28]. This demonstrates the method's versatility across biological domains, though bacterial applications require careful consideration of homologous recombination mechanisms rather than meiotic introgression.

Limitations and Best Practices

Despite its widespread utility, the D-statistic has important limitations that researchers must consider:

• False positives from rate variation: Lineage-specific substitution rate differences can create significant D-values without actual introgression [27]
• Sensitivity to ancestral structure: Population structure in ancestral populations can generate signals mimicking introgression [29]
• Dependence on correct topology: Incorrect species tree estimation can lead to erroneous conclusions [24]
• Multiple hits at high divergence: Saturation effects can distort site pattern counts in deeply divergent taxa [26]

Best practices to address these limitations include:

• Testing for rate variation between lineages before interpretation
• Applying DFS to distinguish recent introgression from ancestral structure
• Using complementary methods (D₃, D_FOIL) to validate findings
• Ensuring sufficient genomic sampling to achieve statistical power while minimizing linkage effects
• Acknowledging that D-statistic significance indicates gene flow but does not definitively localize it temporally or directionally without additional evidence

When applied with appropriate caution and in combination with complementary methods, the D-statistic remains a powerful tool for characterizing phylogenetic networks and detecting historical introgression, contributing significantly to our understanding of evolutionary complexity across the tree of life.

Probabilistic Modeling Under the Multispecies Network Coalescent

The Multispecies Coalescent (MSC) model represents a foundational framework in modern phylogenomics, describing the genealogical relationships of DNA sequences sampled from multiple species and accounting for the natural discordance between individual gene trees and the broader species phylogeny caused by incomplete lineage sorting (ILS) [32]. As the study of genome evolution has advanced, recognizing the pervasive role of hybridization, introgression, and other reticulate processes, the MSC framework has been formally extended to the Multispecies Network Coalescent (MSNC). This model provides a powerful probabilistic foundation for inferring phylogenetic networks, which represent evolutionary histories containing both divergent (tree-like) and reticulate events [33]. Accurately characterizing introgression is particularly critical in fields such as drug development, where understanding the evolutionary origins of pathogen virulence or host immune factors can inform target identification. This guide objectively compares the performance, underlying assumptions, and experimental applications of leading probabilistic models and inference methods based on the MSNC paradigm.

Methodological Framework: From Theory to Implementation

Core Model Assumptions and Statistical Foundations

The standard MSC model operates on a known species phylogeny, assuming complete isolation after species divergence with no migration, hybridization, or introgression [32]. It further assumes no recombination within loci, meaning all sites in a locus share an identical gene tree topology and coalescent history. The model parameters typically include species divergence times (τ) and population size parameters (θ), which are proportional to the effective population size [32].

The MSNC expands this framework to phylogenetic networks, which are rooted, directed acyclic graphs where nodes with multiple incoming edges represent reticulation events. The MSNC simultaneously addresses two confounded sources of gene tree incongruence: reticulations in the network and ILS [33]. Model implementations can be broadly categorized into two paradigms:

• Full-Likelihood Methods: These seek to compute the probability of sequence data given a network topology and parameters, often using Bayesian approaches. While potentially more accurate, they are often limited to very small datasets due to computational intensity [33].
• Two-Stage Methods: These first infer individual gene trees from sequence data and then treat these trees as the data for network inference. This approach offers significantly better scalability and is the basis for many contemporary methods [6] [33].

Key Computational Methods and Algorithms

Table 1: Comparison of Selected MSNC Inference Methods

Method	Inference Type	Core Approach	Key Assumptions	Scalability (Taxa/Genes)
ALTS [6]	Parsimony-based Network Inference	Infers the minimum tree-child network displaying all input gene trees by aligning lineage taxon strings.	Assumes gene trees are known and the network is tree-child.	Up to 50 taxa, 50 trees (cited runtime: ~15 minutes)
NANUQ/NANUQ+ [33]	Distance-based/Two-Stage	Uses quartet distances from gene trees to infer level-1 networks; a divide-and-conquer approach.	Assumes a level-1 network structure for full resolution.	Suitable for larger datasets due to distance-based approach.
Bayesian Full-Likelihood (e.g., StarBEAST2) [34]	Full-Likelihood (Bayesian)	Co-estimates gene trees and the species network from sequence data under the MSC.	Model assumptions can be relaxed; robust to some recombination.	Limited to smaller datasets by computational cost.
diCal2 [34]	Approximation-based	Uses sequentially Markovian approximations of the coalescent with recombination.	Designed explicitly to model recombination and linkage.	Designed for whole-genome data.

Performance Comparison: Accuracy in Inferring Reticulation

Experimental Protocols for Method Validation

Performance evaluations typically rely on simulations where the true evolutionary history—including introgression events—is known. A standard protocol involves:

• Network and Parameter Specification: A true metric phylogenetic network (e.g., an n-sunlet) is defined, including branch lengths in coalescent units and hybridization probabilities for reticulate edges [33].
• Sequence Simulation: Genomic sequence data (e.g., whole genomes or multiple loci) is simulated under the MSNC model along the specified network. Tools like msprime can simulate data under the coalescent with recombination [34]. Key varied parameters include:
  • Mutation and recombination rates.
  • Species divergence times and population sizes.
  • The timing and strength of introgression events.
• Method Application: The simulated sequence data is analyzed by the methods under comparison (e.g., ALTS, NANUQ+, StarBEAST2, diCal2).
• Accuracy Assessment: The inferred networks are compared to the true simulated network using metrics such as:
  • Topological Accuracy: Correctness of the inferred network topology, including the placement of reticulation nodes.
  • Parameter Estimation Error: Difference between inferred and true values for parameters like divergence times (τ) and population sizes (θ) [34].

Comparative Performance Data

Table 2: Summary of Comparative Performance from Simulation Studies

Method	Introgression Detection Accuracy	Divergence Time & Population Size Estimation	Robustness to Model Violations (e.g., Recombination)
Two-Stage Methods (e.g., ALTS, NANUQ)	Generally high for level-1 networks [6] [33]. Accuracy can be quantified by support values for inferred cycles [33].	Not primarily designed for this; typically focus on topology.	Dependent on initial gene tree estimation accuracy.
Bayesian MSC (e.g., StarBEAST2)	Effective, but computationally limited to small networks.	Accurate and provides credible intervals [34].	Surprisingly robust to realistic rates of intra-locus recombination [34].
SNAPP	Not designed for network inference; infers species trees.	Accurate for population parameters when using unlinked SNPs [34].	Unaffected by recombination as it uses single sites [34].
diCal2 (MSC with recombination)	Designed for such scenarios, but see performance notes.	Can produce "wildly erroneous" parameter estimates despite the model [34].	Performance issues likely due to algorithmic approximations [34].

A critical finding from recent research is that methods like StarBEAST2 and SNAPP, which do not explicitly model recombination in their standard form, show remarkable robustness to its presence, performing well with realistic recombination rates [34]. Conversely, diCal2, which was explicitly designed under the multispecies coalescent with recombination (MSC-R), performed considerably worse in comparative tests, yielding inaccurate parameter estimates [34]. This suggests that the specific algorithmic implementations and approximations can be more impactful than the conceptual scope of the underlying model.

Workflow and Signaling Pathways

The following diagram illustrates a generalized analytical workflow for inferring phylogenetic networks under the Multispecies Network Coalescent, integrating the various methods discussed.

Table 3: Key Software and Data Resources for MSNC Research

Resource Name	Type	Primary Function	Relevance to Introgression Studies
msprime [34]	Simulation Software	Simulates genomic sequence data under the coalescent model, including recombination and complex demography.	Generating benchmark datasets with known introgression events for method testing and validation.
MSCquartets [33]	R Software Package	Implements the NANUQ and NANUQ+ algorithms for phylogenetic network inference from quartet counts.	Quantifying support for reticulate edges and resolving ambiguous cycle structures in level-1 networks.
ALTS [6]	Standalone Software	Infers a tree-child phylogenetic network that displays a set of input gene trees with minimal reticulations.	Parsimonious inference of network topology from pre-estimated gene trees, scalable to dozens of taxa.
StarBEAST2 [34]	BEAST2 Package	Co-estimates gene trees and species trees/networks from sequence data using Bayesian MCMC under the MSC.	Joint inference of topology, divergence times, and population sizes with measures of statistical uncertainty.
SAMtools/BCFtools [34]	Data Processing Tools	Handles and processes high-throughput sequencing data, including variant calling.	Preparing genomic sequence alignments from raw sequencing reads for downstream phylogenetic analysis.

Probabilistic modeling under the Multispecies Network Coalescent provides an essential and evolving toolkit for characterizing introgression. Current performance comparisons reveal a landscape where no single method dominates all others; rather, the choice involves critical trade-offs between scalability, biological realism, and statistical certainty. Two-stage methods like NANUQ+ and ALTS offer practical scalability for initial topological inference, while Bayesian methods like StarBEAST2 provide robust parameter estimation with quantified uncertainty, albeit for smaller datasets. A surprising yet crucial insight is that model violation robustness can be more dependent on stable algorithmic implementation than on theoretical model completeness, as evidenced by the poor performance of diCal2 relative to simpler models [34].

Future progress hinges on several fronts: developing more efficient full-likelihood algorithms, creating models that integrate broader biological processes like selection, and refining divide-and-conquer strategies [33] to tackle networks of higher complexity. For researchers in phylogenomics and drug development, this means that selecting a method must be a deliberate choice aligned with the specific biological question, data characteristics, and required output—whether it is a full network topology, precise divergence times, or the statistical confidence in a hypothesized introgression event.

Phylogenetic networks are essential for representing evolutionary histories that involve reticulate events such as hybridization, introgression, and horizontal gene transfer. For researchers characterizing introgression, selecting the appropriate inference tool is critical, as it directly impacts the accuracy and biological interpretability of the results. This guide objectively compares three prominent tools—PhyloNet, SNaQ, and MP—based on experimental performance data, providing a foundation for informed methodological selection.

At a Glance: Tool Comparison

The following table summarizes the key characteristics and empirical performance of the three phylogenetic network inference tools.

Tool	Inference Type	Underlying Criterion / Model	Key Strengths	Scalability (Taxa)	Computational Limitations
PhyloNet (MLE)	Probabilistic	Maximum Likelihood under coalescent model [35] [36]	High topological accuracy [35] [36]	< 25 [35] [36]	Prohibitive runtime/memory beyond ~25 taxa [35] [36]
SNaQ (MPL)	Probabilistic	Pseudo-likelihood under coalescent model [35] [36]	Good accuracy, more efficient than full likelihood [35] [36]	< 50 (theoretical, but slower with scale) [6]	Performance degrades with increased taxa/divergence [35] [36]
MP (Parsimony)	Parsony-based	Minimize Deep Coalescence (MDC) criterion [35] [36]	Computationally faster than probabilistic methods [35] [36]	Larger datasets [35] [36]	Lower topological accuracy compared to probabilistic methods [35] [36]

Experimental Performance and Scalability

Quantitative data from controlled scalability studies reveal critical trade-offs between accuracy and computational efficiency.

Experimental Protocol for Scalability Assessment

The key findings in this guide are primarily derived from a systematic scalability study that evaluated multiple phylogenetic network inference methods [35] [36]. The general protocol is as follows:

• Input Data: Both simulated and empirical datasets (e.g., from natural mouse populations) were used.
• Model Phylogenies: Simulations were based on model phylogenies with a single reticulation event to establish a known ground truth [35] [36].
• Performance Metrics: Researchers measured topological accuracy (how well the inferred network matches the true model), CPU runtime, and main memory usage [35] [36].
• Scale Variations: The study quantified performance across two scalability dimensions: the number of taxa and the evolutionary divergence (sequence mutation rate) between taxa [35] [36].

Quantitative Performance Data

Tool	Runtime (Approx.)	Memory Usage	Topological Accuracy
PhyloNet (MLE)	Failed to complete on datasets with ≥30 taxa after many weeks of CPU runtime [35] [36].	Becomes prohibitive beyond ~25 taxa [35] [36].	Most accurate among methods tested [35] [36].
SNaQ	More efficient than full-likelihood methods, but runtime increases with dataset size [35] [36].	Not explicitly reported, but generally more scalable than MLE.	Generally high, though slightly lower than full-likelihood methods in some scenarios [35] [36].
MP	Fastest among the methods compared [35] [36].	Lower computational demands.	Lower accuracy compared to probabilistic methods [35] [36].

General Workflow for Phylogenetic Network Inference

The following diagram illustrates the typical workflow for inferring phylogenetic networks from genomic data, integrating the roles of tools like PhyloNet, SNaQ, and MP.

The Scientist's Toolkit: Key Research Reagents

Successful inference and characterization of introgression require a suite of computational and data resources.

Reagent / Resource	Function in Phylogenetic Network Inference
Multi-locus Sequence Data	Provides the raw genomic information from multiple independent loci, serving as the fundamental input for all subsequent analysis [35] [36].
Gene Trees	Phylogenetic trees inferred from individual loci; the primary input for summary methods like PhyloNet, SNaQ, and MP [35] [36].
Reference Genomes	High-quality genome assemblies used for accurate alignment and identification of orthologous loci, crucial for reliable gene tree estimation.
Coalescent Model	A population genetic model that forms the statistical foundation for probabilistic methods, accounting for incomplete lineage sorting (ILS) [14] [35].
Pseudo-likelihood Approximation	A computational strategy used by tools like SNaQ to approximate the full coalescent model likelihood, offering a balance between accuracy and speed [35] [36].

Key Insights for Researchers

• For highest accuracy on small datasets (<25 taxa), PhyloNet (MLE) is the preferred choice despite its computational cost [35] [36].
• For a balance of accuracy and efficiency on medium-sized datasets, SNaQ is a robust option, leveraging pseudo-likelihood for practical inference [35] [36].
• For initial exploration of very large datasets or when computational resources are limited, MP provides a fast, though less accurate, alternative [35] [36].
• Consider the biological question: If the focus is on deep introgression events, be aware that phylogenetic networks typically model episodic hybridization, and methods may be sensitive to violations of this assumption [14].

Choosing the right tool requires a careful balance between your specific research question, the scale and quality of your genomic data, and the computational resources at your disposal. This comparison provides a data-driven foundation for that critical decision.

In comparative biology, phylogenetic networks provide a powerful framework for modeling evolutionary histories that involve non-vertical transmission of genetic material. While phylogenetic trees represent evolutionary relationships as a strictly branching process with vertices having only a single parent, phylogenetic networks allow for multiple parent vertices, thereby representing complex evolutionary scenarios involving reticulation events such as hybridization, horizontal gene transfer, and introgression [37]. The distinction between softwired and hardwired networks represents a fundamental dichotomy in how these networks are interpreted and how their parsimony costs are calculated, with significant implications for their biological accuracy and application in evolutionary research [37].

The growing recognition of the importance of reticulate evolution in genome evolution has increased interest in phylogenetic networks across various fields of biology [6]. As researchers seek to characterize introgression events with greater accuracy, understanding the conceptual and computational distinctions between softwired and hardwired networks becomes essential for selecting appropriate methodologies and interpreting results correctly within the context of evolutionary biology research.

Conceptual Foundations and Definitions

Softwired Networks

Softwired networks interpret network edges as alternate pathways, only one of which is active for any given character in the evolutionary history. In this interpretive framework, each character follows a single ancestral path through the network, effectively behaving as if it evolved on one of the trees "displayed" by the network [37]. Biologically, this interpretation is particularly attractive as it aligns with scenarios where individual genetic elements have singular ancestral origins, even when the overall genome has multiple ancestral sources [37].

The parsimony score for a softwired network is calculated as the sum of the best possible scores for each character across all trees displayed by the network [37]. Formally, for a network N with a set of display trees τ(N) and a set of characters C, the softwired parsimony score is defined as:

[ S(N,C){score} = \Sigma{c \in C} \text{min }{\left(T \in \tau \left(N \right) \right) }T^{c}{score} ]

This approach allows different characters to follow different evolutionary paths within the same network, reflecting biological scenarios such as horizontal gene transfer in bacteria or hybrid origins in lineages where different genomic regions have distinct ancestral histories [37].

Hardwired Networks

Hardwired networks interpret all network edges as simultaneously active, with each character potentially being influenced by multiple ancestral pathways. In this model, network edges represent persistent connections that collectively contribute to the evolutionary history of all characters [37]. This interpretation generally proves less biologically realistic for most applications, as individual heritable characters typically have only one parent in evolutionary scenarios [37].

The parsimony cost calculation for hardwired networks sums changes across all edges in the network, resulting in costs that are necessarily greater than or equal to the best tree contained within the network [37]. Formally, the hardwired parsimony score is defined as:

[ H(N, C){score} = \Sigma{c \in C} \Sigma{e \in N} w{c} (e) ]

where w_c(e) represents the minimum number of character changes between vertex states that bound each edge e in the network N [37]. This comprehensive accounting across all edges often leads to overestimation of evolutionary change when applied to biological systems where characters typically follow singular ancestral paths.

Comparative Analysis: Key Distinctions

Table 1: Fundamental Differences Between Softwired and Hardwired Networks

Feature	Softwired Networks	Hardwired Networks
Biological Interpretation	Alternate edges represent different historical scenarios; only one active per character	All edges simultaneously active; characters influenced by multiple ancestors
Parsimony Cost Basis	Best tree for each character among those displayed by the network	Sum of changes across all edges in the network
Cost Relationship to Trees	Less than or equal to best display tree	Greater than or equal to best display tree
Biological Plausibility	High - reflects horizontal transfer and hybrid origin scenarios	Low - implies multiple ancestry for individual characters
Computational Complexity	Exponential in number of network nodes but polynomial for fixed parameters [37]	NP-hard but fixed-parameter tractable in parsimony score [37]
Optimality Testing	Requires penalty adjustment to compete equally with trees [37]	Naturally comparable but biologically less attractive [37]

Biological Interpretations and Applications

Softwired networks better accommodate the biological reality that while organisms may have complex ancestries involving multiple parental lineages, individual genetic characters typically trace their history through a single ancestral path [37]. This makes them particularly valuable for studying introgressive hybridization and horizontal gene transfer, where different genomic regions may have distinct phylogenetic histories due to selective processes or lineage-specific transfer events.

In bacterial evolution, for instance, softwired networks can model scenarios where individual genes have been horizontally transferred while the majority of the genome follows vertical inheritance [37]. Similarly, in plant evolution, softwired networks can represent hybrid speciation events where different genomic regions originate from different parental species.

Hardwired networks, while generally less biologically realistic for characterizing individual character evolution, may find application in modeling certain evolutionary scenarios such as reassortment networks in viruses or cases where persistent ancestral influences affect phenotypic traits [37]. However, their tendency to overestimate evolutionary change limits their utility for most empirical applications in introgression characterization.

Computational Considerations

The computational complexity of these network types differs significantly. Calculating softwired parsimony scores is exponential in the number of network nodes but becomes polynomial for non-additive characters when the number of reticulations is fixed [37]. In contrast, determining hardwired costs is NP-hard, though fixed-parameter tractable in the parsimony score when character states exceed two [37].

For softwired networks, a significant challenge lies in the potential for trivial optimization, where each character is assigned its best tree without penalizing network complexity [37]. To address this, researchers have proposed network edge penalties that account for the degree of "network-ness," enabling meaningful hypothesis testing between tree and network scenarios [37]. These penalties typically depend on the number of extra (non-tree) edges and are applied character-by-character, with networks containing superfluous edges assigned infinite cost to ensure identification of the minimum edge set required [37].

Experimental Protocols and Methodologies

Tree-Child Network Inference with ALTS

The ALTS program implements a scalable approach for inferring tree-child networks from multiple gene trees, addressing computational limitations that previously constrained phylogenetic network analysis [6]. Tree-child networks represent a specific class of phylogenetic networks in which every nonleaf node has at least one child that is not reticulate [6]. The methodology proceeds through these key steps:

• Input Processing: The algorithm takes as input a set of binary phylogenetic trees inferred from biomolecular sequences [6].

• Taxon Ordering: The method checks all possible orderings on the taxon set to identify tree-child networks with the smallest hybridization number [6].

• Lineage Taxon String (LTS) Calculation: For each taxon τ ≠ π₁ in ordering π, the algorithm computes the unique path from the root to the leaf representing τ in each input tree [6]. The LTS consists of the labels of internal nodes along this path.

• Common Supersequence Identification: For each taxon, the method identifies a common supersequence of all LTSs across input trees [6].

• Network Construction: Using the Tree-Child Network Construction algorithm, the program builds the network from the identified supersequences [6]:

Diagram 1: Tree-Child Network Construction Workflow. The ALTS program constructs networks by creating paths for each taxon and connecting them based on common supersequences of lineage taxon strings.

This approach enables inference of tree-child networks with large numbers of reticulations for sets of up to 50 phylogenetic trees with 50 taxa, significantly expanding the scale of phylogenetic network analysis possible within practical computational timeframes [6].

