Discrete Trait Analysis vs. Structured Birth-Death Models: A Comprehensive Guide for Biomedical Researchers

Abstract

This article provides a comprehensive comparison of Discrete Trait Analysis (DTA) and Structured Birth-Death Models (SBDM), two foundational methods in phylogenetic inference for studying trait evolution and population dynamics. Tailored for researchers, scientists, and drug development professionals, it covers the core principles, methodological applications, and practical challenges of both approaches. Drawing on current research and software advancements, the guide offers a clear framework for selecting and optimizing these models to track pathogen spread, quantify transmission dynamics, and inform public health strategies, with a focus on real-world use cases in infectious disease and genomic epidemiology.

Core Concepts: Understanding Discrete Trait Analysis and Structured Birth-Death Models

Discrete Trait Analysis (DTA) is a phylogenetic comparative method used to infer the evolutionary history of discrete characteristics—such as geographic location, disease state, or morphological feature—across a phylogeny. In essence, DTA treats these traits as if they were evolutionary characters that can "mutate" from one state to another (e.g., from geographic region A to region B) along the branches of a tree [1] [2]. This approach allows researchers to reconstruct the ancestral states of these traits at internal nodes of the phylogeny, providing insights into historical evolutionary processes, migration patterns, and trait associations.

The method operates by modeling trait evolution using a continuous-time Markov chain (CTMC), typically defined by a transition rate matrix (Q-matrix) that describes the rate of change between all possible pairs of discrete states [3]. DTA has become a widely used technique in fields ranging from viral phylogeography, where it helps trace the spread of pathogens, to macroevolution, where it investigates correlations between phenotypic traits [4] [1]. Its popularity stems from its computational efficiency and intuitive analogy to substitution processes in molecular evolution, though this very analogy also underpins its primary limitations when applied to population-level processes such as migration [2].

Core Principles and Methodological Framework

The Statistical Foundation of DTA

The statistical engine of Discrete Trait Analysis is a Markov model that describes the instantaneous rates of change between discrete character states. The core component is the Q-matrix, a square matrix where each off-diagonal element qij represents the instantaneous rate of change from state i to state j. The diagonal elements are set such that each row sums to zero, ensuring proper probabilistic interpretation [3]. The likelihood of observing a particular pattern of trait evolution across a phylogeny can be calculated by considering the product of probabilities along all branches, integrating over all possible ancestral states at internal nodes.

Model selection plays a crucial role in DTA, with researchers typically comparing different structures of the Q-matrix:

• Equal Rates (ER): All transition rates are equal
• Symmetric (SYM): Forward and backward rates between any two states are equal
• All Rates Different (ARD): Each possible transition has a distinct rate parameter [3]

The choice among these models depends on biological rationale and can be evaluated using statistical criteria such as Akaike Information Criterion (AIC) or Bayesian model comparison [3] [5].

Implementation Workflows

The practical implementation of DTA typically follows a structured workflow, whether for ancestral state reconstruction or phylogeographic analysis:

Data Preparation: The process begins with assembling two core components: a phylogenetic tree with branch lengths (typically time-calibrated) and a dataset of discrete traits for the tips of the tree. These traits must be carefully coded into discrete states (e.g., 0/1 for binary traits, or specific labels for multi-state traits) [3] [5].

Model Specification and Fitting: Researchers specify the structure of the transition rate matrix based on biological hypotheses or employ model selection to determine the best-fitting structure. The model is then fitted to the data using maximum likelihood or Bayesian inference methods [5].

Ancestral State Reconstruction: Once the model is fitted, marginal or joint ancestral state reconstruction is performed to estimate the probability of each discrete state at internal nodes of the phylogeny [5].

Visualization and Interpretation: The results are typically visualized by projecting the reconstructed states onto the phylogeny, often using color coding or other visual cues to represent state changes across evolutionary history [3].

The following diagram illustrates this generalized workflow for a DTA analysis:

DTA Versus Alternative Phylogeographic Models

Conceptual Comparison of Modeling Approaches

Discrete Trait Analysis represents just one approach to modeling trait evolution and population history. When compared with structured coalescent and birth-death models, fundamental differences emerge in their underlying assumptions and mathematical foundations.

The table below summarizes the key distinctions between these approaches:

Feature	Discrete Trait Analysis (DTA)	Structured Coalescent Models	Structured Birth-Death Models
Core Concept	Treats location/trait as evolving like a discrete character [2]	Models genealogy within structured population considering lineage migration [2]	Models lineage birth/death events across different structured populations [4]
Computational Demand	Low to moderate [2]	High to very high [4] [2]	Moderate to high [4]
Handling of Sampling Bias	Sensitive to uneven sampling across states [6] [2]	Better accounts for variable sampling intensity [2]	Can incorporate sampling proportions [4]
Population Size Inference	Not directly inferred	Infers effective population sizes per deme [2]	Infers birth, death, and sampling rates [4]
Typical Applications	Discrete trait evolution, phylogeography with limited demes [4] [2]	Accurate migration history, outbreak source attribution [2]	Emerging outbreak dynamics, serially sampled data [4]
Key Limitations	Assumes independence of trait evolution from tree-generating process [4] [2]	Computationally intensive with many demes [4] [2]	May require strong priors for convergence [4]

Performance Comparison: Empirical Evidence

Experimental comparisons between DTA and structured coalescent models reveal significant performance differences, particularly in scenarios involving biased sampling or outbreak origin estimation.

A seminal study by Dellicour et al. highlighted these disparities through simulations and empirical analyses [2]. When investigating the zoonotic transmission of Ebola virus, DTA implausibly suggested sustained undetected human-to-human transmission over four decades, while the structured coalescent analysis correctly identified repeated seeding from a large unsampled non-human reservoir population [2]. This case exemplifies how model misspecification in DTA can lead to fundamentally incorrect biological conclusions.

Another critical evaluation focused on root state classification accuracy—the ability to correctly infer the geographic origin of an outbreak at the root of the phylogeny [6]. This research demonstrated that DTA performance peaks at intermediate sequence dataset sizes and that common metrics like Kullback-Leibler divergence can provide misleading support for models with finer discretization schemes, unrelated to actual classification accuracy [6].

The following diagram illustrates the fundamental conceptual differences in how these models approach lineage history:

Experimental Protocols and Validation Frameworks

Standard DTA Implementation Protocol

A robust DTA implementation for ancestral state reconstruction typically follows this experimental protocol:

Data Curation and Alignment:

• Obtain or infer a time-calibrated phylogenetic tree with branch lengths
• Code discrete traits consistently across all tips
• Ensure trait data alignment with tree tip labels
• Address missing data using appropriate methods (e.g., partial assignment or modeling uncertainty) [3] [5]

Model Selection and Fitting:

• Specify multiple candidate Q-matrix structures (ER, SYM, ARD)
• Fit models using maximum likelihood or Bayesian inference
• Conduct statistical model comparison using AIC, BIC, or Bayes factors
• Select best-fitting model for subsequent reconstruction [3] [5]

Ancestral State Reconstruction:

• Perform marginal ancestral state estimation using the selected model
• Calculate state probabilities at all internal nodes
• Optionally, perform joint reconstruction to identify the most probable overall history
• Assess uncertainty in reconstructions through bootstrap or Bayesian credibility intervals [5]

Validation and Sensitivity Analysis:

• Conduct posterior predictive simulations to assess model adequacy
• Perform sensitivity analyses on key prior assumptions
• Validate reconstructions using known historical events or fossil data when available [6] [5]

Performance Evaluation Experiments

To quantitatively evaluate DTA performance against alternative methods, researchers have developed standardized simulation protocols:

Root State Classification Accuracy:

• Simulate phylogenetic trees under known migration models
• Generate trait data with known root states
• Apply DTA and structured models to reconstruct root states
• Compare accuracy rates across methods and conditions [6]

Sampling Bias Sensitivity Assessment:

• Simulate data with controlled sampling inequalities across demes
• Systematically vary the degree of sampling bias
• Measure reconstruction accuracy degradation under bias
• Compare robustness between DTA and structured approaches [2]

Migration Rate Estimation Precision:

• Simulate data with known migration rate matrices
• Recover rates using different inference methods
• Calculate deviation between true and estimated parameters
• Assess statistical consistency and precision across methods [2]

Essential Research Toolkit for DTA

Successful implementation of Discrete Trait Analysis requires familiarity with both conceptual frameworks and practical software tools. The following table outlines key resources in the DTA researcher's toolkit:

Tool Category	Specific Software/Package	Primary Function	Key Considerations
Bayesian Evolutionary Analysis	BEAST 2 [4] [7]	Bayesian phylogenetic inference with discrete trait models	Supports both DTA and structured coalescent approximations; packages include BEASTCLASSIC, MASCOT, GEOSPHERE [4]
R Comparative Methods Packages	phytools [3] [5]	Phylogenetic comparative methods including ancestral state reconstruction	Provides functions for plotting, simulation, and model fitting; integrates well with other R packages [5]
R Comparative Methods Packages	corHMM [3]	Hidden Markov Models for phylogenetic comparative analysis	Specializes in correlated trait evolution; efficient for complex model fitting [3]
Model Selection Frameworks	AIC/BIC [3]	Statistical model comparison	Standard approach for comparing DTA model structures (ER, SYM, ARD) [3]
Simulation Tools	Phylogenetic simulation packages	Generating synthetic data under known models	Essential for method validation and power analysis [6] [2]

Discrete Trait Analysis represents a powerful but nuanced approach for investigating the evolutionary history of discrete characteristics across phylogenies. Its computational efficiency and intuitive framework make it well-suited for exploratory analyses, discrete phenotypic trait evolution, and situations with limited computational resources or few discrete states [4] [3]. However, the method's sensitivity to sampling bias and its fundamental assumption of independence between trait evolution and the tree-generating process necessitate careful application [2].

For researchers studying population-level processes such as migration or epidemic spread, structured coalescent models or their approximations (e.g., BASTA) generally provide more accurate inference, particularly when sampling is uneven or the number of demes is manageable [2]. The emerging generation of phylogenetic software, including BEAST X with its Hamiltonian Monte Carlo samplers, promises to reduce the computational barriers to these more sophisticated approaches [7].

Ultimately, method selection should be guided by biological context, sampling structure, and inferential goals. Discrete Trait Analysis remains a valuable tool in the phylogenetic toolkit when applied judiciously to questions aligned with its theoretical foundations and with appropriate caveats regarding its limitations.

Structured Birth-Death Models (SBDMs) represent a significant advancement in phylodynamic analysis by integrating population dynamics directly with lineage sorting in structured populations. Unlike approaches that treat discrete traits as independently evolving characters, SBDMs explicitly model how birth (speciation), death (extinction), and migration processes between subpopulations shape phylogenetic trees. This comprehensive analysis compares SBDMs against discrete trait analysis (DTA), examining their theoretical foundations, performance characteristics, and practical applications through current experimental data and case studies. The findings demonstrate that while DTA offers computational efficiency, SBDMs provide superior accuracy for inferring migration history and root state locations, particularly in epidemiological investigations and evolutionary studies of pathogens.

Phylodynamic methods aim to quantify past population dynamics from genetic sequencing data, with particular importance for understanding the spread of infectious diseases in structured populations [8]. When analyzing pathogens, the host population may be geographically structured, or the pathogen population may consist of different subpopulations, such as drug-sensitive and drug-resistant variants. Understanding how these subpopulations interact—whether separated by geographic distance, host characteristics, or other barriers—represents a key determinant in understanding how epidemics spread and evolve [8].

Two primary classes of models exist for phylodynamic analysis of structured populations: structured birth-death models (SBDMs) and discrete trait analysis (DTA). These approaches differ fundamentally in their theoretical foundations and biological assumptions. SBDMs, implemented in packages such as BDMM (Birth-Death Mixture Model) for BEAST2, are based on birth-death processes that explicitly model speciation, extinction, and migration rates between demes (subpopulations) [9] [8]. In contrast, DTA treats sampling locations as discrete traits that evolve along branches of the phylogenetic tree in a manner analogous to the substitution of alleles at a genetic locus, often described as the "Mugration" model [2].

The core distinction lies in how each approach integrates the tree-generating process with migration dynamics. SBDMs incorporate migration directly into the population dynamic process that generates the tree, while DTA models migration as a separate process occurring upon an already-existing tree [4] [2]. This fundamental difference has profound implications for model accuracy, computational requirements, and appropriate application domains.

Theoretical Foundations and Model Specifications

Structured Birth-Death Models: Mathematical Framework

Structured Birth-Death Models are continuous-time Markov processes that track the number of individuals in different subpopulations through time [10] [11]. In macroevolution and epidemiology, these "individuals" typically represent species or infected hosts. The model defines several key parameters operating within and between d discrete types (demes):

• Birth rate (λ): The per-lineage rate of speciation or infection generation
• Death rate (μ): The per-lineage rate of extinction or recovery
• Migration rate (m): The rate at which lineages move between demes

In the multi-type birth-death model with sampling as implemented in the BDMM package, the process begins at time 0 (the origin) with one individual of type i with probability hᵢ [8]. The time interval (0,T) is partitioned into n epochs through time points 0 < t₁ < ... < tₙ₋₁ < T, allowing rate parameters to change at predefined intervals. Each individual of type i at time t (where tₖ₋₁ ≤ t < tₖ) gives birth to a new individual of type j at rate λᵢⱼ,ₖ, migrates to type j at rate mᵢⱼ,ₖ (with mᵢᵢ,ₖ = 0), dies at rate μᵢ,ₖ, and is sampled at rate ψᵢ,ₖ [8]. At specific sampling times tₖ, each individual of type i is sampled with probability ρᵢ,ₖ. Upon sampling, individuals are removed from the infectious pool with probability rᵢ,ₖ [8].

The probability density of the resulting sampled phylogeny is computed by numerically integrating a system of differential equations backward along all branches to the origin of the tree [8]. This computation involves calculating the probability flow through the tree while accounting for all possible migration histories and population dynamics.

Table 1: Key Parameters in Structured Birth-Death Models

Parameter	Symbol	Description	Units
Speciation/Birth Rate	λᵢⱼ,ₖ	Rate at which lineage in deme i gives birth to lineage in deme j during epoch k	events/time
Extinction/Death Rate	μᵢ,ₖ	Rate at which lineages in deme i are lost during epoch k	events/time
Migration Rate	mᵢⱼ,ₖ	Rate at which lineages migrate from deme i to deme j during epoch k	events/time
Sampling Rate	ψᵢ,ₖ	Rate at which lineages in deme i are sampled through time during epoch k	events/time
Sampling Probability	ρᵢ,ₖ	Probability of sampling lineages in deme i at time tₖ during epoch k	dimensionless
Removal Probability	rᵢ,ₖ	Probability that sampling removes lineage from infectious pool	dimensionless

Discrete Trait Analysis: Underlying Assumptions

Discrete Trait Analysis (DTA) operates on fundamentally different principles from SBDMs. In DTA, the geographic location or other discrete trait of interest is treated as a character state that evolves along the branches of a phylogenetic tree according to a continuous-time Markov process [2]. The model assumes that:

• The relative size of subpopulations drifts over time, such that subpopulations can become lost (extinct) or fixed (the sole remaining subpopulation) without constraints
• Sample sizes across subpopulations are proportional to their relative sizes
• The migration process is conceptually separated from the coalescent process

The DTA model inherits assumptions appropriate for the independent mutation of loci within lineages but profoundly at odds with classical population genetics models of migration [2]. Specifically, it does not account for the effects of population structure on the coalescent process itself, treating the tree as fixed rather than shaped by the population dynamics it aims to infer.

Performance Comparison: Experimental Data and Case Studies

Quantitative Performance Metrics

Recent empirical studies have directly compared the performance of SBDMs and DTA across multiple metrics, revealing significant differences in accuracy, computational efficiency, and robustness to sampling bias.