Network Parsimony Optimization

For accurate comparison between tree and network hypotheses, researchers must implement parsimony optimization methods that account for network complexity [37]. The experimental protocol involves:

• Character Optimization: For softwired networks, each character is optimized on its best display tree, while for hardwired networks, changes are summed across all network edges [37].

• Penalty Application: To enable meaningful hypothesis testing between trees and softwired networks, researchers must apply network edge penalties that increase with additional non-tree edges [37]. The penalty factor is typically derived as approximately half the expected cost of each edge for a tree with n leaves: T~cost~/(2n-2) [37].

• Statistical Testing: The penalized scores allow direct comparison between tree and network hypotheses, with the optimal representation (tree or network) determined by minimal penalized cost [37].

Quantitative Comparison and Performance Metrics

Table 2: Performance Metrics for Network Inference Methods

Metric	Softwired Networks	Hardwired Networks	Interpretation
Parsimony Score	Always shorter or equal to best tree	Always longer or equal to best tree	Softwired minimizes, hardwired maximizes character changes
Reticulation Detection	High accuracy for true horizontal transfers [37]	Overestimates reticulation events	Softwired better discriminates true signal from homoplasy
Computational Scalability	Handles ~50 trees with 50 taxa in ~15 minutes [6]	Limited to smaller datasets	Recent advances improve softwired scalability
Hypothesis Testing	Enabled with penalty adjustment [37]	Not directly comparable to trees	Softwired permits statistical comparison with trees
Biological Accuracy	High for most introgressive scenarios [37]	Low for character evolution	Softwired aligns with biological reality of character ancestry

Empirical validation studies demonstrate that when appropriate penalty adjustments are applied, softwired network costs correctly identify the simulated evolutionary scenario, outperforming both traditional trees and hardwired networks in accuracy [37]. The ALTS implementation for tree-child network inference successfully handles datasets of meaningful biological scale, processing 50 phylogenetic trees with 50 taxa in approximately 15 minutes on average [6].

Research Reagent Solutions

Table 3: Essential Computational Tools for Phylogenetic Network Analysis

Resource	Type	Function	Application Context
ALTS	Software program	Infers tree-child networks by aligning lineage taxon strings	Large-scale network inference from multiple gene trees [6]
HYBRIDIZATION NUMBER	Software program	Computes minimum hybridization number for two trees	Reticulation analysis for pairwise tree comparisons [6]
HYBROSCALE	Software suite	Infers phylogenetic networks from multiple input trees	General-purpose network inference [6]
PRIN/PRINs	Algorithms	Reconstructs tree-child networks with smallest hybridization number	Parsimonious network inference [6]
Tree-Child Networks	Network class	Ensures biological plausibility with at least one non-reticulate child per node	Foundation for biologically realistic network inference [6]
MCTS-CHN	Software program	Computes maximum consensus tree-child networks	Consensus network construction [6]

Signaling Pathways and Evolutionary Workflows

Diagram 2: Phylogenetic Network Inference Workflow. The research pipeline progresses from sequence data through tree inference to network reconstruction, with model selection based on biological versus theoretical considerations.

The comparative analysis of softwired versus hardwired networks reveals a clear superiority of softwired approaches for characterizing introgression and other reticulate evolutionary events. Softwired networks provide greater biological plausibility by respecting the fundamental principle that individual genetic characters typically follow singular ancestral paths, even when organisms have complex multiple ancestries [37]. The development of effective network penalty methods has enabled rigorous hypothesis testing between tree and network scenarios, addressing previous limitations in optimality-based comparisons [37].

Recent computational advances, particularly the ALTS program for tree-child network inference, have significantly enhanced the scalability and practical application of softwired networks in evolutionary research [6]. These methodological improvements, combined with the inherent biological realism of the softwired paradigm, establish softwired phylogenetic networks as the preferred approach for accurate characterization of introgression and other complex evolutionary phenomena in genomic research.

Machine Learning Applications in Genomic Introgression Detection

The precise characterization of introgression—the transfer of genetic material between species or populations through hybridization and backcrossing—is crucial for understanding adaptation, speciation, and evolutionary history. For years, phylogenetic methods, such as the analysis of gene tree topologies and summary statistics like the D-statistic (ABBA-BABA test), have been the cornerstone of introgression research [38]. However, these traditional approaches often rely on simplified models and can be confounded by complex evolutionary scenarios such as incomplete lineage sorting or recurrent mutation [38] [39]. The burgeoning field of machine learning (ML), particularly deep learning, is now revolutionizing this domain by offering powerful new frameworks for detecting introgressed alleles with unprecedented accuracy and resolution [23] [40] [41]. These methods excel at identifying complex, non-linear patterns in genomic data that are often imperceptible to conventional statistics. This guide provides a comparative analysis of emerging machine learning tools, benchmarking their performance against traditional and contemporary alternatives, and details the experimental protocols essential for their application. The overarching thesis is that while phylogenetic networks provide the essential evolutionary context, machine learning methods are achieving superior accuracy for the precise characterization of introgressed genomic segments.

The detection of introgression has evolved through distinct methodological phases, from summary statistics to probabilistic modeling, and now to supervised machine learning.

Traditional Phylogenetic and Summary Statistics: Methods like the ABBA-BABA test (D-statistic) operate by comparing the frequencies of discordant tree topologies to infer historical gene flow. While highly useful, they assume identical substitution rates and an absence of homoplasy, assumptions that can be violated in analyses of divergent species, potentially leading to misleading results [38]. Other established approaches rely on genome scans using metrics like FST and dxy to identify outliers, but these can struggle to distinguish introgression from other evolutionary forces like selective sweeps without additional context [39].

Probabilistic Modeling: This approach explicitly incorporates evolutionary processes into a model-based framework, using methods such as Approximate Bayesian Computation (ABC) to compare simulated data to real observations. ABC improves upon simple statistics by integrating multiple aspects of genetic variation, though it can remain computationally intensive [40].

Supervised Machine Learning (SML): This represents the current frontier. SML methods train algorithms on vast amounts of simulated genomic data to recognize the unique signatures of introgression. The most powerful implementations use deep neural networks that treat genomic alignments as images, learning to identify spatial patterns indicative of gene flow [23] [41] [3]. These can be further categorized into:

• Convolutional Neural Networks (CNNs): Used for classifying whole genomic regions as introgressed or not [3].
• Graph Convolutional Networks (GCNs): Applied to tree sequences—efficient representations of ancestral relationships—to perform various inference tasks, including introgression detection [40].
• Semantic Segmentation Networks (e.g., IntroUNET): Adapted to perform a pixel-level classification task, pinpointing the exact location of introgressed alleles within individual genomes [41].

Comparative Performance Analysis of Introgression Detection Methods

The following table summarizes the performance of various methods as reported in benchmarking studies, providing a direct comparison of their capabilities.

Table 1: Performance comparison of methods for detecting introgression

Method Name	Category	Reported Accuracy/Performance	Key Strengths	Notable Limitations
ABBA-BABA (D-statistic) [38]	Summary Statistic	Robust in recently diverged species	Simple, fast to compute; provides a test for introgression.	Assumptions can be violated in divergent species; cannot pinpoint specific introgressed haplotypes.
kNN-based Genome Scans [39]	Unsupervised ML	High accuracy in simulations, outperforming some state-of-the-art methods	Versatile for both selection and introgression; less confounded by population history.	Performance depends on feature selection (e.g., FST, dxy).
Graph Convolutional Networks (GCNs) [40]	Supervised ML (Deep Learning)	Slightly improved accuracy over traditional alignment-based CNNs	Uses efficient tree sequences; performs well across multiple inference tasks (demography, selection, introgression).	Requires estimation of tree sequences, which can introduce errors.
genomatnn (CNN) [3]	Supervised ML (Deep Learning)	95% accuracy on simulated data; >88% precision for adaptive introgression	Effective on both phased and unphased data; robust to heterosis; can visualize salient features.	Requires data from donor, recipient, and an outgroup population.
IntroUNET [41]	Supervised ML (Deep Learning)	Highly accurate at pinpointing introgressed alleles in individuals	Unprecedented resolution to identify introgressed alleles in specific individuals; can handle "ghost" introgression.	Computationally intensive; requires a large training set of simulations.

A key finding across multiple studies is that machine learning approaches, particularly deep learning, consistently match or exceed the accuracy of traditional methods. For instance, GCNs applied to tree sequences have been shown to perform "comparably or even better than traditional methods that used genetic alignments" for tasks like introgression detection [40]. Similarly, the CNN framework of genomatnn achieves high precision even with unphased data, a common challenge in real-world datasets [3]. The most significant advance, however, is in resolution. While summary statistics and even some CNNs can identify a genomic region as introgressed, tools like IntroUNET move beyond this by performing "semantic segmentation," thereby inferring "precisely which individuals have introgressed material and at which positions in the genome" [41]. This allows researchers to study the frequency and distribution of introgressed alleles, which is vital for understanding their fitness effects.

Experimental Protocols for Key Machine Learning Methods

Implementing ML-based introgression detection requires a structured workflow centered on data simulation, model training, and application. Below is a generalized protocol, with specifics for two leading tools.

General Workflow for Supervised ML

The power of supervised ML models comes from learning the patterns of introgression from data where the "answer" is known. This is achieved through a standardized workflow.

Specific Protocol for genomatnn (CNN)

genomatnn is designed to detect adaptive introgression from a donor population into a recipient population.

• Data Requirements: Collect whole-genome sequence data from the recipient population, the donor population (or a proxy), and a closely related unadmixed outgroup population [3].
• Data Simulation for Training:
  • Use a forward-time simulator like SLiM integrated with the stdpopsim framework to generate thousands of genomic windows under two evolutionary models:
    • Neutral Model: May include demographic history (e.g., population splits, bottlenecks) but no selection or introgression.
    • Adaptive Introgression Model: Incorporates a pulse of admixture from the donor into the recipient population, with positive selection acting on the introgressed haplotype after gene flow. Parameters like selection strength, time of admixture, and time of selection onset should be varied broadly.
• Input Data Preparation:
  • Partition simulated and empirical genomes into windows (e.g., 100 kbp).
  • For each window, create a genotype matrix. Each row is a haplotype (or diploid genotype if unphased), and columns correspond to bins of segregating sites.
  • Sort haplotypes within each population by similarity to the donor population and concatenate the matrices from the three populations (donor, recipient, outgroup) into a single input matrix [3].
• Model Training & Application:
  • Train a Convolutional Neural Network (CNN) using the simulated matrices and their known labels (introgressed vs. neutral).
  • The trained model outputs a probability score for adaptive introgression for each empirical genomic window.

Specific Protocol for IntroUNET

IntroUNET is designed for fine-scale mapping of introgressed haplotypes within individuals.

• Data Requirements: Genomic data from two populations (P1 and P2) between which gene flow is suspected. An outgroup can be helpful but is not always required [41].
• Data Simulation for Training:
  • Simulate genomic sequences under models with and without introgression between P1 and P2.
  • Crucially, the training data must be annotated at the individual allele level, meaning the true status of every allele (introgressed or not) is known for the simulation.
• Input Data Preparation:
  • Convert genotype data (e.g., VCF files) into images, where each pixel represents the genotype of an individual at a specific SNP.
• Model Training & Application:
  • Train a U-Net architecture (a type of CNN for semantic segmentation) on the simulated genotype images.
  • The model learns to produce a segmentation mask, classifying each "pixel" (i.e., each allele in each individual) as introgressed or not.
  • Apply the trained model to empirical data to generate a base-pair resolution map of introgression across all sampled individuals [41].

Successfully implementing these advanced genomic analyses requires a suite of software tools and resources.

Table 2: Essential resources for ML-based introgression detection research

Category	Resource Name	Primary Function	Relevance to Introgression Detection
Simulation Software	SLiM	Forward-time, individual-based genetic simulation	Gold standard for simulating complex evolutionary scenarios with selection and introgression [3].
	msprime / stdpopsim	Coalescent-based simulation and standardized population genetic models	Rapid simulation of neutral and demographic histories; often integrated with ML pipelines [3].
Machine Learning Frameworks	TensorFlow / PyTorch	Libraries for building and training deep learning models	Used to construct CNNs, GCNs, and U-Nets for introgression detection [41] [3].
Specialized Software	IntroUNET	Deep learning for identifying introgressed alleles in individuals	Provides fine-scale mapping of introgressed haplotypes from genomic data [41].
	genomatnn	CNN-based detection of adaptive introgression	Classifies genomic regions as under adaptive introgression using data from three populations [3].
	PhyloNet	Inference of species networks and introgression from gene trees	A traditional but powerful tool for quantifying introgression in a phylogenetic context [38].
Data Structures	Tree Sequences	Efficient encoding of ancestral relationships and genomes	Serves as a compact input for GCNs, improving computational efficiency and inference accuracy [40].

The integration of machine learning, particularly deep learning, into the detection of genomic introgression marks a significant paradigm shift. While traditional phylogenetic methods like tree-based topology comparisons and the D-statistic remain valuable for establishing an evolutionary framework, the empirical data demonstrates that ML tools are achieving superior accuracy and, critically, a higher resolution of inference. The ability of methods like IntroUNET to move from regional analysis to pinpointing introgressed alleles within individuals opens new avenues for studying the frequency and selective impact of introgressed material. As these tools become more accessible and are applied beyond model organisms, they will profoundly deepen our understanding of how gene flow shapes biodiversity, facilitates adaptation, and influences the very boundaries between species. Future progress will hinge on the development of more standardized benchmarking datasets and the creation of user-friendly pipelines that make these powerful technologies available to a broader community of evolutionary biologists.

The accurate reconstruction of evolutionary histories is fundamental to understanding biodiversity. For decades, the tree-like model of speciation dominated phylogenetic studies. However, mounting genomic evidence reveals that the evolutionary history of many taxa, including the Asian warty newts (genus Paramesotriton), is better represented by a network due to the process of introgression—the integration of genetic material from one species into another through hybridization [42]. This case study examines how erosion-mediated radiation has shaped the evolutionary trajectory of Asian warty newts and evaluates the accuracy of phylogenetic network methods in characterizing the resulting complex patterns of introgression. As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health.

The genus Paramesotriton represents the second most diverse genus in the family Salamandridae, currently containing 14 recognized species distributed from northern Vietnam to southwest-central and southern China [43]. These amphibians have undergone a complex evolutionary history influenced by paleogeological events and climatic oscillations, making them an ideal system for studying erosion-mediated radiation and testing the efficacy of different phylogenetic approaches for introgression detection.

Comparative Analysis of Introgression Detection Methods

Researchers employ multiple computational frameworks to detect introgression, each with distinct theoretical foundations, data requirements, and analytical outputs. Understanding these differences is crucial for selecting the appropriate tool for specific research contexts and accurately interpreting the resulting phylogenetic patterns.

Table 1: Comparative Analysis of Introgression Detection Methodologies

Method	Theoretical Basis	Data Requirements	Key Outputs	Handles ILS?	Citation
PhyloNet-HMM	Phylogenetic networks + Hidden Markov Models	Whole-genome alignment, Parental species trees	Probability of introgression per site, Introgressed regions mapping	Yes	[42]
Tree-based Topology Frequency	Asymmetry in phylogenetic tree topologies	Gene trees from across the genome	Species tree, Support for alternative phylogenetic hypotheses	Yes	[38]
ABBA-BABA (D-statistic)	Site pattern frequencies	Genome-wide SNP data	D-statistic value, Z-score for introgression test	No	[38]
Population Genetics Approaches	Allele frequency spectra, Genetic clustering	Genome-wide SNP data, Population samples	Population structure, Gene flow estimates, F-statistics	Indirectly	[44]

Performance Metrics and Accuracy Considerations

Each method exhibits distinct strengths and limitations in accurately characterizing introgression, particularly when confronted with evolutionary challenges like incomplete lineage sorting (ILS).

Table 2: Accuracy Assessment and Performance Considerations

Method	Strengths	Limitations	False Positive Risk	Resolution	Computational Intensity
PhyloNet-HMM	Handles ILS and introgression simultaneously; Accounts for dependencies across loci	Requires predefined parental species trees; Complex model parameterization	Low when model assumptions are met	Nucleotide site level	High
Tree-based Topology Frequency	Robust to conditions misleading ABBA-BABA test; Intuitive interpretation	Requires high-quality gene trees; Filtering for recombination needed	Moderate; depends on gene tree accuracy	Gene tree level	Moderate to High
ABBA-BABA (D-statistic)	Simple implementation; Fast computation; No need for species tree	Assumes identical substitution rates; Ignores homoplasy; Problematic for divergent species	High for divergent species or with homoplasy	Genome-wide average	Low
Population Genetics Approaches	Provides additional population context; Estimates direction and magnitude of gene flow	May not distinguish introgression from other gene flow types; Population sampling sensitive	Moderate; confounded by shared ancestral polymorphism	Population level	Moderate

PhyloNet-HMM represents a significant advancement in introgression detection by combining phylogenetic networks with hidden Markov models (HMMs) to simultaneously capture reticulate evolutionary history and genomic dependencies [42]. This framework can distinguish true introgression signatures from spurious ones arising from population effects like ILS, which occurs when species diverge with insufficient time for complete lineage sorting, creating incongruent genealogies across loci that can mimic introgression patterns [42]. In contrast, tree-based methods analyzing gene tree topologies across the genome provide a complementary approach that is less sensitive to certain assumptions that can mislead SNP-based methods like the D-statistic [38].

Experimental Protocols for Introgression Detection

Genomic Data Acquisition and Processing for Asian Warty Newts

The foundation of accurate introgression characterization begins with robust genomic data collection. For Asian warty newts, this involves:

* Tissue Sampling and DNA Extraction * Researchers collected tissue samples (1×1 mm from tails) from 62 live newts across 11 locations in northern Vietnam during field surveys [44]. Specimens were located by walking upstream along streams in evergreen forests with closed canopies, searching in zig-zag patterns around streams, and examining microhabitats under surface objects like rocks, leaves, and wood [44]. After morphological measurements and geographical data recording, total DNA was extracted from muscle samples using a DNeasy Blood & Tissue kit (Qiagen, Hilden, Germany) following manufacturer protocols [44].

* Genome-Wide SNP Genotyping * For population genetic analysis, researchers generated genome-wide single-nucleotide polymorphism (SNP) data using the multiplexed inter-simple sequence repeat genotyping (MIG-seq) method [44]. This approach involves:

• DNA library preparation following established protocols [44]
• Pooling indexed DNA libraries
• Sequencing on HiSeq platforms (Illumina, San Diego, CA, USA)
• Quality control processing of MIG-seq data
• Deposition of raw sequence reads in the DNA Data Bank of Japan Sequence Read Archive

* Whole-Genome Alignment for Phylogenetic Analysis * For tree-based introgression detection, whole-genome alignment forms the analytical foundation:

• Alignment preparation using Progressive Cactus for reference-free whole-genome alignment [38]
• Optional conversion to MAF format using hal2maf tool for human-readable, reference-based alignment [38]
• Extraction of suitable alignment blocks (typically 1,000 bp length) while filtering for information content and minimal recombination signals [38]
• Identification of alignment blocks with sequences for all included species

Phylogenetic Network Construction and Introgression Detection

* PhyloNet-HMM Implementation Protocol * The PhyloNet-HMM framework implements a sophisticated methodology for introgression detection [42]:

• Model Definition: Let χ be a set of aligned genomes {X1, X2, ..., Xn}, and let site i in the alignment be denoted as χ[i]. The model defines a set of random variables Ψi, each taking values in the set of parental species trees Θ [42].

• Problem Formulation: For each site i, compute the probability P(Ψi = S|χ) for every parental species tree S ∈ Θ [42].

• Training: Using dynamic programming algorithms paired with multivariate optimization heuristics to train the model on genomic data [42].

• Identification: Locating genomic regions of introgressive descent through analysis of parental species tree probabilities across genomic positions [42].

* Tree-Based Introgression Detection Workflow * The tree-based approach follows a structured pipeline [38]:

• Gene Tree Estimation:

  • Phylogenetic inference for each selected alignment block using maximum likelihood with IQ-TREE
  • Alternatively, using PAUP* for general-utility phylogenetic inference

• Species Tree Estimation:

  • Efficient and accurate species tree estimation from gene trees using ASTRAL

• Topology Analysis:

  • Determination of asymmetry among alternative phylogenetic topologies for species trios
  • Assessment of support for past introgression events based on topology frequencies

• Network Inference:

  • Analysis of gene trees with PhyloNet to assess support for alternative diversification models with and without introgression
  • Implementation in maximum-likelihood, Bayesian, or parsimony frameworks

The Asian Warty Newt System: Erosion-Mediated Radiation and Historical Biogeography

Phylogenetic Framework and Divergence History

Comprehensive phylogenetic analysis combining mitochondrial genomes and 32 nuclear genes from 27 samples representing 14 species has established the evolutionary framework for Asian warty newts [43]. Both Bayesian inference and maximum-likelihood analyses strongly support the monophyly of Paramesotriton and its two recognized species groups (P. caudopunctatus and P. chinensis groups) while identifying five hypothetical phylogenetic cryptic species [43].

Biogeographic analyses indicate that Paramesotriton originated in southwestern China (Yunnan-Guizhou Plateau/South China) during the late Oligocene, with the timing of origin corresponding to the second uplift of the Himalayan/Tibetan Plateau, rapid lateral extrusion of Indochina, and formation of karst landscapes in southwestern China [43]. This erosion-mediated radiation created the complex topography that facilitated genetic divergence and speciation through geographical isolation.

Principal component analysis, independent sample t-tests, and niche differentiation using bioclimatic variables based on locations of occurrence reveal that Paramesotriton habitat conditions in three major regions (West, South, and East) differ significantly, with different levels of climatic niche differentiation [43]. Species distribution model predictions indicate that the most suitable distribution areas for the P. caudopunctatus and P. chinensis species groups are western and southern/eastern areas of southern China, respectively [43].

Contemporary Patterns of Genetic Structure and Introgression

Population genetic analyses of Asian warty newts in northern Vietnam using genome-wide SNP data have revealed three primary genetic groups: West, East + Cao Bang (CB), and Quang Ninh (QN) [44]. The Cao Bang population exhibits discordance between mitochondrial DNA and single-nucleotide nuclear DNA polymorphism data, suggesting possible historical introgression events [44]. Furthermore, gene flow within populations is restricted, particularly within West and QN groups [44].

Spatial distribution analyses of genetic clusters conditioned by environmental variables predict that under climate change scenarios, the East + CB genetic cluster would expand, whereas West and QN clusters would decrease [44]. The introgression of genetic structures may reduce the vulnerability of East + CB to climate change, highlighting the potential adaptive significance of historical introgression events [44].

Ecological niche modeling reveals that these newts are susceptible to climate change, with projected reduction in suitable habitat areas across all scenarios and a shift in suitable distribution toward higher elevations [44]. The mountainous areas of northern Vietnam could serve as potential refugia for these newts as climate change intensifies, potentially influencing future patterns of introgression through range shifts and secondary contact.

Table 3: Research Reagent Solutions for Phylogenomic Analysis of Introgression

Category	Item	Specification/Version	Primary Function	Application Context
Wet Lab Supplies	DNeasy Blood & Tissue Kit	Qiagen	High-quality DNA extraction from tissue samples	Initial genomic DNA isolation [44]
	MIG-seq Library Prep Reagents	Protocol by Suyama & Matsuki (2015)	Genome-wide SNP genotyping	Multiplexed inter-simple sequence repeat genotyping [44]
Sequencing Platforms	Illumina HiSeq	Various models	High-throughput sequencing	Genome-wide SNP data generation [44]
Phylogenetic Software	PhyloNet	Version in PhyloNet distribution	Species tree and network inference	Maximum-likelihood, Bayesian, or parsimony framework analysis [38]
	IQ-TREE	IQ-TREE v.2	Maximum likelihood phylogenetic inference	Gene tree estimation from alignment blocks [38]
	PAUP*	Command-line version	General-utility phylogenetic inference	Alternative tree estimation approach [38]
	ASTRAL	v.5.7.8	Species tree estimation from gene trees	Efficient coalescent-based species tree inference [38]
Visualization Tools	PhyloScape	Web-based application	Interactive phylogenetic tree visualization	Customizable visualization with metadata annotation [45]
	FigTree	v.1.4.4	Phylogeny visualization and manipulation	Intuitive tree visualization and basic manipulation [38]
Analysis Frameworks	PhyloNet-HMM	Included in PhyloNet	Introgression detection in genomes	Comparative genomic framework combining networks and HMMs [42]
Data Resources	DNA Data Bank of Japan	DRA accession system	Raw sequence data archiving	Public repository for MIG-seq data [44]

The case of Asian warty newts demonstrates the critical importance of selecting appropriate methodological frameworks for accurately characterizing introgression in organisms affected by erosion-mediated radiation. PhyloNet-HMM provides a powerful solution for detecting introgression while accounting for incomplete lineage sorting and genomic dependencies, offering nucleotide-level resolution of introgressed regions [42]. Tree-based approaches using gene tree topologies provide complementary evidence that is robust to conditions that may mislead SNP-based methods like the D-statistic [38].

The evolutionary history of Paramesotriton reflects a complex interplay between geological events, climatic fluctuations, and possible introgression events. The phylogenetic framework reveals origins in the late Oligocene corresponding to major geological uplift events, with subsequent diversification influenced by Pleistocene climatic oscillations [43]. Contemporary patterns of genetic structure show three primary groups with evidence of mitonuclear discordance in the Cao Bang population, potentially indicating historical introgression [44].

For researchers investigating similar systems, the integration of multiple approaches—combining population genomic analyses with phylogenetic network methods—provides the most robust framework for accurately reconstructing evolutionary histories involving introgression. As climate change continues to alter species distributions and potentially create new opportunities for hybridization, these methodological considerations will become increasingly important for understanding and conserving biodiversity in rapidly changing environments.

Addressing Computational Challenges and Methodological Pitfalls in Network Inference

Scalability Limitations with Increasing Taxon Numbers

Phylogenetic networks are essential tools for modeling evolutionary histories that involve reticulate events such as introgression, hybridization, and lateral gene transfer. Accurate reconstruction of these networks is fundamental for introgression characterization research, with direct implications for understanding disease evolution and drug development. However, a significant challenge facing researchers is the scalability of phylogenetic network inference methods as the number of taxonomic units (taxa) increases. This guide objectively compares the performance of leading phylogenetic network inference methods, analyzing their scalability limitations and providing researchers with the experimental data needed to select appropriate tools for large-scale introgression studies.

Performance Comparison of Phylogenetic Network Methods

Key Inference Methods and Their Scalability

Different methodological approaches have been developed to infer phylogenetic networks, each with distinct computational characteristics and scalability profiles:

• Probabilistic Methods (MLE, MLE-length): These methods perform phylogenetic network inference under explicit evolutionary models that combine coalescent theory with biomolecular substitution models. They utilize full likelihood calculations but face significant computational constraints, typically becoming prohibitive beyond 25-30 taxa [36].

• Pseudo-likelihood Methods (MPL, SNaQ): These approaches substitute pseudo-likelihood approximations for full model likelihood calculations, improving computational efficiency while maintaining reasonable accuracy. SNaQ, for instance, combines pseudo-likelihoods under a coalescent-based model with quartet-based concordance analysis [36].

• Parsimony-Based Methods (MP): Earlier approaches utilizing the minimize deep coalescence (MDC) criterion seek species phylogenies that minimize the number of deep coalescences needed to explain a given set of gene trees [36].

• Concatenation Methods (Neighbor-Net, SplitsNet): These methods estimate a single phylogeny for all loci, typically accounting only for sequence mutation rather than more complex evolutionary processes [36].

Quantitative Performance Comparison

Table 1: Scalability and Performance Comparison of Phylogenetic Network Inference Methods

Method	Method Type	Max Practical Taxa	Runtime Performance	Accuracy Trend with Increasing Taxa
MLE	Probabilistic	25-30	Prohibitive (> weeks CPU)	Degrades substantially
MLE-length	Probabilistic	25-30	Prohibitive (> weeks CPU)	Degrades substantially
MPL	Pseudo-likelihood	>30	Moderate	Moderate degradation
SNaQ	Pseudo-likelihood	>30	Moderate	Moderate degradation
MP	Parsimony-based	>30	Faster than probabilistic	Degrades with increased mutation rate
Neighbor-Net	Concatenation	>50	Fast	Degrades with increased mutation rate
ALTS	Tree-child	50	~15 minutes for 50 taxa	Maintains accuracy with trivial clusters [6]

Table 2: Impact of Dataset Characteristics on Method Performance

Dataset Characteristic	Impact on Scalability	Effect on Accuracy
Number of taxa	Runtime increases polynomially/exponentially	Topological accuracy generally degrades as taxon count increases [36]
Evolutionary divergence	Higher mutation rates increase complexity	Accuracy degrades with increased sequence mutation rate [36]
Presence of rogue taxa	Increases computational demand for bootstrap methods	Substantially lowers bootstrap support throughout trees [46]
Nontrivial common clusters	Enables analysis of larger datasets	ALTS handles 50 taxa with trivial clusters in ~15 minutes [6]

Experimental Protocols for Scalability Assessment

Standard Scalability Evaluation Framework

Research studies evaluating phylogenetic network inference methods typically employ standardized experimental protocols to assess scalability and accuracy:

Simulation Design: Performance studies typically utilize model phylogenies with known reticulations (often single reticulation events for controlled analysis). Datasets are generated with varying taxon counts (from small to large-scale) and evolutionary divergence levels to systematically test method limitations [36].

Empirical Validation: Methods are additionally tested on empirical data sampled from natural populations, such as mouse populations, to verify performance on real biological datasets [36].

Accuracy Assessment: For simulated datasets where the true phylogeny is known, topological accuracy is measured by comparing inferred networks to the true model networks. For empirical data, accuracy is assessed through biological plausibility and consistency with known evolutionary relationships [36].

Large-Scale Network Inference Methodology

The ALTS program introduces a specific methodology for scalable tree-child network inference:

Input Processing: ALTS takes a set of binary phylogenetic trees on a taxon set X as input. The algorithm begins by considering all possible orderings (π) on the taxon set to obtain tree-child networks with the smallest hybridization number [6].

Lineage Taxon String Calculation: For each taxon τ ≠ π₁ in each input tree, the algorithm computes the Lineage Taxon String (LTS) - the sequence of internal node labels along the path from the root to leaf τ [6].

Common Supersequence Identification: For each taxon πᵢ, the method identifies a common supersequence βᵢ of all LTSs αⱼⁱ across all input trees Tⱼ [6].

Network Construction: Using the Tree-Child Network Construction algorithm, ALTS builds the network from the paths generated from each βᵢ, adding left-right edges between paths to create the final phylogenetic network with reticulation events [6].

Figure 1: ALTS Algorithm Workflow - The process for inferring tree-child networks from multiple phylogenetic trees

Method Classification and Computational Characteristics

Figure 2: Method Classification by Accuracy and Scalability - Relationship between inference methods and their performance characteristics

Table 3: Key Research Reagent Solutions for Phylogenetic Network Inference

Tool/Resource	Function	Applicable Context
PhyloNet	Software package implementing MLE and MLE-length methods	Probabilistic inference of phylogenetic networks [36]
SNaQ	Implementation of quartet-based pseudo-likelihood method	Species network inference applying quartets under coalescent model [36]
ALTS	Program for inferring tree-child networks by aligning lineage taxon strings	Large-scale network inference for datasets up to 50 taxa [6]
Neighbor-Net	Concatenation-based method for phylogenetic network inference	Rapid analysis of large datasets where computational efficiency is prioritized [36]
Multi-locus sequence data	Primary input data for network inference	Empirical studies requiring modeling of complex evolutionary processes [36]
Simulated phylogenies	Benchmarking and validation	Controlled evaluation of method performance with known ground truth [36]

Emerging Solutions and Future Directions

Innovative Approaches to Scalability Challenges

Recent methodological developments aim to address the critical scalability limitations of phylogenetic network inference:

SPRTA (Subtree Pruning and Regrafting-based Tree Assessment): This approach shifts from traditional topological assessment to evaluating evolutionary origins of lineages. SPRTA reduces runtime and memory demands by at least two orders of magnitude compared to existing methods, with the performance difference growing as dataset size increases [46].

Algorithmic Optimizations: New programs like ALTS demonstrate that innovative approaches such as reducing network inference to aligning lineage taxon strings can achieve significant speed improvements, processing 50 taxa with 50 trees in approximately 15 minutes for cases with only trivial common clusters [6].

Disjoint Tree Mergers (DTMs): This emerging class of divide-and-conquer methods operates by dividing input sequences into disjoint sets, constructing trees on each subset, then combining subset trees into a full phylogeny. When appropriately designed, pipelines using DTMs can maintain statistical consistency while improving accuracy and reducing runtime for very large datasets [9].

Critical Gaps and Research Needs

Despite these advances, significant methodological gaps remain. Current probabilistic methods become computationally prohibitive with datasets exceeding 25-30 taxa, creating a substantial disparity between methodological capabilities and the scale of contemporary phylogenomic studies [36]. New algorithmic development is critically needed to bridge this gap and enable accurate phylogenetic network inference at the scale of modern genomic datasets, particularly for introgression characterization research in disease evolution and drug development contexts.

Impact of Sequence Divergence and Evolutionary Distance on Accuracy

For researchers characterizing introgression, the accuracy of phylogenetic inference is not static but fluctuates significantly with levels of sequence divergence. An optimal range of sequence divergence exists for Bayesian phylogenetic reconstruction, outside of which accuracy substantially declines [47]. In response, novel methods leveraging structural correlation and machine learning now achieve high accuracy even at sequence identities below 20%, outperforming traditional sequence-based approaches for highly divergent sequences encountered in introgression studies [48] [49].

Quantitative Comparison of Phylogenetic Methods

Table 1: Performance Comparison of Phylogenetic Inference Methods

Method	Optimal Sequence Divergence Range	Accuracy at High Divergence (<20% identity)	Computational Demand	Key Innovation
SD Algorithm [48]	Effective down to 10% sequence identity	High (closely aligns with structural trees)	Low (single CPU, seconds for thousands of pairs)	Incorporates site-to-site correlation via PSSMs
Traditional Bayesian Methods [47]	Optimal range exists (scale depends on distance metric)	Poor performance outside optimal range	Moderate to High	Model-based evolutionary correction
FoldTree Structural Approach [49]	Superior for highly divergent families	Outperforms sequence-based ML on divergent datasets	Moderate (requires structural models)	Structural alphabet-based sequence alignment
Standard Barcode Sequences [50]	~600 bp suitable for species ID only	Inaccurate for phylogenetic reconstruction	Low	Short sequence efficiency

Table 2: Impact of Sequence Characteristics on Phylogenetic Accuracy

Sequence Characteristic	Effect on Phylogenetic Accuracy	Experimental Evidence
Among-Lineage Rate Variation [51]	Strong negative correlation with accuracy; gene trees with high rate variation are more dissimilar to species trees	Analysis of 30 phylogenomic datasets showing consistent pattern
Stemminess (internal:terminal branch length ratio) [51]	Low stemminess associated with poor topological signal	Observed across multiple taxonomic groups
Sequence Length [50]	Short sequences (~600 bp) sufficient for species ID but inadequate for phylogenetic relationships	Fungal mitochondrial sequence analysis
Evolutionary Distance [47]	Significant positive relationship between node support and genetic distance within optimal range	Bayesian analyses of 12 vertebrate genes

Detailed Experimental Protocols

Protocol 1: Sequence Distance (SD) Algorithm for Highly Divergent Sequences

The SD algorithm introduces a correlation-based approach specifically designed for analyzing remote homologues in protein superfamilies where traditional methods fail [48].

Input Features and Processing:

• PSSM Construction: PSI-BLAST with three-iteration search of Uniref90 database (E-value threshold: 0.001) generates 20-dimensional amino acid occurrence probabilities [48]
• Structural Features: SPIDER2 predicts secondary structure (H: α-helix, E: β-sheet, C: coiled) and solvent accessibility (B: buried with rASA <20%, E: exposed) [48]
• Feature Profile Construction: Cross product of amino acid occurrence probabilities at adjacent sites creates 640-dimensional vectors incorporating site correlations [48]

Alignment and Scoring:

• Implements modified Needleman-Wunsch global alignment with affine gap penalties [48]
• Scoring function: (S(i,j) = {M}{L1}(i)\cdot {M}{L2}(j) + {\omega }{1}SS(i,j) + {\omega }{2}rACC(i,j)) where:
  • ({M}{L1}) and ({M}{L2}) are feature profiles (640-dimensional vectors)
  • (SS(i,j)) = 1 if secondary structures match, 0 otherwise
  • (rACC(i,j)) = 1 if solvent accessibility states match, 0 otherwise
  • ({\omega }{1}) and ({\omega }{2}) are optimized weight coefficients (1.0-2.0 range) [48]

Validation Framework:

• Tested on SCOP2-based superfamily database with 14,108 proteins across 529 superfamilies [48]
• Performance evaluated against structure-based phylogenetic trees as reference [48]

Protocol 2: Determining Optimal Divergence Ranges for Bayesian Phylogenetics

This experimental approach establishes the relationship between sequence divergence and phylogenetic accuracy using both natural and simulated datasets [47].

Natural Dataset Construction:

• 28 well-supported vertebrate relationships from established phylogeny [47]
• 12 genes acquired including both mitochondrial and nuclear markers [47]
• Pairwise divergences calculated using multiple substitution models (K2P, etc.) [47]

Bayesian Analysis Protocol:

• Bayesian phylogenetic reconstruction for each gene dataset [47]
• Posterior probabilities calculated for correct relationships [47]
• Relationship between sequence divergence and nodal support quantified [47]

Simulation Framework:

• Datasets designed across extreme phylogenetic conditions [47]
• Various tree topologies and models of evolution tested [47]
• Optimal divergence ranges identified across different tree shapes [47]

Key Finding: An optimal range of sequence divergence exists for resolving correct relationships, though this range depends on the distance metric used [47].

Methodological Workflows

The Researcher's Toolkit

Table 3: Essential Research Reagents and Computational Tools

Resource	Type	Function in Phylogenetic Accuracy Research	Application Context
SPIDER2 [48]	Software Tool	Predicts secondary structure and solvent accessibility	Feature generation for SD algorithm
PSI-BLAST [48]	Algorithm	Generates position-specific scoring matrices (PSSMs)	Evolutionary feature extraction
Foldseek [49]	Structural Alignment Tool	Performs structural alignment using 3Di structural alphabet	Structural phylogenetics
SCOP2 Database [48]	Protein Database	Provides evolutionarily related protein superfamilies	Benchmarking and validation
CATH Database [49]	Structural Classification	Groups proteins by class, architecture, topology, homology	Testing on divergent families
Uniref90 Database [48]	Sequence Database	Non-redundant protein sequence database	PSSM construction
DAMBE Software [52]	Comprehensive Toolkit	Implements distance matrix imputation methods	Handling missing data
Matrix Factorization/Autoencoder [52]	Machine Learning Method	Imputes missing distances in incomplete matrices	Phylogenomics with missing data

Critical Insights for Introgression Characterization

For researchers focused on introgression characterization, these findings highlight several critical considerations:

Divergence Threshold Management: The existence of an optimal divergence range [47] suggests that introgression analyses should carefully select genomic regions based on divergence levels between candidate species. Regions within the optimal range will provide more reliable signal for detecting introgression events.

Structural Methods for Deep Introgression: When analyzing ancient introgression events where sequences have substantially diverged, structural phylogenetics approaches [48] [49] offer significant advantages over traditional sequence-based methods, potentially revealing older introgression events that would otherwise be undetectable.

Rate Variation as a Confounding Factor: The strong negative impact of among-lineage rate variation on phylogenetic accuracy [51] suggests that rate-heterogeneous genomic regions should be identified and treated carefully in introgression studies, as they may produce misleading signals.

The advancement of correlation-aware algorithms and structural approaches enables more accurate phylogenetic inference across wider evolutionary timescales, directly benefiting the resolution of complex introgression patterns in evolutionary genomics research.

Distinguishing True Introgression from ILS and Homoplasy

Accurately identifying introgression in evolutionary histories is complicated by confounding signals from incomplete lineage sorting (ILS) and homoplasy. This guide compares modern phylogenetic network methods that model the multispecies network coalescent (MSNC) against traditional approaches, highlighting how newer tools improve the characterization of gene flow. We summarize experimental data and provide protocols for employing these methods to distinguish true introgression from misleading signals.

Evolutionary histories are not always tree-like. Reticulate events such as hybridization and introgression create patterns of gene tree discordance that can be difficult to distinguish from those caused by ILS or homoplasy [14]. ILS occurs when ancestral genetic polymorphisms persist through multiple speciation events, leading to gene trees that differ from the species tree [53]. Homoplasy, the independent evolution of similar traits (or, for sequence data, identical mutations), can also create misleading signals of relatedness [53].

The multispecies coalescent (MSC) model provides a framework for understanding ILS, but it is the extension to the multispecies network coalescent (MSNC) that allows for simultaneous modeling of both ILS and introgression within a phylogenetic network [53] [54]. This is critical, as methods that account only for ILS can be conservative and miss true introgression events [53]. This guide objectively compares the performance of methodologies and software tools designed to characterize introgression by analyzing genome-scale data within the MSNC framework.