Table 2: Performance Comparison of SBDM vs. DTA

Performance Metric	Structured Birth-Death Models	Discrete Trait Analysis
Root State Classification	Higher accuracy, particularly with intermediate sequence dataset sizes [6]	Lower accuracy, sensitive to sampling bias [6] [2]
Migration Rate Estimation	More accurate across simulated and empirical datasets [2]	Often inaccurate, particularly with biased sampling [2]
Computational Efficiency	More demanding, especially with many demes [4] [8]	Faster computation, enabling analysis of large datasets [4] [2]
Sampling Bias Sensitivity	Robust to uneven sampling across demes [2]	Highly sensitive to uneven sampling [2]
Maximum Dataset Size	~250 sequences in initial implementation, now improved to 500+ [8]	Effectively unlimited with sufficient computational resources
Theoretical Foundation	Based on population genetics principles [8] [2]	Based on phylogenetic character evolution [2]

Case Study: Ebola Virus Transmission Dynamics

A compelling illustration of the practical implications of model choice comes from the analysis of Ebola virus genomic data [2]. When investigating the zoonotic transmission of Ebola virus, structured coalescent methods (conceptually similar to SBDMs) correctly inferred that successive human Ebola outbreaks were seeded by a large unsampled non-human reservoir population. In contrast, the discrete trait analysis implausibly concluded that undetected human-to-human transmission had allowed the virus to persist over the past four decades [2].

These diametrically opposed conclusions have significant implications for public health policy and intervention strategies. The DTA results would suggest focusing resources on detecting and interrupting human transmission chains, while the SBDM results correctly highlight the importance of understanding and monitoring the animal reservoir to prevent future spillover events. This case study underscores how model misspecification in phylogeographic analyses can lead to fundamentally incorrect inferences with real-world consequences.

Algorithmic Improvements in BDMM

Recent algorithmic enhancements to the BDMM package have substantially improved its practical utility. Initial versions were limited to analyzing datasets of approximately 250 genetic sequences due to numerical instability caused by underflow in probability density calculations [8]. Important algorithmic changes have dramatically increased the number of genetic samples that can be analyzed while improving numerical robustness and computational efficiency [8].

These improvements allow for enhanced precision of parameter estimates, particularly for structured models with a high number of inferred parameters. Additional model extensions include support for homochronous sampling events at multiple time points (not only the present), removal of the requirement that individuals are necessarily removed upon sampling, and more flexible migration rate specification through piecewise-constant changes through time [8].

Experimental Protocols and Methodologies

Standard Implementation of SBDM Analysis

The implementation of Structured Birth-Death Models using the BDMM package in BEAST2 follows a standardized workflow with specific requirements at each stage:

Software Requirements:

• BEAST2 v2.7.4 or later for Bayesian evolutionary analysis
• BEAUti2 for generating XML configuration files
• BDMM package (with automatic installation of MultiTypeTree and MASTER dependencies)
• Tracer v1.7.2 for parameter analysis
• TreeAnnotator for summary tree production
• IcyTree for tree visualization [9]

Data Preparation Protocol:

• Sequence data in FASTA format with labels containing sampling metadata
• Temporal information encoded as last element in underscore-delimited sequence names
• Location data encoded as second element in underscore-delimited sequence names
• Setting tip dates through the "Use tip dates" option with "after last" delimiter configuration
• Configuring tip locations using the "Guess" function with "split on character" and group 2 selection [9]

Model Configuration Specifications:

• Substitution Model: JC69 with 4 gamma categories for rate variation
• Gamma Category Count: 4 for discrete gamma approximation
• Shape Parameter: Estimated with starting value of 1.0
• Proportion of Invariant Sites: Fixed to alignment-specific value (e.g., 0.867)
• Clock Model: Strict clock with rate set to 0.005 substitutions/site/year
• Tree Prior: Multi-type birth-death model with appropriate epoch specification [9]

MCMC Analysis Parameters:

• Chain length: 10,000,000 to 100,000,000 steps depending on dataset size
• Log parameters: Every 1,000 to 10,000 steps
• Burn-in: 10% of chain length
• Convergence assessment: Effective Sample Size (ESS) > 200 for all parameters [9]

Model Selection and Validation Framework

Robust comparison between SBDM and DTA approaches requires a systematic validation framework:

Simulation-Based Calibration:

• Simulate phylogenetic trees under known birth-death parameters
• Generate sequence data along simulated trees using appropriate substitution models
• Analyze simulated data with both SBDM and DTA approaches
• Compare estimated parameters to known values to assess accuracy and bias [2]

Empirical Data Benchmarking:

• Select empirical datasets with known epidemiological history
• Analyze with both SBDM and DTA approaches
• Compare inferred parameters to historical records
• Assess consistency across multiple independent datasets [2]

Sensitivity Analysis Protocol:

• Vary sampling schemes across demes to assess robustness
• Test different epoch configurations for time-varying parameters
• Evaluate prior sensitivity through alternative prior specifications
• Assess convergence across multiple independent MCMC runs [8] [6]

Visualization and Workflow Diagrams

Structured Birth-Death Model Workflow

SBDM Analysis Workflow: This diagram illustrates the standard workflow for implementing Structured Birth-Death Models in BEAST2, from data preparation through final visualization.

Model Comparison Framework

Model Comparison Framework: This diagram contrasts the theoretical foundations, applications, performance characteristics, and limitations of SBDM (green) versus DTA (red) approaches.

Table 3: Essential Research Tools for SBDM Implementation

Tool/Resource	Function	Application Context
BEAST2 Platform	Bayesian evolutionary analysis using MCMC	Primary inference framework for SBDM and DTA [9]
BDMM Package	Implements multi-type birth-death model	Phylodynamic inference in structured populations [9] [8]
BEAUti2	Graphical configuration of BEAST2 XML files	Setting up analysis parameters and model specifications [9]
Tracer	MCMC diagnostics and parameter summary	Assessing convergence and summarizing posterior distributions [9]
TreeAnnotator	Summary tree production from posterior tree distribution	Generating maximum clade credibility trees [9]
MultiTypeTree Package	Defines colored trees for structured populations	Required dependency for BDMM analyses [9]
MASTER Package	Stochastic simulation of birth-death processes	Model validation and simulation-based calibration [9]
IcyTree	Browser-based phylogenetic tree visualization	Rapid visualization of phylogenetic trees with annotations [9]

Discussion and Future Directions

The comparative analysis presented here demonstrates that Structured Birth-Death Models and Discrete Trait Analysis represent fundamentally different approaches to phylogeographic inference with distinct strengths and limitations. SBDMs provide a more principled foundation based on population genetics principles and generally offer superior accuracy for inferring migration history and root state locations, particularly in scenarios with biased sampling across demes [2]. However, this accuracy comes at the cost of increased computational demands, which has historically limited applications to smaller datasets.

Recent algorithmic improvements to BDMM have substantially addressed these limitations, enabling analysis of datasets containing several hundred genetic sequences [8]. These advances, coupled with the development of approximate methods like BASTA (BAyesian STructured coalescent Approximation) that maintain accuracy while improving computational efficiency, suggest a promising trajectory for SBDM methodologies [2].

For researchers and drug development professionals, model selection should be guided by the specific research question and data characteristics. When accurate reconstruction of migration history and outbreak origins is paramount—particularly in public health contexts where inferences directly inform intervention strategies—SBDMs represent the preferred approach despite their computational demands. In exploratory analyses or applications where computational efficiency is a primary concern, DTA may still offer utility, though conclusions should be interpreted with appropriate caution regarding potential biases.

Future methodological development will likely focus on further improving the scalability of SBDMs to accommodate the increasingly large genomic datasets generated by modern surveillance systems, while maintaining the theoretical rigor and statistical accuracy that distinguish them from alternative approaches.