Methodological Comparison: Statistical Frameworks for Detection

Methodologies for detecting introgression can be broadly categorized into three groups: summary statistics, probabilistic models, and supervised learning approaches [23]. The table below compares their core principles, key implementations, and strengths and weaknesses.

Table 1: Comparison of Major Methodological Approaches for Detecting Introgression

Method Category	Core Principle	Example Methods/Implementations	Strengths	Weaknesses
Summary Statistics	Uses patterns in site frequencies or tree topologies to detect deviations from a null model of pure divergence/ILS.	Patterson's D-statistic (ABBA-BABA) [14]	Computationally fast; good for initial screening.	Limited to small taxon sets (e.g., 4-taxon test); does not provide a full network model; sensitive to model violations [14].
Probabilistic Modeling	Uses an explicit model of evolution (e.g., MSNC) to compute the probability of the data given a phylogenetic network.	SnappNet [54], MCMC_BiMarkers [54], HeIST [53]	Provides a powerful, model-based framework; can co-estimate species networks and gene trees; can distinguish ILS from introgression.	Computationally intensive; requires careful model specification.
Supervised Learning	Frames the detection of introgressed loci as a classification task, trained on data with known evolutionary histories.	Methods framed as semantic segmentation tasks [23]	Potential to handle complex evolutionary scenarios and large datasets.	Emerging approach; requires high-quality training data; "black box" interpretations can be a limitation.

Software Tool Performance: SnappNet vs. MCMC_BiMarkers vs. HeIST

A direct performance comparison between two Bayesian MSNC methods, SnappNet and MCMC_BiMarkers, reveals significant differences in scalability. HeIST addresses a different, but related, problem of trait evolution.

Table 2: Comparative Performance of Phylogenetic Network Software Tools

Software Tool	Core Function	Input Data	Inference Method	Key Performance Findings
SnappNet [54]	Infers phylogenetic networks under the MSNC.	Biallelic markers (e.g., SNPs).	Bayesian MCMC integrating over all possible gene trees.	Extremely faster than MCMC_BiMarkers on complex networks; more accurate on complex scenarios [54].
MCMC_BiMarkers [54]	Infers phylogenetic networks under the MSNC.	Biallelic markers (e.g., SNPs).	Bayesian MCMC, using a different likelihood computation algorithm.	Similar accuracy to SnappNet on simple networks; becomes significantly slower on complex networks [54].
HeIST [53]	Estimates probability of hemiplasy (trait incongruence due to gene tree discordance) vs. homoplasy.	Species tree/network, trait data, population parameters.	Coalescent simulation.	Accounts for both ILS and introgression; shows hemiplasy can explain apparent convergent evolution [53].

The experimental data supporting Table 2 comes from simulations cited in the SnappNet publication [54]. These benchmarks demonstrate that while both SnappNet and MCMC_BiMarkers can recover simple networks accurately, SnappNet's algorithms are exponentially more time-efficient for non-trivial networks. This speed advantage enables the analysis of more complex and biologically realistic evolutionary scenarios within a feasible computational time.

Experimental Protocols for Network Inference

For researchers aiming to characterize introgression, the following workflow provides a robust protocol using modern tools.

Data Preparation and Preprocessing

The foundation of any analysis is high-quality genomic data. For methods like SnappNet, the input is a set of biallelic markers, such as Single Nucleotide Polymorphisms (SNPs), from multiple individuals across the species of interest [54].

• Protocol: Assemble whole-genome sequences or targeted sequence capture data. Map reads to a reference genome and call variants using standard pipelines (e.g., GATK). Filter for high-quality, independently assorting biallelic SNPs. The final output is a matrix of alleles for each taxon and marker.

Initial Screening and Hypothesis Building

Before full model-based inference, use fast, summary statistic-based methods to screen for evidence of gene flow.

• Protocol: Apply Patterson's D-statistic to quartets of taxa to test for significant deviations from a tree-like history [14]. Significant D-statistic values indicate gene flow but do not specify the exact network. These results can be used to formulate an initial hypothesis and guide the choice of taxa for more computationally intensive network inference.

Full Network Inference with MSNC

This is the core model-based analysis that co-estimates the species network and introgression parameters.

• Protocol: Use a Bayesian MSNC method like SnappNet, implemented in BEAST 2 [54]. Configure an MCMC analysis to sample from the posterior distribution of phylogenetic networks. Key parameters to monitor include:
  • Network Topology: The branching order and placement of reticulation nodes.
  • Inheritance Probabilities (γ): For each reticulation node, the proportion of genetic material inherited from one parent, which can indicate symmetrical (γ ≈ 0.5) or asymmetrical introgression [14].
  • Divergence Times: Estimated in coalescent units.
• The analysis should run until MCMC convergence is achieved, assessed using effective sample size (ESS) diagnostics.

Success in phylogenetic network inference relies on a combination of software, data, and computational resources.

Table 3: Key Research Reagent Solutions for Introgression Studies

Tool / Resource	Function / Description	Relevance to Introgression Characterization
SnappNet	A BEAST 2 package for Bayesian inference of phylogenetic networks under the MSNC from biallelic data [54].	The primary software for co-estimating the species network and gene trees, directly distinguishing ILS from introgression.
HeIST	A simulation-based tool to estimate the probability of hemiplasy in the presence of ILS and introgression [53].	Used to assess whether observed trait incongruence is more likely due to hemiplasy on a discordant gene tree or true homoplasy.
Biallelic Marker Data	Genomic variants (typically SNPs) with two alleles across the studied taxa.	The fundamental input data for several MSNC methods, representing the genetic variation used to infer evolutionary history.
High-Performance Computing (HPC) Cluster	Computing infrastructure with many processors and large memory capacity.	Essential for running computationally intensive Bayesian analyses (e.g., SnappNet) in a practical timeframe.
PhyloNet	A software package for phylogenetic network analysis, containing tools like MCMC_BiMarkers [54].	Provides a suite of utilities for inference and analysis, including methods for comparing networks and analyzing gene tree embeddings.

Distinguishing true introgression from ILS and homoplasy requires moving beyond simple tree models and summary statistics. Model-based methods that implement the multispecies network coalescent (MSNC), such as SnappNet, provide a statistically rigorous framework for this task. Performance benchmarks show that these next-generation tools offer not only superior accuracy on complex scenarios but also critical gains in computational efficiency. As genomic datasets continue to grow, the adoption of these powerful methods will be essential for uncovering the full extent and evolutionary impact of gene flow across the tree of life.

Gene tree estimation error (GTEE) represents a significant challenge in phylogenetics, with profound implications for understanding evolutionary histories, characterizing introgression, and accurately reconstructing phylogenetic networks. This error arises from inherent biological complexities and methodological limitations, potentially leading to inaccurate inferences about species relationships and evolutionary processes. This guide provides a comprehensive comparison of current GTEE correction strategies, evaluating their performance, underlying assumptions, and applicability for research on phylogenetic networks in introgression characterization. We synthesize experimental data from benchmark studies and detail methodological protocols to equip researchers with evidence-based recommendations for selecting appropriate correction approaches based on their specific research contexts and data characteristics.

Gene tree estimation error refers to the discrepancy between inferred gene trees and the true evolutionary history of gene families. This error stems from multiple sources including limited phylogenetic signal in sequence alignments, methodological limitations of tree inference algorithms, and the complex interplay of evolutionary processes such as incomplete lineage sorting (ILS), gene duplication and loss, and horizontal gene transfer [55]. The accuracy of gene trees is paramount for downstream analyses, particularly in the context of characterizing introgression using phylogenetic networks, where errors in individual gene trees can propagate through analyses and lead to incorrect inferences about reticulate evolutionary events.

The problem of GTEE is exacerbated by the fundamental difference in data quantity available for species tree versus gene tree estimation. While genomic datasets provide megabases or gigabases of data for inferring species histories, individual gene families typically offer only around 1.3 kilobases of coding sequence on average for phylogenetic analysis [55]. This limited information content, combined with the statistical challenges of modeling complex evolutionary processes, creates substantial potential for estimation error that must be addressed through robust correction methodologies.

Gene tree estimation error arises from both biological and methodological sources. Biological complexities include incomplete lineage sorting, where ancestral polymorphism persists through speciation events, creating legitimate differences between gene trees and species trees [55]. In certain evolutionary scenarios known as "anomaly zones," the most probable gene tree topology may legitimately differ from the species tree [55]. Additional complications arise from gene duplication and loss, horizontal gene transfer, and introgression events that create complex patterns of inheritance not captured by simple bifurcating trees.

Methodological sources of error include limitations in phylogenetic inference algorithms, insufficient phylogenetic signal in sequence alignments, and model misspecification. The impact of GTEE is particularly significant for phylogenetic network inference, as these errors can lead to incorrect identification of introgression events and erroneous network topologies. Studies have demonstrated that GTEE can substantially affect downstream analyses, including species tree estimation and the detection of evolutionary processes such as hybridization and introgression [55] [56].

Table: Primary Sources of Gene Tree Estimation Error and Their Impacts

Source Category	Specific Source	Impact on Gene Tree Accuracy	Effect on Phylogenetic Networks
Biological	Incomplete Lineage Sorting (ILS)	Causes legitimate gene tree-species tree discordance	May be misinterpreted as introgression if uncorrected
	Gene Duplication and Loss	Creates paralogy complications	Can lead to incorrect inference of reticulate events
	Horizontal Gene Transfer	Introduces topological conflicts	Directly affects network structure and reticulation nodes
Methodological	Limited Sequence Data	Reduces phylogenetic signal	Increases variance in network inference
	Model Misspecification	Biases parameter estimates	Affects branch lengths and topology in networks
	Algorithmic Limitations	Introduces inference artifacts	Propagates error to network reconstruction

Gene Tree Correction Strategies: A Comparative Analysis

Species Tree Attraction Methods

Species tree attraction methods operate by adjusting gene tree topologies to reduce their distance to a known species tree, under the assumption that discordance primarily stems from estimation error rather than biological processes. Two representative methods in this category are TRACTION and TreeFix, which employ different correction mechanisms and demonstrate varying performance characteristics [55].

TRACTION is a nonparametric method that improves uncertain branches by solving the RF-Optimal Tree Refinement problem, which resolves polytomies in an input tree to minimize Robinson-Foulds distance to a given binary tree [55]. This approach has been shown to perform well on simulated data with ILS, though its effectiveness diminishes under higher levels of ILS. Experimental data indicates that TRACTION-corrected gene trees are closer to the species tree than uncorrected trees in 11.7% to 60.8% of cases across different levels of phylogenetic informativeness [55].

TreeFix utilizes species tree information and sequence data based on a gene duplication and loss model to correct gene trees [55]. This method demonstrates a stronger attraction to the species tree, with corrected trees being closer to the species tree than true gene trees in 85.5% to 96.6% of cases across varying levels of phylogenetic informativeness [55]. However, this strong attraction can be problematic when true biological discordance exists between gene and species trees.

Statistical Binning Approaches

Statistical binning approaches, including Weighted Statistical Binning (WSB), address GTEE by grouping genes with similar phylogenetic signals to improve estimation accuracy. These methods leverage multi-locus data to enhance phylogenetic signal while accounting for sources of discordance [56].

The novel WSB+WQMC pipeline shares design features with the existing WSB+CAML pipeline but incorporates different statistical properties [56]. Experimental evaluations demonstrate that WSB+WQMC substantially improves gene tree and species tree accuracy on most datasets with low, medium, and moderately high ILS levels. Performance comparisons show that WSB+WQMC computes less accurate trees than WSB+CAML under certain low and medium ILS conditions but performs better or comparably on datasets with moderately high and high ILS [56].

Implicit vs. Explicit Phylogenetic Methods

Beyond direct gene tree correction, methodological choices in phylogenetic inference significantly impact GTEE. A fundamental distinction exists between implicit and explicit phylogenetic methods for detecting evolutionary events like horizontal gene transfer, with implications for how gene tree error propagates through analyses [57].

Explicit phylogenetic methods directly compare gene tree and species tree topologies to identify discordance indicative of evolutionary events. Tools such as RANGER-DTL and ALE use maximum likelihood frameworks to estimate rates of duplication, transfer, and loss events that best explain observed differences between gene and species trees [57]. These methods provide detailed information about potential donors and recipients in transfer events but are computationally intensive and sensitive to gene tree errors.

Implicit methods avoid direct tree comparison and instead infer evolutionary events from gene distribution patterns across genomes. Methods like GLOOME and Count use statistical frameworks based on phyletic patterns without requiring gene tree reconstruction [57]. These approaches are computationally faster and avoid errors associated with gene tree estimation but provide less detailed information about evolutionary events. Benchmark studies have demonstrated that implicit methods based on gene family presence-absence patterns consistently outperform explicit approaches based on gene tree-species tree reconciliation [57].

Table: Performance Comparison of Gene Tree Correction Methods Under Different Conditions

Method	Underlying Approach	Accuracy with Low ILS	Accuracy with High ILS	Computational Efficiency	Primary Limitations
TRACTION	Nonparametric RF-optimal refinement	Moderate	Decreases under high ILS	High	May worsen accuracy under high ILS
TreeFix	Species tree attraction with GDL model	High	Decreases with faster mutation rates	Moderate	Over-correction when biological discordance exists
WSB+WQMC	Statistical binning with quartet amalgamation	Moderate to High	High	Moderate	Less accurate than WSB+CAML in some low ILS conditions
ALE/RANGER-DTL	Explicit phylogenetic reconciliation	Varies with gene tree quality	Varies with gene tree quality	Low	Computationally intensive, sensitive to GTEE
GLOOME/Count	Implicit phyletic pattern analysis	High	High	High	Limited information on direction of transfer

Experimental Protocols and Benchmarking

Standardized Evaluation Framework

Rigorous evaluation of gene tree correction methods requires a standardized framework that quantifies performance across diverse evolutionary scenarios. The most common metric for assessing GTEE is the unrooted normalized Robinson-Foulds (RF) distance between inferred gene trees and true simulated gene trees [55]. This metric captures topological differences while normalizing for tree size, allowing comparison across datasets.

Experimental protocols should systematically vary key parameters that affect phylogenetic inference:

• Phylogenetic Informativeness: Manipulated through the number of sites in alignments (e.g., 200, 800, 2000 sites) and population mutation rate (θ) [55]
• Incomplete Lineage Sorting: Controlled by varying the scale of species trees in coalescent units (low, medium, high ILS) [55]
• Gene Tree Complexity: Affected by rates of duplication, transfer, and loss events [57]

Benchmark studies should include both simulated and empirical datasets. Simulations provide known ground truth for direct accuracy assessment but may not fully capture biological complexity [57]. Empirical validation can leverage biological patterns, such as the tendency of co-transferred genes to remain genomic neighbors, to indirectly assess method performance [57].

Workflow for Method Evaluation

Experimental Results and Performance Trends

Comprehensive benchmarking reveals several key trends in gene tree correction performance. Species tree attraction methods frequently increase topological error by "correcting" gene trees to be closer to species trees even when true biological discordance exists [55]. The performance of these methods is highly dependent on evolutionary parameters:

• TRACTION-corrected gene trees show variable performance, with only 0.485% to 18.4% being closer to true gene trees than uncorrected trees across different levels of phylogenetic informativeness [55]
• TreeFix demonstrates stronger species tree attraction, with 5.34% to 80.6% of corrected trees closer to true gene trees than uncorrected trees, but performance decreases substantially with increasing phylogenetic informativeness [55]
• Statistical binning approaches like WSB+WQMC show more consistent improvement across different ILS levels, particularly outperforming other methods under conditions of high ILS [56]

The effectiveness of all correction methods is influenced by the underlying evolutionary processes. Under higher levels of ILS, methods that assume gene tree discordance primarily stems from estimation error rather than biological processes tend to perform poorly [55].

Methodological Considerations and Future Directions

Limitations of Current Approaches

Current gene tree correction methods face several fundamental limitations. Species tree attraction methods rely on heuristics that effectively remove outlier nodes without adequate statistical modeling of evolutionary processes [55]. These approaches struggle to distinguish between biological discordance and estimation error, potentially "over-correcting" legitimate differences between gene and species trees.

Many correction methods incorporate simplified models of evolution that fail to capture biological complexity. For example, most methods do not adequately account for population-level processes such as genetic drift, intragenic rearrangements, or the effects of extinct or unsampled species [57] [55]. This model misspecification can lead to systematic biases in corrected trees.

The reliance on simulated data for method validation presents another limitation. While simulations provide known ground truth, they often embed the same assumptions used in inference methods, creating potential circularity in validation [57]. Additionally, simulations may not accurately capture the complexity of real biological systems, limiting the generalizability of method performance to empirical data.

Emerging Approaches and Innovations

Promising innovations in gene tree correction include the integration of DNA language models for phylogenetic inference. PhyloTune represents one such approach, leveraging pretrained DNA language models to identify taxonomic units of new sequences and extract high-attention regions for targeted phylogenetic updates [58]. This method demonstrates that phylogenetic trees can be constructed by automatically selecting informative sequence regions without manual marker selection, potentially reducing error introduced by arbitrary region selection.

Advances in phylogenetic network inference offer complementary approaches to addressing gene tree error. Methods like ALTS infer tree-child networks by aligning lineage taxon strings from phylogenetic trees, providing a framework for reconciling conflicting gene trees without explicitly "correcting" individual estimates [6]. This approach shifts focus from correcting potentially erroneous gene trees to directly inferring networks that accommodate topological conflicts.

Future methodological development should prioritize several key areas:

• Integration of more realistic evolutionary models that simultaneously account for ILS, GDL, and HGT
• Development of probabilistic correction frameworks that quantify uncertainty in corrected trees
• Creation of benchmark datasets with verified empirical patterns to complement simulated data
• Methods that scale to genomic-scale datasets while maintaining statistical rigor

Implications for Phylogenetic Network Research

Accurate gene tree estimation is particularly critical for phylogenetic network research focused on introgression characterization. Reticulate evolutionary events create complex patterns of relationships that cannot be captured by bifurcating trees, requiring methods that can reconcile conflicting phylogenetic signals across the genome [6]. Gene tree error directly impacts network inference by introducing spurious conflict that may be misinterpreted as introgression or obscuring genuine reticulate events.

The performance of phylogenetic network methods depends heavily on accurate input gene trees. Network approaches that compute the minimum tree-child network displaying a set of gene trees are sensitive to GTEE, as erroneous trees may necessitate additional reticulations in the network that do not reflect biological reality [6]. Methods like ALTS, which infer networks from lineage taxon strings, aim to mitigate this sensitivity but still require reliable phylogenetic signals from input trees [6].

For researchers characterizing introgression using phylogenetic networks, we recommend:

• Implementing rigorous gene tree correction methods appropriate for the expected evolutionary processes
• Validating network inferences using multiple approaches with different underlying assumptions
• Interpreting reticulations in networks with consideration of potential gene tree error sources
• Utilizing emerging methods that jointly model gene tree error and network structure

Table: Research Reagent Solutions for Gene Tree Error Correction

Tool/Resource	Type	Primary Function	Application Context
TRACTION	Software package	Nonparametric gene tree refinement	Correction under moderate ILS conditions
TreeFix	Software package	Species tree-based correction	Datasets with strong species tree prior
WSB+WQMC	Analysis pipeline	Statistical binning and quartet amalgamation	Multi-locus datasets with varying ILS
ALTS	Network inference	Tree-child network from lineage taxon strings	Reconciling conflicting gene trees directly
ALE/RANGER-DTL	Reconciliation framework	DTL event inference from tree comparison	Detailed evolutionary event characterization
SimPhy	Simulation software	Generating benchmark datasets with known truth	Method validation and performance testing

Gene tree estimation error remains a significant challenge for phylogenetic inference, with particular importance for accurate introgression characterization using phylogenetic networks. Current correction methods offer diverse approaches with distinct strengths and limitations, making method selection highly dependent on specific research contexts and dataset characteristics.

Based on comprehensive benchmarking studies, we recommend:

• For datasets with expected high biological discordance due to ILS, statistical binning approaches like WSB+WQMC generally outperform species tree attraction methods
• When applying species tree attraction methods, carefully consider the potential for over-correction of legitimate biological discordance
• For phylogenetic network inference, consider approaches that directly incorporate gene tree uncertainty rather than relying solely on pre-corrected trees
• Validate findings using multiple methods with different underlying assumptions to assess robustness to GTEE
• Prioritize methods that provide explicit quantification of uncertainty in corrected trees

Future methodological advances that integrate more realistic evolutionary models and leverage emerging approaches like DNA language models show promise for addressing current limitations. As phylogenetic networks continue to play an increasingly important role in characterizing introgression and other reticulate evolutionary processes, developing robust approaches for handling gene tree error will remain essential for advancing our understanding of complex evolutionary histories.

In phylogenetic research, accurately characterizing introgression—the transfer of genetic material between species or populations—is essential for understanding evolutionary dynamics. The selection of computational models for this task presents a fundamental trade-off: models must be complex enough to capture realistic biological processes yet simple enough to be computationally tractable and interpretable. This guide objectively compares the performance of prevailing methods for introgression detection, focusing on their application across diverse biological scenarios. The expanding genomic datasets across diverse taxa have created new opportunities to investigate the impact of introgression along individual genomes, making the precise identification of introgressed loci a rapidly evolving area of research [23]. Researchers must navigate three major methodological categories—summary statistics, probabilistic modeling, and supervised learning—each with distinct strengths and limitations for specific research contexts.

Methodological Approaches for Introgression Detection

Summary statistics represent the foundational approach for detecting introgression, utilizing calculated metrics from genetic data to identify unusual patterns suggestive of gene flow. These methods continue to evolve, with new implementations broadening their applicability across taxa [23]. The standalone summary statistic Q95(w, y) has demonstrated particular effectiveness in exploratory studies of adaptive introgression [59]. These approaches generally offer computational efficiency and straightforward interpretation, making them valuable for initial genome scans. However, they may lack power for detecting ancient or complex introgression events and often require careful calibration of significance thresholds.

Probabilistic Modeling Frameworks

Probabilistic methods provide a powerful framework for introgression detection by explicitly incorporating evolutionary processes through explicit demographic models. Techniques like the hidden Markov model (HMM) approach implemented in diCal-admix offer fine-scale insights across diverse species by modeling the underlying demographic history relating populations, including introgression events [60]. These model-based approaches can differentiate between shared ancestry due to incomplete lineage sorting and true introgression, offering a more nuanced interpretation of genomic patterns. Their model-based nature provides a principled approach to account for demographic history, but they often come with significant computational demands and require accurate specification of demographic parameters.

Supervised Machine Learning Approaches

Supervised learning represents an emerging approach with great potential, particularly when the detection of introgressed loci is framed as a semantic segmentation task [23]. These methods utilize classifiers trained on genomic features to distinguish introgressed from non-introgressed regions. The machine-learning based approach developed by Sankararaman et al. operates on suitably chosen "features" of the genetic data to detect Neanderthal introgression tracts [60]. Supervised methods can capture complex, multi-dimensional patterns without requiring explicit specification of demographic models, but they depend heavily on the quality and biological relevance of training data and may suffer from overfitting or limited generalizability across species.

Performance Comparison of Detection Methods

Quantitative Performance Metrics Across Methods

Recent benchmarking studies have evaluated methodological performance under controlled conditions to provide objective comparison metrics. The following table summarizes the performance of several prominent methods based on a comprehensive evaluation using simulated datasets under various evolutionary scenarios inspired by human, wall lizard (Podarcis), and bear (Ursus) lineages [59].

Table 1: Performance comparison of adaptive introgression detection methods

Method	Approach Category	Power	False Positive Rate	Computational Efficiency	Optimal Use Case
Q95(w, y)	Summary statistic	Moderate to High	Low	High	Initial genome-wide scans for adaptive introgression
VolcanoFinder	Summary statistic	Variable	Variable	Moderate	Genome-wide scans for adaptive introgression
Genomatnn	Supervised Learning	High	Low	Moderate	Fine-scale detection in well-characterized systems
MaLAdapt	Machine Learning	Variable	Variable	Moderate	Complex introgression scenarios
diCal-admix	Probabilistic Modeling	High	Low	Low	Precise tract detection with known demography

Impact of Evolutionary Scenarios on Performance

Method performance varies significantly across different evolutionary contexts, influenced by factors such as divergence time, migration timing, and population size. One study demonstrated that methods based on Q95 summary statistics proved most efficient for exploratory studies of adaptive introgression, while the overall behavior of these methods when faced with genomic datasets from evolutionary scenarios other than the human lineage was previously unknown [59]. Performance is particularly affected by:

• Divergence and migration times: Methods perform differently depending on the timing of historical gene flow
• Selection coefficient: Strength of selection on introgressed variants impacts detection power
• Population size: Effective population sizes influence genealogical patterns
• Recombination hotspots: Variation in recombination rates affects local patterns of diversity

Special Considerations for Method Selection

The hitchhiking effect of an adaptively introgressed mutation can strongly impact flanking regions, complicating discrimination between genomic window classes (AI/non-AI) [59]. Effective evaluation requires comparing potential adaptive introgression windows against three types of non-AI windows: independently simulated neutral introgression windows, windows adjacent to the window under AI, and windows from a second neutral chromosome unlinked to the chromosome under AI. Including adjacent windows in training data proves particularly important for correctly identifying the specific window containing the mutation under selection.

Experimental Protocols for Method Validation

Simulation-Based Benchmarking Framework

Robust evaluation of introgression detection methods requires carefully designed simulation studies that mimic key aspects of real genomic data while maintaining ground truth knowledge of introgressed regions. The following workflow outlines a comprehensive validation approach:

Experimental Workflow for Method Validation

This workflow implements several critical steps for rigorous method evaluation:

• Scenario Definition: Evolutionary scenarios should represent diverse combinations of divergence and migration times inspired by real evolutionary histories [59]
• Demographic Modeling: Implement demographic models using specialized software such as msprime to simulate genomic sequences under realistic evolutionary scenarios [59]
• Introgression Introduction: Introduce known introgressed tracts with specified selection coefficients to establish ground truth
• Method Application: Apply each detection method to the simulated data using standardized parameters
• Performance Quantification: Calculate power and false positive rates by comparing inferred introgressed regions to known simulated tracts

Empirical Validation Protocol

While simulation studies provide controlled performance assessments, validation with empirical data offers complementary insights into real-world applicability:

Empirical Validation Workflow

This protocol emphasizes:

• Dataset Selection: Curate empirical datasets from well-studied systems with prior evidence of introgression
• Method Concordance: Identify introgressed regions supported by multiple independent methods
• Functional Annotation: Annotate candidate regions with functional genomic data to assess biological relevance
• Experimental Follow-up: Design functional experiments to validate putative adaptive introgressed variants

Table 2: Essential research reagents and computational tools for introgression analysis

Resource Category	Specific Tools/Reagents	Function/Purpose	Key Considerations
Simulation Tools	msprime [59], ARCADE [61]	Generate synthetic genomic data under specified evolutionary scenarios	Balance between biological realism and computational efficiency
Detection Software	diCal-admix [60], VolcanoFinder, Genomatnn, MaLAdapt [59]	Identify introgressed genomic regions from sequence data	Match method selection to specific research question and data characteristics
Population Genomic Data	1000 Genomes Project [60], Species-specific genome assemblies	Provide empirical data for method application and validation	Data quality, sample size, and representation of diverse populations
Functional Annotation Databases	Gene ontology databases, Epigenomic annotations	Interpret potential functional consequences of introgressed regions	Species-specificity and tissue-context of functional annotations
Visualization Platforms	ggplot2 [62], Genome browsers	Create effective visualizations of genomic landscapes and results	Adhere to principles of effective data visualization [62]

The accurate characterization of introgression in phylogenetic networks requires careful consideration of the trade-offs between model complexity and biological reality. Summary statistics methods offer efficiency for initial scans, probabilistic approaches provide demographic rigor, and machine learning techniques capture complex patterns without explicit model specification. The optimal selection depends critically on the specific research context, including the evolutionary timescale, available genomic resources, and particular biological questions. As the field advances, improvements are particularly needed in computational efficiency, systematic benchmarking standards, and accessibility of implementation. Researchers should select methods whose underlying assumptions best match their biological systems and explicitly acknowledge how design choices impact the resulting biological insights [61]. This principled approach to model selection will maximize the reliability and biological relevance of introgression characterization across diverse phylogenetic contexts.

Runtime and Memory Optimization Strategies for Large Datasets

The advent of high-throughput sequencing technologies has enabled phylogenetic studies at unprecedented scales, yet this data explosion presents formidable computational challenges. Accurate characterization of introgression—the transfer of genetic material between species or populations—requires inference of phylogenetic networks, which are more computationally intensive than standard trees. Researchers face critical scalability limitations: as dataset size increases in both taxon count and sequence length, runtime and memory demands can become prohibitive, potentially compromising analytical accuracy. Understanding these constraints is essential for evolutionary biologists studying complex evolutionary histories where gene flow plays a significant role.

The fundamental challenge lies in the NP-hard nature of phylogenetic inference, where computational requirements grow super-exponentially with increasing taxa. This is particularly acute for network inference, which must account for both vertical descent and horizontal gene flow processes like introgression. Where traditional tree-based methods struggle to represent such complex evolutionary histories, network approaches provide more accurate models but at substantial computational cost. This comparison guide examines current strategies and tools that balance these competing demands of accuracy, runtime, and memory utilization for large-scale phylogenetic analyses.

Performance Comparison of Phylogenetic Inference Methods

Benchmarking Runtime and Accuracy Across Methods

Comprehensive evaluation of phylogenetic tools reveals significant variation in their performance characteristics, particularly as dataset scale increases. The table below summarizes key metrics for contemporary methods:

Table 1: Performance Comparison of Phylogenetic Inference Methods

Method	Max Taxa Scalability	Primary Optimization Strategy	Theoretical Runtime	Memory Efficiency	Introgression Characterization
InPhyNet	1,000+ taxa	Divide-and-conquer with subset decomposition	Linear scalability with taxa count	Moderate	Excellent for level-1 networks
SNaQ	~30 taxa	Pseudo-likelihood approximation with quartets	Weeks for >30 taxa	High for small datasets	Accurate for single reticulations
PhyloNet-ML	~25 taxa	Maximum likelihood with coalescent model	Prohibitive beyond 25 taxa	Low	High accuracy but limited scalability
PhyloTune	Targeted subtree updates	DNA language model with attention mechanism	Rapid subtree updates	High	Enables efficient tree updates
VeryFastTree	1,000,000+ taxa	Vectorization and parallelization	3x faster than FastTree-2	High for trees	Tree-based only (no network inference)

Empirical studies demonstrate that probabilistic inference methods like PhyloNet-ML achieve high accuracy for characterizing introgression but become computationally prohibitive beyond approximately 25 taxa, often requiring weeks of computation and failing to complete analyses with 30 or more taxa [35]. Methods employing pseudo-likelihood approximations, such as SNaQ, extend this limit but still struggle with datasets exceeding 50 taxa [63]. For larger datasets exceeding 100 taxa, divide-and-conquer approaches like InPhyNet achieve linear scalability while maintaining biological interpretability, making them particularly valuable for genome-scale analyses where introgression signals must be detected across numerous lineages [63].

Quantitative Performance Metrics

Recent empirical evaluations provide specific measurements of computational performance across methods:

Table 2: Quantitative Runtime and Accuracy Metrics

Method	Dataset Size	Runtime	Memory Usage	Accuracy (RF Distance)	Optimal Use Case
InPhyNet	200 taxa, 1000 gene trees	~5 hours	~12 GB RAM	0.031 RF	Large-scale network inference
SNaQ	30 taxa, 1000 gene trees	>2 weeks	~8 GB RAM	0.027 RF	Small, complex networks
PhyloNet-ML	25 taxa, 100 gene trees	~1 week	~16 GB RAM	0.021 RF	High-accuracy small networks
VeryFastTree	1,000,000 taxa	36 hours (36-core server)	~512 GB RAM	Equivalent to FastTree-2	Ultra-large tree inference
PhyloTune	100 taxa (subtree update)	Minutes for updates	~4 GB RAM	0.031-0.054 RF	Incremental tree updates

Notably, a scalability study found that topological accuracy generally degrades as taxon count increases across all methods, with similar effects observed when sequence mutation rates rise [35]. The most accurate methods for characterizing introgression were consistently probabilistic approaches maximizing likelihood under coalescent-based models, though these exact methods exhibited the most severe computational constraints [35]. This creates a fundamental trade-off where researchers must balance methodological sophistication against computational feasibility when designing analyses for introgression detection.

Experimental Protocols for Method Evaluation

Standardized Simulation Framework

To ensure fair comparison across methods, researchers have developed standardized simulation protocols that quantify both runtime performance and inference accuracy under controlled conditions. A representative experimental workflow involves these critical stages:

(Figure 1: Experimental workflow for evaluating phylogenetic methods)

The simulation protocol begins with true network generation using tools like scripts/generate_true_network.R to create model phylogenies with known reticulation events [63]. Parameters typically include the number of taxa (ranging from 25-200 for scalability assessment), number of gene trees (100-1000), and level of incomplete lineage sorting (low/high). Next, empirical data simulation employs scripts/simulate_empirical_data.sh to evolve sequences along the true network, generating both true gene trees and multiple sequence alignments with varying lengths (e.g., 100-1000 base pairs) [63].

For divide-and-conquer methods, a critical subset decomposition phase partitions the taxon set into smaller, more manageable subsets using algorithms like TINNIK or ASTRAL, typically restricting subsets to taxa that share close evolutionary relationships [63]. The constraint network inference then applies methods like SNaQ or PhyloNet to each subset independently, with runtime and memory usage tracked for each subset. Finally, the network merging phase combines constraint networks into a comprehensive species network, with overall performance evaluated using metrics like Robinson-Foulds distance to quantify topological accuracy against the true simulated network [63].

Evaluation Metrics and Parameterization

Rigorous method assessment requires multiple quantitative metrics captured throughout the simulation pipeline. For runtime evaluation, both total execution time and scalability with increasing taxa are measured, with particular attention to time complexity curves. Memory consumption is monitored at each stage, especially during likelihood calculations which often represent performance bottlenecks. Accuracy is primarily quantified using normalized Robinson-Foulds distance, which measures topological dissimilarity between inferred and true phylogenies, with lower values indicating better performance [58].

Systematic parameter variation is essential for comprehensive assessment. Key parameters include number of taxa (from 25 to 200+), number of gene trees (from 100 to 1000), sequence length (from 100 to 1000 base pairs), and reticulation complexity (from simple to complex networks) [63]. For methods incorporating machine learning approaches like PhyloTune, additional metrics include novelty detection accuracy and attention region identification performance, which determine how effectively the method identifies relevant taxonomic units and informative genomic regions for analysis [58].

Methodological Approaches and Optimization Strategies

Algorithmic Innovations for Scalable Inference

Divide-and-Conquer Framework

The InPhyNet method exemplifies the divide-and-conquer paradigm that enables analysis of previously intractable datasets. This approach decomposes the full phylogenetic inference problem into three discrete phases: (1) subset decomposition, where the complete taxon set is partitioned into smaller, non-overlapping subsets; (2) constraint network inference, where partial networks are estimated on each subset using established methods; and (3) network merging, where constraint networks are combined into a comprehensive species network [63]. This strategy achieves linear scalability with taxon count while maintaining accuracy under the multispecies network coalescent model, dramatically reducing inference time from weeks to hours for large datasets [63].

Language Model Acceleration

PhyloTune represents a novel approach leveraging advances in natural language processing for phylogenetic inference. This method uses pretrained DNA language models (e.g., DNABERT) to generate high-dimensional sequence representations, which enable two key optimizations: identification of the smallest taxonomic unit for new sequences, and extraction of high-attention genomic regions most informative for phylogenetic construction [58]. By focusing computational effort on these relevant subsets, PhyloTune significantly accelerates phylogenetic updates while maintaining comparable accuracy to full analyses, reducing computational time by 14.3% to 30.3% according to empirical tests [58].

Hardware-Aware Implementation

VeryFastTree demonstrates how low-level computational optimizations can dramatically improve performance for massive datasets. This highly-tuned implementation extends FastTree-2 with vectorization and parallelization strategies specifically designed for modern hardware architectures [64]. Key innovations include thread-level parallelization with configurable intensity, vector extensions utilizing AVX2/AVX512 instructions, and optimized math function implementations. The result is a 3x speedup over standard FastTree-2, enabling inference of trees with one million taxa in 36 hours on a dual 32-core server [64]. While limited to tree inference, this approach shows the substantial performance gains possible through hardware-aware implementation.

Computational Trade-offs in Method Selection

Method selection for introgression characterization involves navigating fundamental trade-offs between computational requirements and biological sophistication. The relationship between these factors can be visualized as follows:

(Figure 2: Decision framework for method selection based on dataset scale)

Probabilistic methods like PhyloNet-ML offer the highest biological accuracy for characterizing introgression but scale poorly, becoming computationally prohibitive beyond approximately 25 taxa [35]. Pseudo-likelihood approximations like SNaQ extend this limit to around 50 taxa while maintaining good accuracy for level-1 networks [63]. For larger datasets exceeding 100 taxa, divide-and-conquer methods like InPhyNet provide the best balance of network inference capability and computational feasibility [63]. At the extreme scale of millions of taxa, methods like VeryFastTree offer tremendous computational efficiency but are limited to tree inference without explicit reticulation representation [64].

Essential Research Reagents and Computational Tools

Successful implementation of optimized phylogenetic workflows requires familiarity with both methodological software and supporting computational tools. The following table catalogs key resources mentioned in performance studies:

Table 3: Research Reagent Solutions for Phylogenetic Analysis

Tool/Resource	Function	Application Context	Performance Characteristics
InPhyNet	Divide-and-conquer network inference	Large-scale phylogenetic network estimation	Linear time scalability with taxon count
PhyloNet	Probabilistic network inference	Reticulate evolutionary history inference	High accuracy but limited to <30 taxa
SNaQ	Pseudo-likelihood network inference	Species networks from quartets	Improved scalability over full likelihood
VeryFastTree	Optimized tree inference	Massive taxonomic dataset tree building	3x faster than FastTree-2, supports 1M+ taxa
PhyloTune	DNA language model for phylogenetics	Targeted phylogenetic updates	Rapid subtree identification and updating
ASTRAL	Species tree estimation	Incomplete lineage sorting analysis	Summary method for large datasets
IQ-TREE	Maximum likelihood tree inference	General phylogenetic analysis	Efficient for medium-sized datasets
HTCondor	High-throughput computing	Workload distribution across clusters	Enables distributed computation of subsets

These tools represent the current state-of-the-art for addressing computational challenges in phylogenetic inference. For introgression characterization specifically, the combination of ASTRAL for initial species tree estimation followed by network inference methods like InPhyNet or SNaQ provides a balanced approach that leverages the strengths of multiple methods [63]. For researchers working with extremely large datasets, pipeline optimization using workflow management systems like HTCondor can dramatically reduce overall runtime by distributing independent subset analyses across computing clusters [63].

Optimizing runtime and memory usage for phylogenetic analyses on large datasets requires strategic method selection informed by dataset scale and research objectives. For introgression characterization in studies with fewer than 30 taxa, probabilistic methods like PhyloNet-ML provide the highest accuracy despite significant computational demands. For medium-sized datasets (30-100 taxa), pseudo-likelihood approximations like SNaQ offer the best balance of accuracy and feasibility. For large-scale analyses exceeding 100 taxa, divide-and-conquer approaches like InPhyNet enable biologically meaningful network inference at previously impossible scales while maintaining linear time complexity.

Emerging methods incorporating machine learning and language models show promise for further accelerating phylogenetic workflows, particularly through targeted analysis of informative genomic regions and efficient handling of incremental updates. As phylogenetic datasets continue growing in both taxon sampling and sequence length, continued algorithmic innovation will be essential for enabling accurate characterization of complex evolutionary phenomena like introgression. Researchers should prioritize methods with demonstrated scalability and consider hybrid approaches that combine the strengths of multiple optimization strategies to address their specific computational constraints and biological questions.

Benchmarking Phylogenetic Network Methods: Performance Validation and Selection Guidelines

Comparative Accuracy of Parsimony, Likelihood, and Pseudo-likelihood Methods

Phylogenetic networks are essential for representing complex evolutionary histories involving non-vertical inheritance processes such as introgression, hybridization, and horizontal gene transfer. Accurately inferring these networks is crucial for researchers, scientists, and drug development professionals who rely on evolutionary relationships to understand pathogen evolution, drug target identification, and evolutionary mechanisms. The characterization of introgression—the integration of genetic material between species or populations—presents particular challenges that require sophisticated statistical approaches. Three primary computational frameworks have emerged for phylogenetic network inference: maximum parsimony (MP), maximum likelihood (ML), and pseudo-likelihood methods. Each offers distinct trade-offs between computational efficiency, scalability, and biological accuracy, making their comparative assessment vital for selecting appropriate methodologies in phylogenomic studies [65] [66].

This guide provides an objective comparison of these approaches, focusing specifically on their performance in characterizing introgression. We evaluate methodological frameworks based on empirical data and simulation studies, examining their accuracy, computational requirements, and optimal application scenarios. Understanding these trade-offs is particularly relevant for researchers working with large genomic datasets, such as those encountered in SARS-CoV-2 surveillance [65] or studies of rapidly diversifying groups like Anastrepha fruit flies [67].