Evolutionary trees serve as the foundational scaffold for investigating the transmission dynamics and evolutionary history of pathogens. Within Bayesian phylogenetic software platforms like BEAST2, Discrete Trait Analysis (DTA) and Structured Birth-Death (SBD) models represent two principal approaches for leveraging these trees to understand spatial spread and population dynamics [12] [7]. While both methods operate on a phylogenetic tree, their core mechanisms, underlying assumptions, and susceptibility to bias differ significantly. This guide provides an objective comparison for researchers, scientists, and drug development professionals, focusing on their application in phylogeography and phylodynamics.

Core Methodological Comparison

The table below summarizes the fundamental characteristics of Discrete Trait Analysis and Structured Birth-Death models.

Table 1: Fundamental Comparison of Discrete Trait Analysis and Structured Birth-Death Models

Feature	Discrete Trait Analysis (DTA)	Structured Birth-Death (SBD) Models
Core Framework	Neutral trait evolution model mapped onto a fixed tree [12].	Tree-generating process; the tree is an output of the model itself [12].
Primary Output	History and rates of trait changes (e.g., location transitions) [12].	Population growth rates, transmission rates, and becoming uninfectious rates [12].
Key Assumption	Trait evolution does not influence the tree's branching structure [12].	Transmission dynamics directly shape the phylogenetic tree [12].
Handling of Bias	Can be sensitive to and produce biased results from uneven geographic sampling [12] [7].	Less subject to sampling biases; better accounts for population structure [12].
Computational Speed	Generally faster due to its conditional nature [12].	Typically more computationally intensive [12].

Experimental Protocols & Performance Data

Simulation-Based Validation of Methodological Performance

A key approach to evaluating these methods involves simulation studies, where "truth" is known. Researchers often use software like MASTER to simulate phylogenetic trees under a controlled structured coalescent model with predefined parameters, such as effective population size (N_e) trajectories and migration rates [12]. These simulated trees then serve as input for inference by both DTA and structured models (e.g., MASCOT-Skyline). Performance is quantified by how accurately each method recovers the known simulated parameters [12].

Table 2: Comparative Performance in Key Analytical Scenarios

Analysis Scenario	DTA Performance	Structured Model Performance	Supporting Evidence
Uneven Geographic Sampling	Biased reconstruction of migration rates and ancestral states [12].	Significantly more robust; mitigates bias by modeling population structure [12].	Simulation studies using SIR models and SARS-CoV-2 data [12].
Inferring Population Dynamics	Not a primary function; models trait evolution conditional on the tree.	Accurately retrieves non-parametric Ne trajectories over time in different locations [12].	Simulation of Ne trajectories from a Gaussian Markov Random Field (GMRF) [12].
Joint Inference	Not designed for joint spatio-temporal inference.	Jointly infers spatial transmission and temporal outbreak dynamics, improving accuracy for both [12].	Development of the MASCOT-Skyline method, which integrates both aspects [12].

Case Study: SARS-CoV-2 Omicron BA.1 Invasion

The application of these methods to real-world data is exemplified by studies of the SARS-CoV-2 Omicron variant. Research leveraging the advanced phylogeographic and phylodynamic models in BEAST X has traced the invasion of the Omicron BA.1 lineage in England [7]. Such analyses often employ discrete-trait phylogeography but are increasingly enhanced by models that parameterize transition rates between locations as functions of epidemiological predictors, helping to address inherent sensitivities to sampling bias [7].

The Scientist's Toolkit: Essential Research Reagents & Software

Successful phylodynamic analysis requires a suite of specialized software tools and computational resources.

Table 3: Key Reagents and Software for Phylogenetic Analysis

Item	Function	Relevance to Methods
BEAST2 / BEAST X	A cross-platform software platform for Bayesian evolutionary analysis sampling trees; the primary engine for inference [12] [7].	Core platform for implementing both DTA and structured models. BEAST X introduces newer, more scalable models [7].
MASCOT	A BEAST2 package implementing the Marginal Approximation of the Structured COalescenT [12].	Enables computationally efficient inference under the structured coalescent. MASCOT-Skyline adds time-varying dynamics [12].
MASTER	A software package for simulating stochastic phylogenetic trees under birth-death or coalescent models [12].	Used for validation studies and assessing model performance against known parameters [12].
BEAGLE	A high-performance computational library for phylogenetic inference [7].	Accelerates likelihood calculations for all models, enabling analysis of larger datasets [7].
HAMILTONIAN MONTE CARLO (HMC)	An advanced Markov chain Monte Carlo (MCMC) algorithm for sampling from complex, high-dimensional posterior distributions [7].	Implemented in BEAST X to improve inference efficiency for complex models like structured coalescents and relaxed random walks [7].

Visualizing Methodological Workflows

The following diagram illustrates the logical relationship and application focus of DTA and SBD models within a phylogenetic framework.

Diagram: Methodological Pathways. DTA and SBD models use the phylogenetic tree as input but answer different biological questions.

Discussion for Research Application

The choice between Discrete Trait Analysis and Structured Birth-Death models is not merely a technicality but a strategic decision that directly influences research conclusions. DTA offers a faster, more accessible path for initial phylogeographic reconstruction, making it suitable for exploratory analyses or when computational resources are limited. However, its known vulnerability to sampling bias necessitates cautious interpretation, particularly with unevenly sampled data [12]. In contrast, SBD models provide a more robust and mathematically coherent framework for questions where the transmission process itself is the primary object of study, as they explicitly model the processes that generate the tree [12]. They are essential for jointly inferring population dynamics and spatial spread, leading to more accurate parameter estimates. The advent of more scalable software like BEAST X and efficient algorithms like Hamiltonian Monte Carlo is making these more complex models increasingly practical for larger datasets [7]. For grant proposals or drug development research where understanding the precise dynamics of pathogen spread is critical, investing in the structured modeling approach is often the more rigorous and reliable choice.

Phylogeographic inference aims to reconstruct the spatial spread and population dynamics of pathogens using genetic sequence data. For researchers and drug development professionals, selecting the appropriate model is critical for accurately identifying outbreak origins and transmission patterns. Two principal methodologies dominate this field: Discrete Trait Analysis (DTA), which models location history as a discrete trait evolving on a phylogeny, and structured birth-death models, which explicitly incorporate population dynamics through birth (speciation/transmission) and death (recovery/removal) rates [4] [2]. The performance of these models varies significantly in accuracy, bias, and computational demand, influenced by factors such as sampling proportion across populations and the underlying biological reality. This guide provides an objective, data-driven comparison to inform model selection for genomic epidemiology.

Theoretical Foundations and Key Terminology

Core Definitions in Phylogeographic Modeling

• Traits and States: In phylogeography, a trait often represents the geographic location of a sampled sequence. These locations are categorized into distinct states (or demes), such as specific countries, regions, or host species [4] [2]. The evolution of this discrete trait over the phylogeny forms the basis for inferring migration history.
• Birth Rates and Death Rates: In structured population models, the birth rate refers to the rate at which new lineages are generated (e.g., through transmission or speciation), while the death rate is the rate at which lineages are removed from the population (e.g., through recovery or death) [2]. These parameters are fundamental to birth-death models, which use them to infer population dynamics and evolutionary history.
• Sampling Proportions: This refers to the fraction of individuals sequenced from each subpopulation (deme). Biased sampling proportions, where some populations are over- or under-represented, are a major source of inference error, particularly for some model types [2].

Model Classifications

Table 1: Core Phylogeographic Model Classifications in BEAST

Model Category	Key Feature	Representative Software/Package
Discrete Trait Models	Treats location as a discrete trait evolving like a mutation; fast but makes population-genetic assumptions [4] [2].	BEAST Classic [4]
Structured Coalescent Models	Accounts for the effect of population structure on the genealogy; more accurate but computationally intensive [2].	MultiTypeTree (MTT) [4]
Approximated Structured Coalescent	Approximates the structured coalescent to maintain accuracy with better computational efficiency [2].	BASTA, MASCOT [4] [2]
Structured Birth-Death Models	Uses birth and death rates in a structured population; appropriate when a birth-death tree prior is justified [4].	BDMM [4]

Performance Comparison: Quantitative Data

Accuracy and Bias in Root State Inference

A critical test for phylogeographic models is accurately identifying the root state (geographic origin) of an outbreak. Simulations based on the structured coalescent reveal significant performance differences.