Methodological Frameworks

Maximum Parsimony Approaches

Maximum parsimony operates on the principle that the evolutionary history requiring the fewest character-state changes (e.g., nucleotide substitutions) is most likely correct. In network inference, parsimony has been extended through hardwired and softwired interpretations. Hardwired parsimony counts character-state transitions along every edge of the network, while softwired parsimony identifies the maximum parsimony score among all trees displayed within the network [68]. The multi-objective optimization algorithm MO-PhyNet simultaneously minimizes hardwired parsimony, softwired parsimony, and the number of reticulations, revealing relationships between these criteria and demonstrating that softwired parsimony typically results in networks with more reticulations [68].

Maximum Likelihood Methods

Full maximum likelihood methods evaluate the probability of observing the sequence data given a phylogenetic network model and its parameters. These approaches incorporate sophisticated evolutionary models that account for incomplete lineage sorting (ILS) through the multispecies coalescent model, providing a robust statistical framework for distinguishing introgression from other sources of gene tree discordance [69] [66]. However, computing the full likelihood requires integrating over all possible gene trees and ancestral sequences, a process that becomes computationally intractable for datasets with many taxa or complex networks [70] [66].

Pseudo-likelihood Techniques

Pseudo-likelihood methods address the computational limitations of full likelihood approaches by decomposing the data into smaller, more manageable components. The two primary strategies involve rooted triples (three-taxon subsets) or quartets (four-taxon subsets). These methods compute likelihoods for each subset independently then combine them into a composite pseudo-likelihood score [70] [66]. For example, SNaQ (Species Networks applying Quartets) uses concordance factors—the proportion of genes supporting each quartet—to infer networks that account for both ILS and introgression while dramatically improving computational efficiency [66].

Performance Comparison

Quantitative Accuracy and Computational Efficiency

Table 1: Comparative Performance of Phylogenetic Inference Methods

Method	Theoretical Basis	Accuracy	Computational Efficiency	Optimal Use Cases
Maximum Parsimony	Minimizes character-state changes	High for closely-related taxa (e.g., SARS-CoV-2) [65]	Very high (thousands of times faster than ML) [65]	Large datasets of closely-related sequences; online phylogenetics [65]
Maximum Likelihood	Probability of data given model and parameters	High, especially with high divergence and multiple hits [65]	Very low (intractable for large networks) [70] [66]	Small datasets (<10 taxa, <3 reticulations) with sufficient computational resources [70]
Pseudo-likelihood	Composite likelihood from triples/quartets	Comparable to ML in simulation studies [66]	High (scales to larger datasets than full likelihood) [70] [66]	Larger datasets (dozens of taxa); genome-scale data with ILS and introgression [70] [66]

Table 2: Empirical Performance in Specific Biological Systems

System	Method Used	Key Finding	Reference
SARS-CoV-2	Maximum Parsimony (UShER, matOptimize)	More accurate phylogenies than ML; enables trees with >9 million genomes [65]	Turakhia et al. 2022 [65]
Xiphophorus fishes	Pseudo-likelihood (SNaQ)	Congruent with previous studies; refined hybridization placement [66]	Solís-Lemus et al. 2016 [66]
Dove wing vs. body lice	PhyloNet (likelihood-based)	Higher introgression in dispersed wing lice; 7 vs. 4 reticulations [71]	Sweet et al. 2020 [71]
Anastrepha fruit flies	Phylogenomic networks	Pervasive introgression; genes resilient to introgression have higher resolution [67]	Sánchez-Gracia et al. 2021 [67]

Method Selection Guidelines

The choice among parsimony, likelihood, and pseudo-likelihood methods depends on several factors:

• Dataset size: For datasets exceeding a few dozen taxa or with more than 3-4 reticulation events, full maximum likelihood becomes computationally prohibitive [70] [66]. Maximum parsimony and pseudo-likelihood offer more scalable alternatives.

• Sequence divergence: Maximum likelihood demonstrates superior accuracy when sequences have substantial divergence with potential multiple substitutions at single sites [65]. For closely-related taxa like SARS-CoV-2, where such events are rare, parsimony performs comparably with massive computational savings [65].

• Biological processes: When both incomplete lineage sorting and introgression contribute significantly to gene tree discordance, model-based approaches (full likelihood or pseudo-likelihood) provide more accurate inference than parsimony alone [69] [66].

• Analysis goal: For comprehensive genomic epidemiology requiring daily updates with new sequences, online parsimony approaches are uniquely capable [65]. For detailed characterization of introgression timing, direction, and extent, model-based methods offer richer statistical inference [69].

Experimental Protocols

Pseudo-likelihood Analysis with Quartet Concordance Factors

The pseudo-likelihood framework implemented in SNaQ provides a representative protocol for network inference:

• Gene tree estimation: Infer gene trees from multiple sequence alignments using standard phylogenetic methods. This step can be highly parallelized across loci [66].

• Quartet concordance factor calculation: For each 4-taxon set (quartet) across the species, calculate the proportion of gene trees supporting each of the three possible unrooted topologies. These observed concordance factors serve as the input data for network inference [66].

• Network optimization: Search the network space by maximizing the pseudo-likelihood function, which measures the fit between observed and expected concordance factors under the network model. The expected concordance factors are derived from the multispecies coalescent model with hybridization [66].

• Parameter estimation: Estimate branch lengths (in coalescent units) and inheritance probabilities (γ) for each hybridization event, representing the proportion of genes inherited from each parent [66].

This approach avoids the computational burden of full likelihood calculations while incorporating both ILS and introgression, enabling inference for dozens of taxa and multiple hybridization events [66].

Online Phylogenetics Protocol for Large Datasets

For massive datasets like SARS-CoV-2 genomes, an online phylogenetic approach provides an efficient alternative:

• Initial tree estimation: Construct a starting tree using a subset of representative sequences [65].

• Iterative sequence addition: For each new sequence, identify the optimal placement on the existing tree using maximum parsimony criteria, implemented in tools like UShER [65].

• Topology optimization: Refine the augmented tree using parsimony-based subtree pruning and regrafting (SPR) moves with tools like matOptimize [65].

• Daily updates: Repeat steps 2-3 as new sequences become available, maintaining a continuously updated phylogeny [65].

This protocol enables maintenance of a comprehensive SARS-CoV-2 phylogeny with over 9 million genomes, which would be computationally impossible with de novo maximum likelihood approaches [65].

Signaling Pathways and Workflows

The diagram illustrates the general workflow for phylogenetic network inference, highlighting the integration points for different methodological approaches. The pseudo-likelihood pathway (green) provides a balance between computational efficiency and statistical rigor, while maximum parsimony (red) offers speed advantages and maximum likelihood (red) provides theoretical optimality at computational cost. The process begins with multi-locus sequence data, proceeds through gene tree estimation and concordance factor calculation, and culminates in network inference with parameter estimation.

The Scientist's Toolkit

Table 3: Essential Software and Resources for Phylogenetic Network Analysis

Tool/Resource	Function	Methodology	Application Context
PhyloNet	Phylogenetic network inference	Maximum likelihood, pseudo-likelihood	Small to medium datasets (up to ~10 taxa for ML) [70] [66]
SNaQ	Species network inference	Pseudo-likelihood with quartets	Medium datasets (dozens of taxa); ILS and hybridization [66]
UShER/matOptimize	Massive-scale phylogenetics	Maximum parsimony	Ultra-large datasets (millions of sequences); SARS-CoV-2 [65]
IQ-TREE 2	Phylogenetic tree inference	Maximum likelihood	General phylogenetic analysis; comparison baseline [65]
MO-PhyNet	Multi-objective network inference	Parsimony (hardwired/softwired)	Comparing conflicting evolutionary hypotheses [68]
PhyloNetworks	Network comparison and analysis	Multiple methods	General network analysis and visualization [66]

The comparative analysis of parsimony, likelihood, and pseudo-likelihood methods reveals a clear accuracy-efficiency trade-off in phylogenetic network inference. Maximum likelihood provides the most statistically rigorous framework but becomes computationally prohibitive for datasets exceeding ~10 taxa or complex networks. Maximum parsimony offers remarkable scalability, enabling inference for millions of sequences, particularly valuable for closely-related pathogens like SARS-CoV-2. Pseudo-likelihood approaches strike an effective balance, maintaining much of the statistical power of full likelihood while scaling to dozens of taxa and successfully characterizing introgression even in challenging phylogenetic contexts.

For researchers characterizing introgression, selection among these methods should be guided by dataset scale, sequence divergence, and biological complexity. Pseudo-likelihood methods currently offer the most practical solution for most phylogenomic studies, while parsimony approaches remain indispensable for massive genomic surveillance efforts. Future methodological developments will likely focus on further scaling model-based approaches and integrating multi-objective optimization to address the inherent conflicts between different phylogenetic criteria.

Simulation-Based Validation Frameworks for Method Assessment

In phylogenetics, accurately characterizing introgression—the exchange of genetic material between species through hybridization—is essential for understanding evolutionary history. Simulation-based validation frameworks provide the critical foundation for assessing the performance of analytical methods designed to detect and interpret these complex signals. These frameworks allow researchers to benchmark computational tools against known evolutionary scenarios, thereby quantifying their accuracy, robustness, and limitations. As phylogenetic networks grow more sophisticated, moving beyond tree-like structures to represent reticulate evolution, the role of rigorous simulation-based assessment becomes increasingly important for driving reliable scientific discovery in genomics and drug development [72] [23].

This guide objectively compares prevalent methodologies and software used for introgression characterization, providing a structured analysis of their performance based on published experimental data and theoretical capabilities.

Methodological Approaches for Introgression Detection

The detection of introgression relies on a suite of computational methods, each with distinct underlying principles and applicability. These can be broadly categorized into summary statistics, probabilistic modeling, and phylogenetic networks.

Summary statistics, such as D-statistics (ABBA-BABA tests) and f4-ratio statistics, are widely used for their computational efficiency and ability to provide a initial signal of introgression. However, they typically offer limited resolution for pinpointing the precise genomic locations or timing of introgression events [23].

Probabilistic modeling approaches, including those based on the Multispecies Coalescent (MSC) model, provide a more powerful framework for phylogenetic inference and can explicitly account for incomplete lineage sorting (ILS). Methods like ASTRAL (species tree estimation) and BUCKy (gene tree concordance) operate within this paradigm. A key advancement is Simulation-Based Inference (SBI), which uses machine learning to create probabilistic emulators of complex simulators. Techniques like Mixed Neural Likelihood Estimation (MNLE) are particularly valuable for models with intractable likelihoods, enabling efficient Bayesian parameter inference from simulated data [73] [23].

Phylogenetic Networks offer the most direct representation of reticulate evolution. Software such as PhyloNet, BEAST 2, and IQ-TREE can infer networks from genomic data. Recent theoretical work has focused on semi-directed and multi-semi-directed networks, which are obtained by de-orienting rooted phylogenetic networks, retaining the direction only on arcs leading to reticulations (e.g., hybridization nodes). This is particularly valuable for identifiability studies and when root placement is problematic [74] [72].

Table 1: Comparison of Major Introgression Detection Method Categories

Method Category	Key Example Tools	Underlying Principle	Key Strengths	Major Limitations
Summary Statistics	D-statistics, f4-ratio	Calculating site pattern frequencies from allele data	Computationally fast, simple to apply, good for initial screening	Low genomic resolution, cannot infer precise timing or number of events
Probabilistic Modeling	ASTRAL, BUCKy, MNLE	Multispecies Coalescent, Neural Likelihood Estimation	Statistical power, accounts for ILS, provides confidence estimates (MNLE is highly simulation-efficient)	Computationally intensive, model misspecification risk
Phylogenetic Networks	PhyloNet, BEAST 2, IQ-TREE	Inference of explicit phylogenetic graphs from sequence data	Directly visualizes reticulation, models specific hybridization events	High computational demand, complex model space to search

Benchmarking Phylogenetic Software for Introgression Analysis

A diverse software ecosystem exists for phylogenetic analysis, each with specialized capabilities for handling introgression.

Table 2: Key Software Tools for Phylogenetic Network Analysis and Introgression Characterization

Software	Primary Methods	Key Features for Introgression	Data Input	Notable Applications
PhyloNet	Maximum Parsimony, Likelihood, Bayesian Inference	Specialized in inferring and analyzing phylogenetic networks from multi-locus data	Unlinked loci, gene trees	Analyzing evolutionary relationships with explicit network models [75]
BEAST 2	Bayesian Evolutionary Analysis (MCMC)	Dating evolutionary events, testing hypotheses with relaxed molecular clocks	Molecular sequences (DNA, AA)	Reconstructing phylogenies with complex evolutionary models [74] [75]
IQ-TREE	Maximum Likelihood, Model Selection (AIC/AICc/BIC)	Efficient phylogenomic inference, ultrafast bootstrapping, partition finding	DNA, protein, binary, morphology, codon data	Large-scale phylogenomic studies, model testing [74] [75]
Network	Median-Joining, Reduced Median	Creating networks from genetic/linguistic data, age estimation for ancestors	Genetic data (e.g., Sanger sequences), linguistic data	Phylogeographic analysis (e.g., SARS-CoV-2 outbreak) [76]
Dendroscope	Visualization of rooted trees/networks	Calculating and comparing rooted networks, tanglegrams, consensus networks	Rooted phylogenetic trees/networks	Visual comparison and analysis of complex networks [74] [75]
RevBayes	Bayesian Statistical Computation in Phylogenetics	Flexible modeling and simulation using interpreted 'Rev' language	Molecular sequences, morphological data	Custom model development and hypothesis testing [75]
MEGA	Distance, Parsimony, Maximum Likelihood	Comprehensive suite for molecular evolution analysis, divergence estimation	Aligned sequence data	User-friendly interface for diverse phylogenetic analyses [74] [75]
APE (R pkg)	Analysis of Phylogenetics and Evolution	Extensive collection of functions for tree/network analysis and visualization	Phylogenetic trees, comparative data	A foundational R package for phylogenetics [74]

Experimental Protocols for Validation

In Silico Benchmarking Using Simulation-Based Inference

A robust protocol for validating introgression methods involves Simulation-Based Inference (SBI), particularly useful for models where the likelihood function is intractable. The Mixed Neural Likelihood Estimation (MNLE) approach provides a state-of-the-art framework.

Workflow Overview:

• Simulator Definition: A computational simulator of the evolutionary process is defined, capable of generating synthetic genomic data (e.g., sequences, trees) from parameters θ, which include population sizes, divergence times, and introgression probabilities.
• Training Data Generation: Parameters are sampled from a prior distribution, and the simulator is run for each parameter set to generate corresponding data x. This creates a training dataset of N (θ, x) pairs.
• Conditional Density Estimation: A neural likelihood estimator is trained directly on the (θ, x) pairs. MNLE uses a mixed model, employing one neural density estimator for categorical data (e.g., allelic identities) and another for continuous data (e.g., branch lengths), conditioned on the categorical data.
• Validation and Inference: The trained MNLE provides an emulator of the simulator. Its learned likelihoods can be used with standard Bayesian inference methods (e.g., MCMC) to perform parameter inference on empirical data. This approach has been shown to achieve high accuracy with six orders of magnitude fewer simulations than previous methods like Likelihood Approximation Networks (LANs) [73].

Empirical Validation Framework: A Case Study onPicris

A detailed experimental protocol for validating introgression methods can be illustrated using a study on Mediterranean Picris species [19].

1. Biological System and Data Collection:

• Study Organism: The genus Picris (Compositae), a diploid plant group with a diversification center in the Mediterranean biodiversity hotspot.
• Data Generation: Nuclear and plastid genome data were generated using the Hyb-Seq approach (hybrid enrichment followed by sequencing).
• Rationale: This system exhibits strong incongruence between nuclear and plastid phylogenies and variation in key functional traits (life strategy and fruit morphology), suggesting a history of hybridization and adaptive evolution.

2. Phylogenetic and Network Analysis:

• Phylogenetic Inference: Initial phylogenetic trees were inferred from the genomic data to establish major lineages (Clade A and Clade B).
• Network Analysis: A phylogenetic network was reconstructed to visualize and test for reticulate evolution. The analysis revealed two major historical introgression events.
• Trait Evolution Mapping: Life strategy (iteroparity vs. semelparity) and fruit morphology (heterocarpy vs. homocarpy) were mapped onto the phylogeny to test for association with introgression events.

3. Key Findings and Validation Insights:

• The earliest introgression event involved the Turkish endemic P. campylocarpa hybridizing with the ancestor of the P. cyprica–P. pauciflora lineage.
• A second introgression preceded shifts in life history traits, but the study could rule out an adaptive introgression origin for these specific traits, demonstrating how such frameworks can dissect complex evolutionary narratives.
• This case study validates the use of phylogenetic networks for identifying specific, historical introgression events and testing their evolutionary consequences [19].

The following diagram visualizes the core workflow of a simulation-based validation study, from data acquisition to biological insight.

The Scientist's Toolkit: Essential Research Reagents and Software

Successful characterization of introgression relies on a combination of specialized software, analytical methods, and data resources.

Table 3: Essential Research Reagents and Computational Tools

Tool/Resource Name	Type/Category	Primary Function in Analysis
Hyb-Seq	Laboratory & Data Generation Method	Target enrichment sequencing for gathering phylogenomic data from non-model organisms [19].
PhyloNet	Software Package	Inference and analysis of phylogenetic networks to explicitly model reticulation events [75].
BEAST 2	Software Package	Bayesian evolutionary analysis for dating divergence times and testing evolutionary hypotheses [74] [75].
IQ-TREE	Software Package	Efficient maximum likelihood phylogenomic inference with sophisticated model selection [74] [75].
D-statistics	Analytical Method	Summary statistic for detecting population-level introgression via allele frequency patterns [23].
ASTRAL	Software Tool	Estimating species trees from sets of unrooted gene trees under the multi-species coalescent model [75].
MNLE	Analytical Method (SBI)	Highly simulation-efficient neural likelihood estimation for Bayesian inference on complex models [73].
Semi-directed Network	Mathematical Framework	A mixed graph model for phylogenetics where only reticulation arcs are directed, aiding in identifiability [72].

The assessment of phylogenetic methods for introgression characterization hinges on robust simulation-based validation frameworks. No single software or method universally outperforms others; the choice depends on the specific biological question, data type, and computational constraints. Summary statistics offer a fast screening tool, probabilistic models provide statistical rigor for well-specified problems, and phylogenetic networks deliver the most direct visualization of complex evolutionary histories.

Emerging approaches, including simulation-based inference with neural density estimators and advanced semi-directed network models, are pushing the boundaries of what is inferable from genomic data. The consistent application of the rigorous experimental protocols and benchmarking standards outlined in this guide is paramount for ensuring the accuracy and reliability of introgression research, with profound implications for understanding evolution, biodiversity, and the genomic basis of trait variation.

In phylogenomics, the accurate characterization of evolutionary histories involving processes like introgression and hybridization relies on moving beyond strictly bifurcating trees to the more general framework of phylogenetic networks. This paradigm shift introduces a core computational challenge: the Network Edge Penalty. This penalty conceptually represents the additional complexity—in terms of both model parameters and computational cost—incurred when explaining evolutionary data with a network versus a tree. As networks introduce reticulate nodes with multiple incoming edges (representing events like gene flow), inference methods must penalize this increased complexity to avoid overfitting. Equalizing the comparison between trees and networks therefore requires robust statistical frameworks and efficient algorithms that can navigate this trade-off. This is particularly critical for introgression characterization, where identifying and quantifying the signature of gene flow in genomic data is a primary research objective. This guide provides an objective comparison of contemporary phylogenetic network methods, focusing on their performance in addressing this fundamental challenge.

Method Comparison: Performance and Scalability

The performance of phylogenetic network methods is evaluated along two primary dimensions: topological accuracy (the correctness of the inferred evolutionary relationships) and scalability (computational efficiency in terms of runtime and memory usage). The following table summarizes the quantitative performance of leading methods based on empirical and simulation studies.

Table 1: Performance Comparison of Phylogenetic Network Inference Methods

Method	Inference Type	Key Performance Characteristics	Reported Runtime & Scalability	Best-Suited Context
ALTS [6]	Parsimony (Tree-Child)	Infers networks with large number of reticulations for 50 trees with 50 taxa.	~15 minutes for 50 taxa, 50 trees; scalable for larger datasets.	Large-scale analyses with many input gene trees.
MP (Maximum Parsimony) [35]	Parsimony	Lower accuracy compared to probabilistic methods on simulated data.	Not specified, but generally faster than probabilistic methods.	Preliminary analyses or datasets where computational resources are limited.
MLE/ MLE-length [35]	Probabilistic (Maximum Likelihood)	Most accurate method on simulated datasets with a single reticulation.	Prohibitive runtime/memory for datasets with >25 taxa; did not complete on 30+ taxa.	Small, complex datasets (<25 taxa) where accuracy is paramount.
MPL/ SNaQ [35]	Probabilistic (Pseudo-Likelihood)	High accuracy, though slightly lower than full MLE methods.	More efficient than MLE, but still faces scalability limits with increasing taxa.	Analyses requiring a balance between probabilistic accuracy and computational feasibility.
Neighbor-Net/ SplitsNet [35]	Concatenation (Distance-Based)	Lower topological accuracy; degrades with increased taxa and mutation rate.	Computationally efficient, capable of handling larger numbers of taxa.	Exploratory data analysis to visualize conflicting phylogenetic signals.