Table 2: Comparative Model Performance on Simulated and Empirical Data

Model / Method	Performance on Simulated Data	Performance on Ebola Virus Data	Key Limitation
Discrete Trait Analysis (DTA)	Highly unreliable root state inference; extremely sensitive to sampling bias [2].	Implausibly concluded decades of undetected human-to-human transmission [2].	Conceptual separation of migration and coalescent processes; assumes population sizes drift over time [2].
Structured Coalescent (MTT)	High accuracy but becomes computationally intractable with >3-4 demes [2].	Correctly inferred human outbreaks seeded by an unsampled non-human reservoir [2].	Computational intensity limits application to complex models [2].
BASTA (Approximated Structured Coalescent)	High accuracy, comparable to full structured coalescent, but with greater computational efficiency [2].	Maintains reliability in complex real-world scenarios like Ebola zoonotic transmission [2].	An approximation, though a close one to the structured coalescent [2].

For DTA, a study evaluating Bayesian phylogeographic models found that root state classification accuracy is highest at intermediate sequence dataset sizes and does not consistently improve with more data. Furthermore, the commonly used Kullback-Leibler (KL) divergence metric was found to increase with both the number of discrete traits and dataset size, but was not a predictor of model accuracy, limiting its utility for assessing performance on empirical data [6].

Computational Efficiency and Sample Size

Table 3: Computational and Practical Requirements

Aspect	Discrete Trait Analysis (DTA)	Structured Birth-Death & Coalescent
Computational Speed	Fast; efficient for large datasets with many demes [2].	Slower; computational demand increases with complexity and number of demes [2].
Sample Size for Robust Inference	Performance can degrade with large, biased samples [6] [2].	A study on HIV migration rate inference found a sample size of at least 1,000 sequences was needed for robust estimation with model-based phylodynamics [13].
Handling of Sampling Bias	Poor; conclusions are highly sensitive to biased sampling across locations [2].	Better; designed to explicitly account for population structure and sampling proportions [2].

Experimental Protocols and Methodologies

Protocol 1: Benchmarking Model Accuracy with Simulations

This protocol is used to evaluate the root state inference accuracy of different models, as referenced in studies by [6] and [2].

• Data Simulation: Generate multiple sequence alignments using a known evolutionary model and a predefined phylogenetic tree with a known root state. Key parameters to vary include:
  • The number of sequences in the dataset (from small to large).
  • The number of possible discrete trait values (e.g., geographic locations).
  • The sampling proportion across different demes, intentionally introducing bias.
• Phylogeographic Inference: For each simulated dataset, perform inference using the models under comparison (e.g., DTA, Structured Coalescent, BASTA). Use software implementations such as BEAST2 with relevant packages.
• Accuracy Assessment: Compare the inferred root state from each model and analysis against the known, simulated root state. Calculate the classification accuracy for each model across multiple simulation replicates.
• Metric Evaluation: Record model selection metrics like KL divergence and assess their correlation with the measured classification accuracy.

Protocol 2: Assessing Model Performance on Empirical Data with Known Histories

This approach tests models against real-world outbreaks where the transmission history is well-documented.

• Dataset Selection: Curate a genomic dataset from a pathogen outbreak with a reliably known origin and spread pattern (e.g., the 2014 Ebola virus outbreak in West Africa).
• Model Application: Analyze the dataset using DTA and structured models (e.g., BASTA, BDMM) in a Bayesian framework.
• Result Validation: Compare the model-inferred origin and key migration events against the known epidemiological history to determine which model produces more plausible and accurate results [2].

Workflow Diagram: Phylogeographic Model Testing Pipeline

The following diagram illustrates the logical workflow for evaluating phylogeographic models, integrating the protocols above.

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Software and Analytical Tools for Phylogeographic Research

Tool Name	Type	Primary Function	Relevance to Model Comparison
BEAST 2 / BEAST X [4] [7]	Software Platform	A comprehensive, open-source package for Bayesian phylogenetic, phylogeographic, and phylodynamic inference.	The primary ecosystem for implementing and comparing DTA, structured coalescent (e.g., MTT, BASTA), and structured birth-death (e.g., BDMM) models.
BASTA Package [2]	Software Package (for BEAST 2)	Implements a Bayesian structured coalescent approximation.	A key tool that balances the accuracy of the structured coalescent with the computational efficiency needed for analyses with more than a few demes.
BDMM Package [4]	Software Package (for BEAST 2)	Implements the structured birth-death model for scenarios where a birth-death tree prior is more appropriate than a coalescent prior.	Essential for comparing the coalescent and birth-death paradigms in structured populations.
MASCOT Package [4]	Software Package (for BEAST 2)	An approximated structured coalescent model that allows migration rates to be informed by predictors (e.g., flight data) via a GLM.	Used for more complex, real-world scenarios where external data can inform migration patterns.
ProteinEvolver [14] [15]	Software Framework	A simulator for forecasting protein evolution using birth-death models integrated with structurally constrained substitution models.	Useful for forward-time simulation of evolutionary trajectories to generate benchmark data under realistic models of selection.

The choice between discrete trait analysis and structured birth-death/coalescent models involves a direct trade-off between computational expediency and statistical accuracy.

• Discrete Trait Analysis (DTA) offers speed and is practical for an initial exploration of datasets with many locations. However, its fundamental assumptions make it highly susceptible to sampling bias, potentially leading to misleading conclusions about the origin and spread of an outbreak [2]. Its use requires caution, and its findings should not be relied upon exclusively for critical public health decisions.
• Structured Models, including the approximated coalescent (BASTA) and structured birth-death (BDMM) models, are more reliable for inferring migration rates and root states because they explicitly account for population structure and sampling proportions [2]. While computationally more demanding, they are less sensitive to sampling bias and provide a more realistic representation of the underlying evolutionary and epidemiological processes.

For researchers and drug development professionals, the recommendation is clear: for robust, publication-quality phylogeographic inference, particularly when investigating outbreak origins or transmission dynamics, structured models should be the preferred choice. The use of DTA should be limited to preliminary analyses or cases where its assumptions are explicitly met, with results interpreted with appropriate caution. The ongoing development of efficient approximations like BASTA and advances in software like BEAST X are making these more accurate models increasingly accessible for complex, real-world analyses [7] [2].

In the study of pathogen evolution and spread, phylogeographic models are indispensable for transforming genetic sequence data into epidemiological insights. The core challenge for researchers lies in selecting the appropriate model to reconstruct transmission dynamics from molecular data. The central thesis in modern methodological research revolves on a key dichotomy: discrete trait analysis versus structured birth-death models [4]. While discrete trait models excel at identifying major transitions between predefined locations, particularly with complex population structures, structured birth-death models incorporate the tree-generating process directly, offering a more dynamic representation of how populations evolve and migrate across landscapes [4]. This guide provides an objective comparison of these approaches, detailing their performance, data requirements, and applicability to specific biological questions in pathogen research.

Model Comparison: Discrete Trait Analysis vs. Structured Birth-Death Models

The choice between discrete and structured models fundamentally shapes the inferences drawn from pathogen genetic data. The table below provides a systematic comparison of these model families based on key analytical characteristics.

Table 1: Core Model Comparison for Phylogeographic Inference

Characteristic	Discrete Trait Analysis	Structured Birth-Death Models
Core Methodology	Models location as an evolving discrete trait on the phylogeny, often using Bayesian stochastic search variable selection [4].	Integrates population structure directly into the tree prior, modeling birth, death, and migration events [4].
Underlying Process	Does not incorporate the tree-generating process; trait evolution is modeled independently along branches [4].	Explicitly models the tree-generating process (birth/death) within and between populations [4].
Computational Demand	Generally faster; often the only feasible option with many demes (>10) [4].	Computationally intensive; pure implementations are limited to 3-4 demes, though approximations exist [4].
Typical Applications	Identifying major migration pathways between countries or regions; outbreaks with many locations of origin [4].	Detailed dynamics within meta-populations; inferring migration rates and population sizes [4].
Informing Mechanisms	Migration matrix can be informed by covariates (e.g., flight data, borders) in models like MASCOT [4].	Rate matrices can be set for different epochs but not yet informed by GLM in all implementations [4].

Experimental Protocols and Methodologies

Protocol for Discrete Trait Analysis

Discrete trait analysis requires careful data preparation and model configuration to ensure robust inference of geographic spread.

• Data Collection and Curation:

  • Genetic Sequence Alignment: Assemble a multiple sequence alignment from pathogen genomes (e.g., SARS-CoV-2, influenza).
  • Trait Annotation: Annotate each sequence with a discrete location trait (e.g., country, region). The granularity should reflect the biological question and data density [4].
  • Covariate Data (Optional): For advanced models like MASCOT, gather relevant covariate data (e.g., flight passenger numbers, geographical adjacency, trade volumes) to inform the migration rate matrix [4].

• Model Specification in BEAST 2:

  • Package Selection: Typically implemented using the BEAST_CLASSIC package for basic analysis or MASCOT for a structured coalescent approximation informed by covariates [4].
  • Clock Model: Select a strict or relaxed molecular clock model based on prior knowledge of the pathogen's evolutionary rate.
  • Site Model: Define the nucleotide substitution model (e.g., HKY, GTR) often with a gamma distribution for among-site rate variation.
  • Trait Model: Set up the discrete trait model for the location data. In MASCOT, specify the GLM to include the collected covariate data to explain migration rates [4].
  • Tree Prior: Use a coalescent or birth-death tree prior that is appropriate for the population history and sampling scheme.

• Analysis and Output:

  • MCMC Execution: Run a Markov Chain Monte Carlo (MCMC) analysis for a sufficient number of steps to achieve convergence (effective sample sizes >200 for key parameters).
  • Posterior Analysis: Use software like Tracer to assess convergence and TreeAnnotator to generate a maximum clade credibility tree.
  • Visualization: Visualize the spatiotemporal spread using tools like SpreaD3, which can map the posterior distribution of ancestral locations onto the phylogeny.

Protocol for Structured Birth-Death Models

This protocol is designed for inferring population dynamics and migration rates in a structured population framework.

• Data Collection and Curation:

  • Genetic Sequences and Traits: Follow the same steps as for discrete trait analysis to obtain an alignment and discrete location traits.
  • Epoch Definition: If analyzing data across different time periods (e.g., before and after a travel ban), define the epochs and prepare any epoch-specific rate matrices [4].

• Model Specification in BEAST 2:

  • Package Selection: Use the BDMM (Birth-Death Migration Model) package [4].
  • Model Parameterization: Define the structured birth-death model by specifying:
    • Birth Rate: The rate of lineage diversification within each population.
    • Death Rate: The rate of lineage extinction (sampling-through-time can inform this).
    • Migration Rates: The rates of movement between the defined populations (demes). These can be symmetric or asymmetric.
  • Epoch Settings: Configure different rate matrices for each defined epoch to model changing migration dynamics over time [4].
  • Priors: Place strong, well-justified priors on at least one model parameter to aid convergence, as the model can be parameter-rich [4].

• Analysis and Sensitivity:

  • MCMC Execution: Run long MCMC chains, as structured models are computationally intensive. Monitor mixing and convergence closely.
  • Sensitivity Analysis: Perform sensitivity analyses on key priors (e.g., the birth rate prior) to ensure results are robust to prior choice [4].
  • Interpretation: Analyze the posterior estimates of migration rates and population sizes to understand the dynamics of the meta-population.

Conceptual Workflows and Logical Relationships

Decision Framework for Model Selection

The following diagram outlines the logical workflow for choosing between discrete and structured phylogeographic models based on the research question and data.

Workflow for a Discrete Trait Phylogeographic Analysis

This diagram illustrates the key steps in a standard discrete trait analysis, from data preparation to the final visualization of results.

Successful phylogeographic analysis relies on a suite of software, data sources, and computational resources. The table below details key components of the modern molecular epidemiologist's toolkit.

Table 2: Key Research Reagent Solutions for Phylogeographic Analysis

Tool/Resource	Type	Primary Function	Relevance to Model Comparison
BEAST 2 [4]	Software Package	A cross-platform program for Bayesian phylogenetic analysis of molecular sequences.	Core platform for implementing both discrete trait and structured birth-death models.
BEAST_CLASSIC [4]	Software Package (BEAST 2)	Contains the standard discrete trait model for phylogeography.	Enables analysis of location as an evolving trait without the tree prior.
MASCOT [4]	Software Package (BEAST 2)	Approximates the structured coalescent and allows migration rates to be informed by a GLM.	Bridges discrete and structured approaches; allows many demes and covariate inclusion.
BDMM [4]	Software Package (BEAST 2)	Implements the structured birth-death model for phylogeographic inference.	The primary package for a full structured birth-death analysis with migration.
Tracer	Software Tool	Analyzes the trace files from BEAST MCMC runs to assess convergence and parameter estimates.	Essential for diagnosing model performance and ensuring robust conclusions from any analysis.
Discrete Location Traits	Data	Categorical data (e.g., country, state) assigned to each genetic sequence.	The fundamental input for defining populations or demes in both model families.
Covariate Data (e.g., flight passenger numbers) [4]	Data	External data used to inform the migration rate matrix in models like MASCOT.	Adds ecological realism to the model, helping to explain why certain migration routes are preferred.

The choice between discrete trait analysis and structured birth-death models is not a matter of one being universally superior, but of aligning the model with the specific biological question and data constraints. Discrete trait models, particularly when enhanced with covariate data in frameworks like MASCOT, offer a powerful and computationally efficient method for reconstructing large-scale spread patterns across many locations. In contrast, structured birth-death models provide a more mechanistically rich framework for inferring the dynamic processes of birth, death, and migration within a meta-population, at a higher computational cost. As the field advances, the integration of these approaches with other data streams, such as human mobility models [16] [17] and detailed social mixing patterns [18], will further refine our ability to reconstruct and forecast the complex spread of infectious diseases.

From Theory to Practice: Implementing DTA and SBDM in Research

Discrete Trait Analysis (DTA) represents a fundamental methodological approach in Bayesian phylogeography, enabling researchers to reconstruct the evolutionary history and dispersal dynamics of discrete characteristics across phylogenetic trees. Within the BEAST2 ecosystem, DTA serves as a computationally efficient method for inferring how traits such as geographical locations or phenotypic states transition through time. This approach must be understood in contrast to alternative frameworks, particularly the structured birth-death models, which offer different theoretical foundations and computational trade-offs. The core distinction lies in their treatment of the tree generating process: DTA operates by modeling trait evolution along the branches of a fixed or co-estimated phylogeny without explicitly linking the trait dynamics to the population processes that shape the tree itself. In contrast, structured models like the structured birth-death model (implemented in the bdmm package) explicitly connect population dynamics in different demes to the tree formation process, providing a more integrated but computationally demanding framework [4].

The discrete trait model in BEAST2 is implemented through the BEAST_CLASSIC package and utilizes Bayesian stochastic variable selection to reduce parameter dimensionality, making it particularly advantageous when analyzing systems with many discrete states or demes [4]. This methodological choice becomes particularly significant when designing studies to trace pathogen spread, species migration, or the evolution of drug resistance, where accurately modeling transition rates between states can illuminate critical patterns in evolutionary and epidemiological dynamics. For researchers operating within the constraints of limited computational resources or those requiring rapid analytical turnaround, DTA often presents a pragmatic solution, though its theoretical simplifications must be acknowledged and justified within the specific biological context under investigation.