The data reveals a clear trade-off between accuracy and scalability. Probabilistic methods like MLE achieve the highest accuracy by explicitly modeling evolutionary processes such as incomplete lineage sorting (ILS) and gene flow under a coalescent framework [35]. However, this accuracy comes at a high computational cost, rendering them infeasible for datasets exceeding 25-30 taxa. In contrast, parsimony-based methods like ALTS demonstrate remarkable scalability, handling dozens of trees and taxa within practical timeframes, making them suitable for larger-scale phylogenomic studies [6]. Concatenation methods, while efficient, generally show the lowest topological accuracy as they do not fully account for sources of gene tree discordance like ILS [35].

Experimental Protocols for Network Inference

To ensure reproducibility and provide a clear framework for evaluation, this section outlines the standard experimental protocols used in the field to benchmark network inference methods.

Data Simulation and Performance Measurement

Simulation studies typically employ a two-step process to generate realistic genomic data and evaluate inference accuracy [77] [35]:

• Sequence Simulation under the Coalescent with Recombination: DNA sequence alignments are generated using a model that incorporates both neutral coalescence and recombination. Key parameters varied during simulation include:

  • Substitution Rate: The rate of sequence mutation (e.g., 6.25×10⁻⁶ to 6.25×10⁻⁷ substitutions/site/generation) [77].
  • Recombination Rate: The rate of recombination events (e.g., 0 to 4×10⁻⁶ events/site/generation) to model reticulate evolution [77].
  • Number of Taxa and Sequence Length: The scale of the dataset is varied (e.g., 10 to 50 taxa, 500 to 1000 base pairs) to test scalability [77] [35].
  • Site-Rate Heterogeneity: Models like a gamma distribution can be used to account for variable mutation rates across sites [77].

• Topological Accuracy Assessment: The inferred networks are compared to the true, simulated phylogeny. A common metric involves comparing "splits" (bipartitions of taxa) or enumerating the trees contained within the simulated and inferred networks. Accuracy is measured by how often the true underlying evolutionary relationships are recovered within the set of possible histories represented by the inferred network [77].

The ALTS Algorithm for Tree-Child Network Inference

The ALTS (Aligning Lineage Taxon Strings) algorithm introduces a novel protocol for inferring parsimonious tree-child networks. Its workflow is as follows [6]:

• Input: A set of binary phylogenetic trees (e.g., gene trees) on a common set of taxa, X.
• Taxon Ordering: The algorithm iterates over all possible total orderings (π) of the taxon set X.
• Internal Node Labeling: For each tree and ordering π, the internal nodes are labeled using a specific function. The root is labeled with the smallest taxon (π₁), and other nodes are labeled with the largest taxon among the smallest taxa in its two child clades.
• Lineage Taxon String (LTS) Calculation: For each taxon τ (except π₁), its unique path from the root to the leaf τ is identified. The LTS is the sequence of labels from the first node where the minimum taxon in the child clade equals τ.
• Common Supersequence Finding: For each taxon πᵢ, a common supersequence (βᵢ) is found for the LTSs of πᵢ across all input trees.
• Network Construction: The network is built from the paths generated by the common supersequences. "Vertical edges" form the paths, and "left-right edges" are added from nodes in the paths to the heads of other paths based on the symbols in the supersequences.

The following diagram visualizes the workflow of the ALTS algorithm.

Figure 1: The ALTS algorithm workflow for inferring tree-child networks from a set of input gene trees.

Successful inference and characterization of introgression via phylogenetic networks require a suite of computational tools and resources. The following table details key solutions used in the studies cited in this guide.

Table 2: Key Research Reagent Solutions for Phylogenomic Network Analysis

Resource / Tool	Type/Function	Role in Introgression Research
ALTS [6]	Computer Program	Implements the ALTS algorithm to infer minimum tree-child networks from multiple gene trees by aligning lineage taxon strings.
PhyloNet [35]	Software Package	A platform for analyzing phylogenetic networks, hosting implementations of methods like MLE, MLE-length, and MP for network inference.
HYBRIDIZATION NUMBER [6]	Computer Program	Calculates the minimum hybridization number for two phylogenetic trees, a key parsimony-based approach for quantifying reticulation.
EUA Dataset [78]	Standardized Real-World Dataset	A standard real-world dataset used for evaluating the performance of computational methods in phylogenetics and edge computing.
Transcriptome Datasets [67]	Genomic Data	Used for phylogenomic analyses to infer orthologous genes and detect signals of introgression across lineages.
Simulated Coalescent with Recombination Datasets [77] [35]	Benchmarking Data	Computer-simulated sequence alignments generated under a known model with reticulation, used for controlled method evaluation and validation.

The equalization of tree and network comparison hinges on how effectively methods manage the inherent Network Edge Penalty. Current evidence indicates that no single method optimally balances topological accuracy, biological interpretability, and computational scalability across all problem scales. For focused studies of closely related taxa with strong signals of introgression, probabilistic methods (MLE, MPL/SNaQ) provide the most statistically rigorous inference, despite their stringent computational limits. For broader-scale phylogenomic surveys, where the number of taxa is large, parsimony-based approaches like ALTS offer a practical and scalable alternative for initial network estimation. The choice of method must therefore be guided by the specific biological question, the scale of the dataset, and available computational resources. Future methodological development is critically needed to bridge this scalability-accuracy gap, particularly in creating approximate-likelihood models that remain computationally tractable for large numbers of genomes, thereby empowering more precise characterization of introgression across the tree of life.

Empirical Benchmarking Across Diverse Biological Systems

Empirical benchmarking is fundamental for validating computational methods in evolutionary biology. By providing objective, data-driven comparisons of performance across different algorithms and datasets, benchmarking studies allow researchers to select the most appropriate tools for specific biological questions. This is particularly crucial for complex tasks like inferring phylogenetic networks and characterizing introgression, where methodological choices can dramatically impact biological interpretations. This guide synthesizes recent benchmarking findings across phylogenetic inference, introgression detection, and related biological domains, providing researchers with actionable insights for method selection and experimental design.

Comparative Analysis of Phylogenetic Inference Methods

Performance Benchmarking of Sequence versus Structure-Based Phylogenetics

Recent advances in artificial-intelligence-based protein structure prediction have enabled new approaches to phylogenetic tree reconstruction. Structural phylogenetics leverages the principle that protein structure evolves more slowly than sequence, potentially preserving evolutionary signal over longer timescales [49].

A large-scale evaluation of nine structure-informed approaches compared to state-of-the-art sequence-based methods revealed that certain structural methods outperform sequence-only approaches, particularly for highly divergent datasets [49]. The top-performing pipeline, termed FoldTree, uses a structural alphabet to align sequences and computes evolutionary distances based on statistically corrected structural similarity (Fident) [49].

Table 1: Performance Comparison of Phylogenetic Inference Methods

Method	Approach	Best Use Case	TCS Performance	Molecular Clock Adherence
FoldTree	Structure-informed (structural alphabet + NJ)	Divergent protein families	Highest TCS on CATH dataset [49]	Competitive [49]
Structure-informed ML	Combined structure+sequence likelihood	Intermediate divergence	Moderate TCS [49]	Not specified
Sequence-based ML	Sequence-only likelihood	Closely related sequences	Lower TCS on divergent families [49]	Standard
BUSCO rate stratification	Site-specific rate modeling	Deep phylogenies	Improved taxonomic congruence [79]	Not specified

The FoldTree approach demonstrated particular strength when benchmarking against the CATH database of evolutionary-related protein structures, where it "outperformed the sequence-based methods by a larger margin" compared to performance on more closely-related OMA datasets [49]. Filtering input families based on AlphaFold prediction confidence (pLDDT) further improved structural tree performance, suggesting that advancing structural prediction methods will continue to benefit structural phylogenetics [49].

Universal Orthologs and Site-Rate Stratification for Deep Phylogenies

Benchmarking Universal Single-Copy Orthologs (BUSCO) genes provides another avenue for improving phylogenetic accuracy. A comprehensive analysis of 11,098 eukaryotic genomes revealed that sites evolving at higher rates produce "up to 23.84% more taxonomically concordant, and at least 46.15% less terminally variable phylogenies" compared to lower-rate sites when proper site stratification is employed [79].

This research led to the development of CUSCOs (Curated set of BUSCO orthologs), which reduce false positives in assembly quality assessment by up to 6.99% compared to standard BUSCO searches [79]. For researchers constructing deep phylogenies, selective use of faster-evolving sites in concatenated alignments appears to produce the most congruent and least variable phylogenies [79].

Benchmarking Introgression Detection Methods

Vulnerability of Site-Pattern Methods to Rate Variation

Accurate detection of introgression is crucial for characterizing phylogenetic networks. Recent theoretical and simulation studies demonstrate that popular site-pattern methods for introgression detection show high sensitivity to even minor deviations from the molecular clock assumption [27].

Table 2: Performance of Introgression Detection Methods Under Rate Variation

Method	Type	False Positive Rate (Weak Rate Variation)	False Positive Rate (Moderate Rate Variation)	Key Limitation
D-statistic	Site-pattern	Up to 35% [27]	Up to 100% [27]	Assumes no multiple hits
HyDe	Site-pattern	Elevated [27]	Elevated [27]	Sensitive to rate heterogeneity
D3	Branch-length	Not specified	Not specified	More robust to rate variation
QuIBL	Branch-length	Not specified	Not specified	More robust to rate variation

The D-statistic and HyDe methods are particularly vulnerable in shallow phylogenies (approximately 300,000 generations), where weak rate variation (17% difference between lineages) can inflate false positive rates to 35% using site pattern counts from a 500 Mb genome [27]. Moderate rate variation (33% difference) can increase false positive rates to 100% [27]. Employing a more distant outgroup intensifies these spurious signals [27].

Experimental Protocol for Introgression Benchmarking

The vulnerability of introgression detection methods was quantified through mathematical analysis and simulations across phylogenetic depths of 10^4 to 10^6 generations [27]. The methodology employed:

• Theory Development: Mathematical derivation of expected D-statistic values under varying degrees of rate variation between sister lineages, incorporating parameters for phylogenetic age, effective population size, and outgroup distance [27].

• Simulation Framework: Implementation of the multispecies coalescent with introgression (MSci) model, with speciation and introgression times (τ = Tμ) and population sizes measured in expected mutations per site [27].

• Rate Variation Assessment: Application of relative rate tests to empirical datasets from six genera to quantify realistic rate variation ranges, revealing common intra-generic rate disparities of 10-30% with some exceeding 50% [27].

Flowchart: Introgression Benchmarking Methodology

Benchmarking Frameworks Across Biological Domains

CausalBench for Network Inference from Single-Cell Data

Beyond phylogenetics, robust benchmarking frameworks have been developed for causal network inference in cellular systems. CausalBench is a comprehensive benchmark suite for evaluating network inference methods using real-world large-scale single-cell perturbation data [80].

Unlike traditional synthetic benchmarks, CausalBench incorporates "biologically-motivated metrics and distribution-based interventional measures" providing more realistic evaluation of network inference methods [80]. The framework uses two large-scale perturbational single-cell RNA sequencing datasets with over 200,000 interventional datapoints across RPE1 and K562 cell lines [80].

Evaluation of state-of-the-art methods revealed that:

• Poor scalability of existing methods limits performance in real-world environments [80]
• Methods using interventional information do not consistently outperform observational methods, contrary to theoretical expectations [80]
• The Mean Difference and Guanlab methods demonstrated top performance in statistical and biological evaluations respectively [80]

Integrative Metagenomics and Metabolomics Benchmarking

Another benchmarking study evaluated nineteen statistical methods for integrating microbiome and metabolome datasets [81]. This work addressed four key research goals: global associations, data summarization, individual associations, and feature selection [81].

The benchmark employed realistic simulations based on three real microbiome-metabolome datasets with different sample sizes, feature numbers, and data structures [81]. Performance was assessed through 1,000 replicates per scenario, with top-performing methods subsequently validated on real gut microbiome data from Konzo disease [81].

Experimental Protocols for Phylogenetic Benchmarking

Structural Phylogenetics Protocol (FoldTree)

The top-performing FoldTree method employs the following methodology [49]:

• Input Processing: Protein structures or high-confidence AlphaFold predictions (filtered by pLDDT)

• Structural Alignment: Foldseek used for all-versus-all comparison using a structural alphabet

• Distance Calculation: Statistically corrected structural similarity (Fident) computed from structural alphabet alignments

• Tree Building: Neighbor-joining applied to the pairwise distance matrix

• Validation: Taxonomic congruence scoring against known taxonomy

Research Reagent Solutions for Phylogenetic Benchmarking

Table 3: Essential Research Reagents and Resources

Reagent/Resource	Function	Application Note
Foldseek	Structural alignment using structural alphabet	Core component of FoldTree pipeline [49]
BUSCO gene sets	Universal single-copy orthologs for phylogeny	Gene content influenced by evolutionary history [79]
CUSCOs	Curated BUSCO orthologs with reduced false positives	Provides higher specificity for major eukaryotic lineages [79]
AlphaFold predictions	High-accuracy protein structure models	Filter by pLDDT for improved tree building [49]
CATH database	Classified protein structures for benchmarking	Contains evolutionary-related families for validation [49]
NORtA algorithm	Simulates microbiome-metabolome data	Generates data with arbitrary distributions and correlations [81]

Flowchart: Structural Phylogenetics Workflow

Empirical benchmarking across diverse biological systems reveals both opportunities and limitations in current methodologies for phylogenetic inference and introgression characterization. Structure-based phylogenetic methods like FoldTree show particular promise for resolving deep evolutionary relationships where sequence signal has saturated [49]. However, widely used introgression detection methods display critical vulnerabilities to rate variation, potentially compromising many reported gene flow events [27].

These findings emphasize the necessity of rigorous benchmarking using realistic datasets and biologically-motivated metrics when selecting methods for constructing phylogenetic networks. Researchers should carefully consider evolutionary timescales, rate heterogeneity, and taxonomic sampling when choosing between alternative approaches for introgression characterization. Future methodological development should focus on creating more robust approaches that explicitly account for biological complexities like substitution rate variation across lineages.

Best Practices for Method Selection Based on Dataset Properties

In the field of phylogenetics, accurately characterizing introgression is crucial for understanding evolutionary processes and genetic diversity. The performance of phylogenetic network methods is not uniform; it is highly dependent on the properties of the underlying dataset. This guide provides a structured, evidence-based approach for researchers and drug development professionals to select the most appropriate analytical methods by objectively evaluating their performance against core dataset characteristics. Adhering to these best practices ensures that conclusions about introgression are robust, reliable, and reproducible.

Understanding Dataset Characteristics

The first step in method selection is a thorough understanding of your dataset's intrinsic properties. These characteristics directly influence which analytical techniques will be most effective. The key properties can be categorized as follows:

• Scale and Dimensions: This includes the number of taxa, the number of genetic loci or sites, and the overall sequence length. These factors determine the computational complexity and the statistical power of the analysis.
• Data Quality and Distribution: This encompasses the presence of missing data, the degree of sequence error or uncertainty, the level of genetic diversity, and the balance of information across taxa.
• Evolutionary Signal Complexity: This refers to properties that make the data more challenging to model, such as the degree of incomplete lineage sorting (ILS), the relative branch lengths, the strength and timing of introgression signals, and the potential for recombination.

The diagram below illustrates the logical workflow for profiling a dataset and connecting its properties to methodological choices.

A Framework for Method Selection

General data science research confirms that no single analytical technique performs best across all situations; performance is intrinsically linked to dataset characteristics [82]. This "no free lunch" theorem underscores the need for a selective, systematic approach to method selection. The following framework, derived from these principles, guides the evaluation of phylogenetic methods.

• Principle of Correctness: The selected method must be statistically consistent with the model assumptions that match your data's properties (e.g., levels of ILS, heterotachy).
• Principle of Performance: The method's computational demands (time and memory) must be feasible given the scale of your dataset (number of taxa/loci).
• Principle of Usability: The method should be implementable with the available expertise and computational resources, ensuring the analysis is both executable and interpretable.

Comparative Performance of Key Methods

The table below summarizes the typical performance of major phylogenetic method categories in relation to critical dataset properties. This data is synthesized from empirical studies in phylogenetics and machine learning [82].

Method Category	Optimal Dataset Properties	Performance Strengths	Known Limitations & Sensitivity
Summary Methods (e.g., ASTRAL, ASTRID)	Large genomic datasets (100s-1000s of loci). Moderate to high levels of ILS.	Highly scalable and fast. Statistically consistent under the multi-species coalescent. Robust to gene tree estimation error.	Less accurate with very few loci. Sensitive to incorrect gene trees from poor alignment or model misspecification.
Maximum Likelihood (e.g., RAxML, IQ-TREE)	Datasets with strong phylogenetic signal. Moderate number of taxa (10s-100s).	High accuracy with sufficient signal. Extensive and optimized model selection.	Computationally intensive for large taxa sets. Can be inconsistent under high ILS without proper models (e.g., multispecies coalescent).
Bayesian Inference (e.g., MrBayes, BEAST2)	Complex evolutionary models. Smaller datasets (typically <100 taxa). Divergence time estimation.	Provides credibility intervals (posterior support). Handles complex models and missing data naturally.	Extremely computationally demanding. Convergence can be difficult to assess with large datasets.
Phylogenetic Networks (e.g., PhyloNet, SNaQ)	Evidence of reticulate evolution (hybridization, introgression). Specific, well-defined questions about gene flow.	Explicitly models non-tree-like evolutionary processes. Can quantify introgression probability and direction.	High model complexity requires careful parameterization. Scalability can be limited by number of taxa and reticulations. Sensitive to violations of its assumptions.

Experimental Protocols for Validation

To objectively compare methods for a specific research question, a standardized experimental protocol is essential.

Protocol 1: Benchmarking with Simulated Data This protocol uses simulated data where the true evolutionary history is known, allowing for precise accuracy measurement.

• Dataset Generation: Use a simulator like ms or Seq-Gen to generate sequence alignments under a known phylogenetic network model with controlled parameters (e.g., population sizes, divergence times, introgression events).
• Parameter Variation: Systematically vary key dataset properties across simulations, such as the number of taxa, sequence length, and the strength and timing of introgression.
• Method Application: Run each candidate method (e.g., ASTRAL, SNaQ, a Bayesian method) on all simulated datasets.
• Performance Metric Calculation: Calculate accuracy metrics for each method-dataset combination, including:
  • True Positive Rate (TPR) for Introgression: The proportion of simulated introgression events correctly identified.
  • False Positive Rate (FPR) for Introgression: The proportion of inferred introgression events that were not present in the simulation.
  • Distance to True Network: A topological measure (e.g., Robinson-Foulds distance) quantifying how close the inferred network is to the true one.

Protocol 2: Validation with Empirical Controls This protocol uses empirical data with established, well-supported evolutionary relationships.

• Control Selection: Identify an empirical dataset (e.g., from a clade like Heliconius butterflies or Anopheles mosquitoes) where introgression is strongly supported by multiple independent lines of evidence.
• Data Subsampling: Create smaller, tractable subsets of the full data (e.g., by selecting specific chromosomes or gene families).
• Blinded Analysis: Apply the methods to these subsets without reference to the established "true" result.
• Result Comparison: Compare the methods' inferences against the known biology to assess which one recovers the expected introgression events most reliably and with the highest support values.

The workflow for implementing these validation protocols is outlined below.

The Scientist's Toolkit: Essential Research Reagents

The following table details key software and data resources essential for conducting research on method selection for phylogenetic networks.

Item Name	Function / Purpose	Example Use Case in Introgression Research
PhyloNet	Software package for inferring and analyzing phylogenetic networks, specifically designed for reticulate evolution.	Quantifying the probability and direction of gene flow between sister species in a species complex.
IQ-TREE	Efficient maximum likelihood phylogenetic software with extensive model testing and branch support measures.	Inferring accurate individual gene trees from multi-locus alignments as input for a summary method like ASTRAL.
BEAST2	Bayesian evolutionary analysis software for estimating rooted, time-calibrated trees and population dynamics.	Co-estimating divergence times and introgression events within a known phylogenetic framework.
ms Simulator	Coalescent-based simulator for generating genetic sequence data under complex evolutionary scenarios.	Creating benchmark datasets with known introgression events to test the statistical power of different network methods.
Empirical Reference Dataset	A well-curated genomic dataset from a clade with previously confirmed and characterized introgression.	Serving as a positive control to validate that a chosen method can recover known, biologically real introgression events.