Theoretical Framework: DTA versus Structured Models in Evolutionary Inference

Fundamental Methodological Divergences

The choice between Discrete Trait Analysis and structured population models represents a critical branch point in phylogenetic study design, with each approach embodying distinct philosophical and statistical assumptions about the evolutionary process. DTA conceptualizes trait evolution as a separate process that occurs along the branches of a phylogenetic tree, typically modeled using continuous-time Markov chains that describe the stochastic transition between discrete states. This methodological separation allows for computational efficiency but makes the fundamental assumption that the trait's evolutionary dynamics are conditionally independent of the underlying tree-generating process given the tree topology and branch lengths. While this simplification enables the analysis of complex multi-state systems, it potentially ignores important feedbacks between population dynamics and trait evolution [4].

In contrast, structured models like the Multi Type Tree (MTT) implementation of the structured coalescent or the birth-death migration model (BDMM) explicitly incorporate the effect of population structure on the tree generation process itself. These models treat the discrete traits (e.g., geographical locations) as integral components that shape the genealogical history through their influence on migration rates and population sizes. The structured coalescent, for instance, models how lineages coalesce within demes and migrate between them, creating a more biologically realistic but computationally intensive framework. Similarly, the birth-death serial sampling model in the bdmm package incorporates temporal epochs, allowing migration and birth rates to vary across predefined time intervals, capturing dynamic processes like seasonal migration patterns or changing connectivity between populations [4] [9].

Performance Trade-offs in Practical Applications

The theoretical distinctions between these approaches manifest in tangible performance trade-offs that researchers must navigate when designing phylogenetic studies. The computational burden of structured models increases dramatically with the number of demes, with pure implementations of the structured coalescent becoming computationally intractable beyond 3-4 demes. Approximation methods like MASCOT (Marginal Approximation of the Structured COalescenT) extend this limit to approximately 10 demes while introducing GLM capabilities to model migration rates as functions of external predictors like flight passenger volumes or trade relationships [4].

Table 1: Computational and Methodological Trade-offs Between Phylogeographic Models

Model Characteristic	Discrete Trait Analysis (DTA)	Structured Birth-Death (BDMM)	Structured Coalescent Approximation (MASCOT)
Theoretical Foundation	Trait evolution independent of tree process	Tight integration of trait and tree processes	Approximation to structured coalescent
Computational Scaling	Scales well with many demes (>10)	Limited to moderate deme numbers	Handles more demes than exact methods
Treatment of Tree Process	Ignores tree generating process	Explicitly models population dynamics	Approximates population structure effects
Data Integration Capabilities	Limited external data integration	Epoch models for rate variation	GLM for migration predictors
Best Application Context	Exploratory analysis, many demes	Few demes with strong population dynamics	Many demes with known migration predictors

The discrete trait model's computational efficiency stems from its treatment of the trait evolution process as separate from the tree prior, significantly reducing the parameter space that must be explored during Markov Chain Monte Carlo (MCMC) sampling. However, this efficiency comes at the cost of biological realism, as the approach does not account for how population structure in the trait of interest might have influenced the phylogenetic tree's shape and branching times. Simulation studies have demonstrated that this disconnect can introduce biases in parameter estimation, particularly when migration rates between demes are high or when the trait exhibits strong influence on population dynamics [4].

Experimental Protocols: Implementing DTA in BEAST2

Software Environment Configuration

Establishing a proper computational environment represents the foundational step in implementing Discrete Trait Analysis within BEAST2. Researchers must first install the core BEAST2 package, which includes the essential BEAUti2 configuration tool, TreeAnnotator for summarizing posterior tree distributions, and associated utilities. The critical additional requirement for DTA is the BEASTCLASSIC package, which contains the discrete trait evolutionary model implementation. Installation occurs through BEAUti2's package manager interface (File > Manage Packages), where users can select and install BEASTCLASSIC, with the system automatically handling any dependencies [19]. Following installation, a BEAUti2 restart is required to activate the newly installed packages and their associated templates.

The broader analytical workflow typically involves several additional software components that facilitate pre-processing, analysis, and post-processing. Tracer provides essential MCMC diagnostics and parameter summary capabilities, allowing researchers to assess chain convergence through Effective Sample Size (ESS) metrics and visualize posterior distributions. For tree visualization and annotation, FigTree offers publication-ready rendering of phylogenetic trees with node annotations, while DensiTree enables qualitative assessment of tree posterior distributions, revealing areas of topological uncertainty or consensus across the MCMC samples [20].

Step-by-Step Analytical Workflow

The implementation of a discrete trait analysis follows a structured pathway from data preparation through to inference and interpretation, with specific considerations at each stage to ensure biologically meaningful and computationally efficient analysis.

Data Preparation and Configuration: The analytical process begins with assembling the molecular sequence alignment in NEXUS or FASTA format and preparing a corresponding trait data set. For the geographical discrete trait analysis exemplified by the primate mitochondrial DNA data set, trait states (e.g., geographical locations) can often be extracted directly from sequence headers using BEAUti's automated parsing capabilities. The Tip Dates panel configures the temporal dimension of the analysis, critical for calibrating evolutionary rates, while the Tip Locations panel assigns discrete trait states to each taxon. The Guess function automates this process by splitting sequence names on delimiters (e.g., underscores) and extracting the relevant trait field [9].

Model Specification: Within the Site Model panel, researchers specify the nucleotide substitution model (e.g., HKY or GTR) with appropriate among-site rate heterogeneity parameters (Gamma category count typically set to 4). The Clock Model panel determines the mode of evolutionary rate variation across branches, with the strict clock representing the simplest assumption and relaxed clocks accommodating rate variation among lineages. Critically, the Tree Prior panel must be configured to Coalescent or Birth-Death models rather than structured tree priors when implementing standard DTA, as the discrete trait evolution is modeled separately from the tree generation process [20] [21].

Trait Model Configuration: The discrete trait model itself is specified through an additional trait partition, which can be added via the + button in the Partitions panel. After importing the trait data as a separate partition, researchers must navigate to the Traits tab to associate this trait data with the tree. The evolutionary model for the discrete trait typically employs Bayesian Stochastic Search Variable Selection (BSSVS), which effectively reduces the number of estimated transition rate parameters by allowing the MCMC to explore different configurations of non-zero rates between states, with Bayes Factor tests identifying well-supported migration pathways [4].

Prior Specification and MCMC Execution: The Priors panel requires careful attention, particularly for the newly added discrete trait rate parameters. Default priors (e.g., Exponential distributions) often provide reasonable starting points, though these should be adjusted based on prior biological knowledge. The MCMC settings (chain length, sampling frequency) in the MCMC panel must be configured to ensure adequate exploration of the parameter space, with chain lengths typically ranging from 10-100 million generations depending on dataset size and complexity. Following MCMC execution, diagnostic tools like Tracer assess convergence (ESS > 200 for all parameters), with TreeAnnotator generating a maximum clade credibility tree from the posterior sample for visualization and interpretation in FigTree or similar software [20].

Table 2: Essential Research Reagent Solutions for DTA Implementation

Research Reagent	Function in Analysis	Implementation Details
BEAST_CLASSIC Package	Provides discrete trait evolutionary model	Install via BEAUti package manager; required for DTA
BEAUti2 Configuration Tool	Generates BEAST2 XML configuration files	Graphical interface for model specification and data import
Tracer	MCMC diagnostic assessment	Evaluates chain convergence via ESS statistics
TreeAnnotator	Summarizes posterior tree distribution	Generates maximum clade credibility trees with node annotations
FigTree/DensiTree	Phylogenetic tree visualization	Renders annotated trees and posterior tree distributions

Comparative Performance Assessment: Empirical Data and Simulation Studies

Methodological Comparison Through Benchmarking

The performance characteristics of Discrete Trait Analysis versus structured models have been elucidated through both empirical applications and carefully designed simulation studies, revealing context-dependent advantages and limitations. A critical benchmark emerges from the analysis of influenza H3N2 evolution, where the structured birth-death model (BDMM) implemented through the bdmm package has demonstrated enhanced precision in reconstructing migration pathways between geographical regions when compared to standard DTA approaches. In these applications, BDMM recovered posterior estimates that more closely aligned with known epidemiological patterns, particularly when incorporating temporal epoch models that accommodated seasonal variation in migration rates [9].