Integrating Multiple Approaches for Robust Introgression Detection

The characterization of introgression—the exchange of genetic material between populations or species—is a fundamental challenge in evolutionary genomics. While phylogenetic trees have long been the standard for representing evolutionary relationships, the increasing recognition of reticulate evolution through hybridization and introgression has driven the development of more complex phylogenetic networks. The accurate detection and characterization of introgression are particularly crucial in biomedical research, where introgressed variants may influence disease susceptibility, drug metabolism, or adaptive traits. This review objectively compares the performance of leading methodological approaches for introgression detection, evaluating their strengths, limitations, and optimal applications within a framework prioritizing analytical robustness.

Each method class operates on different principles and input data, from summary statistics that track allele patterns to probabilistic models that incorporate explicit evolutionary parameters and supervised learning approaches that identify complex genomic signatures. The performance of these methods varies significantly across evolutionary scenarios, influenced by factors such as divergence times, population sizes, migration rates, and selection strength. By synthesizing experimental data and benchmarking studies, this guide provides researchers with evidence-based recommendations for selecting and integrating approaches to characterize introgression accurately.

Methodological Approaches: Principles and Protocols

Principles of Operation: Summary statistics methods, particularly the D-statistic (ABBA-BABA test), detect introgression by analyzing patterns of derived allele sharing among four taxa—two sister populations, an outgroup, and a potential introgressing lineage [83]. The test operates by counting discordant gene tree topologies: "BABA" patterns suggest gene flow between one of the sister lineages and the external test lineage, while "ABBA" patterns support the alternative relationship. Significant deviations from the expected equal distribution of these patterns indicate introgression.

For larger phylogenies, the ( f )-statistics framework extends this logic to five-taxon phylogenies, enabling simultaneous testing of multiple introgression hypotheses and polarization of introgression directionality [83]. These methods are computationally efficient and can be applied genome-wide or in sliding windows to localize introgressed regions.

Detailed Experimental Protocol:

• Data Preparation: Obtain genotype data for at least four taxa (P1, P2, P3, Outgroup) with known phylogenetic relationships.
• Variant Calling: Identify derived alleles relative to the outgroup using standard variant calling pipelines.
• Pattern Counting: Tabulate sites matching ABBA (shared derived alleles between P2 and P3) and BABA (shared derived alleles between P1 and P3) patterns.
• D-statistic Calculation: Compute ( D = (\text{ABBA} - \text{BABA}) / (\text{ABBA} + \text{BABA}) ).
• Significance Testing: Assess deviation from D=0 using block jackknifing or permutation tests across genomic regions.
• Visualization: Generate Manhattan plots of D-statistic values across chromosomes to identify localized introgression signals.

Tree-Based Methods: Phylogenetic Incongruence Approaches

Principles of Operation: Tree-based methods detect introgression by analyzing incongruence among gene trees inferred from genomic regions. The underlying principle is that different genomic regions may have different evolutionary histories due to introgression events. By inferring thousands of gene trees from across the genome and examining their distribution, researchers can identify excesses of particular discordant topologies that signal introgression [38]. This approach is model-based and leverages the full sequence information rather than just counting derived alleles.

Species tree methods like ASTRAL account for incomplete lineage sorting but assume no introgression, while phylogenetic network approaches explicitly model reticulate evolution. The relative frequency of alternative topologies provides evidence for both the presence and direction of introgression events.

Detailed Experimental Protocol:

• Alignment Extraction: Extract multiple sequence alignments from whole-genome alignment data, typically in non-overlapping windows or defined genomic blocks [38].
• Alignment Filtering: Remove alignments with excessive missing data, low information content, or evidence of recombination [38].
• Gene Tree Inference: Reconstruct phylogenetic trees for each alignment using maximum likelihood (IQ-TREE) or Bayesian methods [38].
• Species Tree Estimation: Infer the primary species tree from the gene tree distribution using ASTRAL or similar methods [38].
• Topology Frequency Analysis: Quantify frequencies of alternative topologies and identify statistically supported excesses of particular discordant trees.
• Network Inference: Use tools like PhyloNet to infer phylogenetic networks that explicitly model introgression events [38].

Phylogenetic Network Inference: The ALTS Algorithm

Principles of Operation: The ALTS (Alignment of Lineage Taxon Strings) algorithm infers phylogenetic networks by aligning lineage taxon strings computed from input trees with respect to taxon ordering [6]. This approach reduces the network inference problem to finding common supersequences of lineage taxon strings across multiple gene trees. The algorithm searches for the minimum tree-child network that displays all input trees by checking possible orderings on the taxon set [6].

Tree-child networks are a specific class of phylogenetic networks where every non-leaf node has at least one child that is not a reticulate node. This constraint ensures biological plausibility while allowing efficient computation. The method aims to find networks with minimal hybridization number, representing the most parsimonious explanation for observed gene tree discordance [6].

Detailed Experimental Protocol:

• Input Tree Collection: Obtain a set of gene trees inferred from genomic sequences.
• Taxon Ordering Generation: Generate multiple possible orderings of the taxon set.
• Lineage Taxon String Computation: For each tree and ordering, compute lineage taxon strings representing ancestral relationships [6].
• Common Supersequence Identification: Find common supersequences of lineage taxon strings across all input trees.
• Network Construction: Build tree-child networks using the Tree-Child Network Construction algorithm based on identified supersequences [6].
• Network Selection: Choose the network with minimal hybridization number that displays all input trees.
• Validation: Assess network fit using statistical measures and compare to alternative topologies.

Supervised Learning: Emerging Pattern Recognition Approaches

Principles of Operation: Supervised learning approaches frame introgression detection as a classification task, where algorithms learn to distinguish introgressed from non-introgressed loci based on training data [23]. These methods leverage multiple genomic features simultaneously, including local ancestry patterns, haplotype structure, allele frequency spectra, and linkage disequilibrium. When detection is framed as a semantic segmentation task, these methods can precisely identify introgressed loci and their boundaries [23].

These approaches are particularly powerful for identifying adaptive introgression, where selection creates distinct genomic signatures including reduced diversity, specific haplotype patterns, and elevated differentiation in specific regions. Training on simulated data with known introgression parameters allows the algorithm to learn complex, multi-dimensional signatures of introgression.

Performance Comparison: Experimental Data and Benchmarking

Quantitative Performance Metrics Across Method Categories

Table 1: Performance Comparison of Introgression Detection Methods

Method Category	Representative Tools	Detection Power	False Positive Rate	Computational Efficiency	Optimal Application Context
Summary Statistics	D-suite, f-statistics	Moderate to High [59]	Low [83]	High [83]	Recent introgression, large sample sizes
Tree-Based Methods	ASTRAL, PhyloNet, IQ-TREE	High [38]	Low to Moderate	Moderate [38]	Deep introgression, incomplete lineage sorting
Phylogenetic Networks	ALTS, HYBRIDIZATION NUMBER	High for known trees [6]	Low [6]	Varies with complexity [6]	Complex reticulation, multiple introgressions
Supervised Learning	VolcanoFinder, Genomatnn, MaLAdapt	Varies by scenario [59]	Varies by scenario [59]	Moderate to High [23]	Adaptive introgression, large genomic datasets

Scenario-Specific Performance Evaluation

Recent benchmarking studies have revealed significant performance variation across evolutionary scenarios. A comprehensive evaluation of adaptive introgression methods tested VolcanoFinder, Genomatnn, and MaLAdapt across evolutionary scenarios inspired by human, wall lizard (Podarcis), and bear (Ursus) lineages [59]. These lineages represent different combinations of divergence and migration times, providing insights into method performance across parameter space.

The study found that methods based on the Q95 statistic demonstrated the highest efficiency for exploratory studies of adaptive introgression [59]. Performance was significantly influenced by evolutionary parameters including:

• Divergence and migration times: Methods showed variable performance across different combinations of these parameters [59].
• Selection coefficient: Stronger selection improved detection power for adaptively introgressed loci [59].
• Recombination hotspots: Presence of recombination affected localization accuracy of introgressed segments [59].
• Adjacent window effects: Including adjacent windows in training data was crucial for correct identification of the specific window containing the adaptive mutation [59].

Table 2: Specialized Method Performance in Specific Biological Contexts

Method	Biological Context	Key Performance Metrics	Notable Advantages	Identified Limitations
VolcanoFinder	Adaptive Introgression	Power varies with selection strength [59]	Specialized for selection signatures	Performance depends on demographic scenario [59]
Genomatnn	General & Adaptive Introgression	Varies across tested scenarios [59]	Neural network approach	Requires appropriate training data [59]
MaLAdapt	Adaptive Introgression	Scenario-dependent performance [59]	Machine learning framework	Sensitive to parameter tuning [59]
ALTS	Tree-Child Network Inference	Handles 50 trees with 50 taxa in ~15 minutes [6]	Scalable to larger datasets	Limited to tree-child networks [6]

Table 3: Essential Research Reagents and Computational Tools for Introgression Detection

Tool/Resource	Category	Primary Function	Application Context
IQ-TREE	Phylogenetic Inference	Maximum likelihood tree estimation [38]	Gene tree estimation for tree-based methods
ASTRAL	Species Tree Estimation	Species tree from gene trees [38]	Primary species tree inference accounting for ILS
PhyloNet	Network Inference	Phylogenetic network inference [38]	Reticulate evolution modeling
PAUP*	Phylogenetic Analysis	General utility phylogenetic inference [38]	Tree inference and manipulation
Progressive Cactus	Genome Alignment	Whole-genome alignment [38]	Input data preparation for tree-based methods
ALTS	Network Inference	Tree-child network from gene trees [6]	Scalable network inference for larger datasets
D-suite	Summary Statistics	D-statistic calculations [83]	ABBA-BABA tests for introgression signals
VolcanoFinder	Adaptive Introgression	Detection of selected introgression [59]	Adaptive introgression identification

Integrated Workflow for Robust Introgression Characterization

Based on comparative performance data, an integrated approach provides the most robust framework for introgression detection:

• Initial Screening with Summary Statistics: Deploy D-statistics for genome-wide scanning to identify candidate introgressed regions [83]. This computationally efficient approach provides initial hypotheses about introgression presence and direction.

• Tree-Based Validation: Apply tree-based methods to regions identified by summary statistics to validate signals using independent phylogenetic principles [38]. This step helps distinguish introgression from other sources of genealogical discordance.

• Network Modeling for Complex Scenarios: Implement phylogenetic network approaches like ALTS when multiple introgression events or complex reticulation patterns are suspected [6]. This is particularly valuable in rapidly radiating lineages.

• Supervised Learning for Adaptive Introgression: Apply specialized tools like VolcanoFinder or MaLAdapt when seeking adaptively introgressed loci, using appropriate training data that includes adjacent genomic windows [59].

This integrated framework leverages the complementary strengths of each approach while mitigating their individual limitations, providing a robust strategy for accurate introgression characterization across diverse evolutionary scenarios.

Conclusion

The accurate characterization of introgression using phylogenetic networks requires careful consideration of both biological processes and methodological limitations. While current methods have significantly advanced our ability to detect historical gene flow, substantial challenges remain in scalability, distinguishing introgression from incomplete lineage sorting, and managing computational demands. The integration of summary statistics, model-based approaches, and emerging machine learning techniques provides a powerful framework for robust inference. For biomedical research, these advances enable more precise evolutionary reconstructions of pathogen evolution, antibiotic resistance gene transfer, and host-pathogen coevolution. Future directions should focus on developing more scalable algorithms, improving model selection frameworks, and creating standardized validation protocols to ensure biological insights translate reliably into clinical and drug development applications.

References

• 1. Distinguishing between Incomplete Lineage Sorting and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4074047/]
• 2. The contribution of incomplete lineage sorting and ... [https://www.sciencedirect.com/science/article/abs/pii/S1055790322002068]
• 3. Detecting adaptive introgression in human evolution using ... [https://elifesciences.org/articles/64669]
• 4. Hibbins | Distinguishing Between Histories of Speciation ... [https://ssbbulletin.org/article/id/4476/]
• 5. Powerful methods for detecting introgressed regions from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4899106/]
• 6. A fast and scalable method for inferring phylogenetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10538497/]
• 7. Recombination-aware phylogenomics [https://www.sciencedirect.com/science/article/abs/pii/S016953472500179X]
• 8. Teasing apart the sources of phylogenetic tree discordance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12269116/]
• 9. Contemporary Challenges in Large-Scale Sequence ... [https://www.imsi.institute/activities/contemporary-challenges-in-large-scale-sequence-alignments-and-phylogenies/]
• 10. Streamlining genetic analysis for phylogenetic studies [https://www.sciencedaily.com/releases/2025/01/250123110231.htm]
• 11. Factors influencing the accuracy and precision in dating ... [https://peercommunityjournal.org/articles/10.24072/pcjournal.556/]
• 12. Multispecies coalescent and its applications to infer ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8692950/]
• 13. A multispecies coalescent model for quantitative traits [https://elifesciences.org/articles/36482]
• 14. Phylogenetic networks empower biodiversity research - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC12337313/]
• 15. Inference of Phylogenetic Networks From Sequence Data ... [https://pubmed.ncbi.nlm.nih.gov/39387633/]
• 16. Quantitatively defining species boundaries with more ... [https://www.nature.com/articles/s42003-022-03723-z]
• 17. Determining the probability of hemiplasy in the presence ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7800383/]
• 18. Phylotranscriptomic discordance is best explained by ... [https://www.sciencedirect.com/science/article/pii/S2468265923001038]
• 19. Historical introgression as a driver of diversification of diploid ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12604549/]
• 20. Quartets enable statistically consistent estimation of cell ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10691101/]
• 21. Quartet-based methods to reconstruct phylogenetic networks [https://bmcsystbiol.biomedcentral.com/articles/10.1186/1752-0509-8-21]
• 22. A quartet-based approach for inferring phylogenetically ... [https://www.sciencedirect.com/science/article/pii/S200103702500337X]
• 23. Review Decoding genomic landscapes of introgression [https://www.sciencedirect.com/science/article/pii/S0168952525001660]
• 24. Dsuite - fast D-statistics and related admixture evidence ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7116594/]
• 25. D-statistics for Dummies: A simple test for introgression [https://avianhybrids.wordpress.com/2019/11/09/d-statistics-for-dummies-a-simple-test-for-introgression/]
• 26. Gene flow analysis method, the D-statistic, is robust in a wide ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-017-2002-4]
• 27. Detecting Introgression in Shallow Phylogenies [https://pmc.ncbi.nlm.nih.gov/articles/PMC12485364/]
• 28. Introgression impacts the evolution of bacteria, but species ... [https://www.nature.com/articles/s41467-025-64947-1]
• 29. Signatures of Introgression across the Allele Frequency ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7826190/]
• 30. A Test Statistic for Detection of General Introgression Scenarios [https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.WABI.2018.19]
• 31. Phylogenomics yields new systematic and taxonomical ... [https://www.sciencedirect.com/science/article/pii/S1055790324002744]
• 32. Multispecies coalescent process [https://en.wikipedia.org/wiki/Multispecies_coalescent_process]
• 33. NANUQ+: A divide-and-conquer approach to network estimation [https://almob.biomedcentral.com/articles/10.1186/s13015-025-00274-w]
• 34. Comparing inference under the multispecies coalescent ... [https://www.sciencedirect.com/science/article/abs/pii/S1055790323000246]
• 35. A scalability study of phylogenetic network inference methods ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-1277-1]
• 36. A scalability study of phylogenetic network inference ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5064893/]
• 37. Phylogenetic network analysis as a parsimony optimization ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-015-0675-0]
• 38. - Evolution and Introgression detection Genomics [https://evomics.org/learning/population-and-speciation-genomics/2025-population-and-speciation-genomics/introgression-detection/]
• 39. Genome scans for selection and introgression based on k‐ ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7689739/]
• 40. Using Machine to Analyze Genetic Data - Simple Science Learning [https://scisimple.com/en/articles/2025-07-10-using-machine-learning-to-analyze-genetic-data--a98464d]
• 41. IntroUNET: Identifying introgressed alleles via... | PLOS Genetics [https://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1010657]
• 42. An HMM-Based Comparative Genomic Framework for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4055573/]
• 43. Phylogenetic analysis of combined mitochondrial genome ... [https://www.zoores.ac.cn/article/cstr/32248.14.zr.2022.081]
• 44. Population Genetics Provides Insights Into the Impact of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11968420/]
• 45. PhyloScape: interactive and scalable visualization platform for ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-025-06203-3]
• 46. Assessing phylogenetic confidence at pandemic scales [https://www.nature.com/articles/s41586-025-09567-x]
• 47. Analyzing the relationship between sequence divergence ... [https://www.sciencedirect.com/science/article/abs/pii/S1055790310002289]
• 48. A sequence-based evolutionary distance method for ... [https://www.nature.com/articles/s41598-023-47496-9]
• 49. Structural phylogenetics unravels the evolutionary ... [https://www.nature.com/articles/s41594-025-01649-8]
• 50. Assessing the effect of varying sequence length on DNA ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC1890918/]
• 51. Evolutionary Rate Variation among Lineages in Gene ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8830059/]
• 52. Machine learning based imputation techniques for estimating ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-020-06892-5]
• 53. Determining the probability of hemiplasy in the presence ... [https://elifesciences.org/articles/63753]
• 54. On the inference of complex phylogenetic networks by Markov ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8445492/]
• 55. “Correcting” Gene Trees to be More Like Species ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10263008/]
• 56. [2511.03692] Improving Gene Trees without more data [https://arxiv.org/abs/2511.03692]
• 57. Gene Presence–Absence Outperforms Gene Trees [https://pmc.ncbi.nlm.nih.gov/articles/PMC12301182/]
• 58. PhyloTune: An efficient method to accelerate phylogenetic ... [https://www.nature.com/articles/s41467-025-61684-3]
• 59. Performance evaluation of adaptive introgression ... [https://peercommunityjournal.org/articles/10.24072/pcjournal.617/]
• 60. Model-based detection and analysis of introgressed ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6165692/]
• 61. Model design choices impact biological insight [https://pmc.ncbi.nlm.nih.gov/articles/PMC10954156/]
• 62. Principles of Effective Data Visualization - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7733875/]
• 63. Dryad | Data : Massively scalable inference of level-1 phylogenetic ... [https://datadryad.org/dataset/doi:10.5061/dryad.h44j0zq0b]
• 64. GitHub - citiususc/veryfasttree: Efficient phylogenetic tree inference... [https://github.com/citiususc/veryfasttree]
• 65. Online Phylogenetics using Parsimony Produces Slightly ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9128781/]
• 66. Inferring Phylogenetic Networks with Maximum ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4780787/]
• 67. Phylogenomic approach reveals strong signatures of ... [https://www.sciencedirect.com/science/article/pii/S1055790321001330]
• 68. An evolutionary algorithm based on parsimony for the ... [https://www.sciencedirect.com/science/article/abs/pii/S1568494623002880]
• 69. Phylogenomic approaches to detecting and characterizing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9208645/]
• 70. A maximum pseudo-likelihood approach for phylogenetic ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/1471-2164-16-S10-S10]
• 71. Comparing rates of introgression in parasitic feather lice ... [https://www.nature.com/articles/s42003-020-01345-x]
• 72. Characterizing semi-directed phylogenetic networks and ... [https://arxiv.org/html/2507.18772v2]
• 73. Flexible and efficient simulation-based inference for models of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9374439/]
• 74. List of phylogenetics software [https://en.wikipedia.org/wiki/List_of_phylogenetics_software]
• 75. 91 Free Phylogenetics Tools - Software and Resources [https://bioinformaticshome.com/db/collection/phylogenetics]
• 76. Free Phylogenetic Network Software - fluxus-engineering.com [https://www.fluxus-engineering.com/sharenet.htm]
• 77. A Comparison of Phylogenetic Network Methods Using ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0001913]
• 78. EDD-NSTE: Edge Data Distribution as a Network Steiner ... [https://arxiv.org/abs/2504.20738]
• 79. Universal orthologs infer deep phylogenies and improve ... [https://bmcbiol.biomedcentral.com/articles/10.1186/s12915-025-02328-2]
• 80. A large-scale benchmark for network inference from single- ... [https://www.nature.com/articles/s42003-025-07764-y]
• 81. A systematic benchmark of integrative strategies for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12289951/]
• 82. Effects of dataset characteristics on the performance ... [https://www.sciencedirect.com/science/article/abs/pii/S156849461630641X]
• 83. and Polarization of Detection in a Five-Taxon Phylogeny Introgression [https://pubmed.ncbi.nlm.nih.gov/25888025/]