Simulation studies examining phylogenetic regression under tree misspecification provide indirect but valuable insights into the robustness of different analytical approaches. Recent investigations have revealed that phylogenetic comparative methods exhibit heightened sensitivity to incorrect tree specification as dataset size increases, with false positive rates soaring to nearly 100% in some misspecified scenarios [22]. This finding has profound implications for DTA, which inherently assumes the correctness of the underlying phylogenetic tree or treats it as fixed during trait evolution modeling. Structured models partially mitigate this concern by co-estimating the tree and trait dynamics, though at substantial computational cost. The application of robust estimators in phylogenetic regression has demonstrated promise in rescuing analyses from tree misspecification, suggesting potential avenues for enhancing the robustness of both DTA and structured approaches [22].

Computational Efficiency and Scalability

The computational burden differential between DTA and structured models represents one of the most practically significant considerations for researchers designing phylogeographic studies. Empirical benchmarks conducted on influenza and rabies virus datasets have demonstrated that DTA implementations typically achieve convergence 3-5 times faster than structured birth-death models for equivalent datasets, making them particularly valuable for exploratory analysis or when computational resources are constrained. This efficiency advantage widens substantially as the number of discrete states increases, with DTA maintaining tractability for systems with 10+ demes where structured models become computationally prohibitive without approximation methods [4].

The introduction of approximation methods like MASCOT for the structured coalescent and the continued refinement of BDMM have narrowed but not eliminated this performance gap. For the critical task of ancestral state reconstruction at internal nodes, which forms the core objective of many discrete trait analyses, both approaches demonstrate similar accuracy under conditions of moderate migration rates and clearly differentiated populations. However, under high migration scenarios or when population structure strongly influences the tree shape, structured models consistently outperform DTA in reconstruction accuracy, justifying their additional computational requirements in these specific biological contexts [4].

Advanced Implementation Strategies and Future Directions

Post-hoc Analysis and Integration Approaches

Advanced implementation strategies for discrete trait analysis have emerged that leverage the computational advantages of DTA while mitigating some of its theoretical limitations. The recently enhanced fixed tree and tree set support in BEAST2, implemented through the FixedTreeAnalysis package, enables a hybrid approach where a previously inferred posterior tree distribution serves as the foundation for subsequent discrete trait analysis. This post-hoc strategy offers significant computational advantages for large datasets, particularly when the primary phylogenetic relationships have been well-established through previous genomic analyses and the research question focuses specifically on trait evolution patterns [23].

The post-hoc approach involves importing a fixed tree or tree set through BEAUti's template system (File > Templates > Fixed Tree Analysis or Tree Set Analysis), then adding the discrete trait partition and configuring the evolutionary model as in standard DTA. When utilizing a tree set drawn from a previous posterior distribution, the MCMC samples trees from this set throughout the analysis, preserving some uncertainty in phylogenetic relationships while dramatically reducing computational time compared to full joint inference. Empirical validation studies have demonstrated that this approach can produce comparable results to joint inference when the fixed trees adequately represent the posterior distribution, though it necessarily ignores potential feedbacks between trait evolution and tree generation [23].

Methodological Integration and Emerging Solutions

The methodological landscape for discrete trait analysis continues to evolve, with several emerging solutions addressing longstanding limitations of standard approaches. For geographical trait analysis, the discrete phylogeographic model in BEAST_CLASSIC remains the standard implementation, but alternative frameworks like the random walk on a sphere model (in the GEO_SPHERE package) offer continuous alternatives that may better reflect biological reality for certain study systems. Similarly, the break-away package implements a founder-dispersal model that assumes one population remains in place while the other migrates at each branching event, producing fundamentally different root location estimates compared to standard random walk models [4].

Future methodological development appears focused on enhancing the integration of external data sources and accommodating more complex evolutionary scenarios. The MASCOT package's GLM capabilities, which allow migration rates to be modeled as functions of predictor variables, represent a promising direction that could potentially be incorporated into DTA frameworks. Similarly, the epoch modeling functionality in BDMM, which accommodates discrete changes in migration rates over time, addresses an important biological reality that currently requires custom implementation in standard DTA [4] [9]. As Bayesian computational methods continue to advance, particularly through Hamiltonian Monte Carlo and other efficient sampling algorithms, the current computational barriers separating DTA from more complex structured models may diminish, potentially enabling more researchers to employ the most biologically appropriate methods regardless of dataset size or complexity.

The comparative analysis of Discrete Trait Analysis and structured birth-death models reveals a landscape of methodological trade-offs rather than absolute superiority of either approach. DTA emerges as the preferred choice for exploratory analyses, systems with numerous discrete states (>10), and situations with computational constraints. Its implementation through the BEAST_CLASSIC package offers a robust, well-supported workflow with relatively straightforward interpretation. In contrast, structured models like BDMM provide enhanced biological realism for systems with strong population dynamics, fewer discrete states, and available prior information about birth, death, and migration parameters.

Strategic implementation of DTA should incorporate several evidence-based practices: (1) utilization of BSSVS to reduce parameter dimensionality and identify well-supported transitions, (2) consideration of post-hoc approaches using fixed tree sets when analyzing large datasets, (3) comprehensive model diagnostics including Bayes Factor tests for migration rates, and (4) sensitivity analyses examining the impact of prior choices on posterior estimates. For research questions where the discrete trait of interest likely influenced population dynamics and thereby shaped the phylogenetic tree itself, structured models warrant their additional computational requirements. Ultimately, the expanding toolkit for discrete trait evolution in BEAST2 provides researchers with multiple pathways to reconstruct evolutionary history, with selection criteria extending beyond statistical performance to encompass biological realism, computational feasibility, and analytical transparency.

Multi-type birth-death models (MTBD), also referred to as structured birth-death models (SBDM), represent a powerful class of phylodynamic models that enable researchers to quantify past population dynamics in structured populations based on phylogenetic trees [8]. These models serve as phylodynamic analogies of compartmental models in classical epidemiology, bridging the gap between traditional epidemiology and pathogen sequence data [24]. The core strength of MTBD models lies in their ability to infer key epidemiological parameters—such as the average number of secondary infections (Re) and infectious time—directly from pathogen phylogenetic trees, which approximate transmission histories [24]. This approach is particularly valuable for emerging epidemics where traditional epidemiological data may be insufficient for accurate parameter estimation. The growing availability of genetic sequencing data has created both opportunities and computational challenges for phylodynamic analyses, driving the development of increasingly sophisticated inference methods and model implementations.

The positioning of MTBD models within the broader landscape of phylogenetic analysis reveals their unique value proposition. While discrete trait analysis (DTA) offers one approach to understanding trait evolution across phylogenies, MTBD models provide a more biologically grounded framework for epidemiological applications by explicitly modeling the population dynamics that generate the observed tree [25]. This distinction becomes particularly important when analyzing emerging infectious diseases, where the stochastic nature of transmission dynamics favors birth-death models over coalescent approaches [24]. The computational framework underlying these models has evolved significantly, with current implementations focusing on maximum likelihood and Bayesian inference methods that can handle increasingly large datasets while maintaining numerical stability.

Model Foundations: Theoretical Framework of Multi-Type Birth-Death Processes

Core Mathematical Framework

The multi-type birth-death model extends the basic birth-death process with sampling to populations divided into a finite number of discrete subpopulations or types [25] [8]. In this formal structure, each individual in the process is characterized by their type membership ( i \in {1 \ldots d} ), where ( d ) represents the total number of possible types. The process is defined by several type-specific parameters that evolve over time intervals ( k \in {1 \ldots n} ) delineated by time points ( 0 < t1 < \ldots < t{n-1} < T ), where ( T ) represents the present time and ( t_0 = 0 ) marks the origin of the process [8].

The model incorporates several key rate parameters that govern the dynamics: birth rates (( \lambda{ij,k} )) representing transmission events from type ( i ) to type ( j ); death rates (( \mu{i,k} )) representing removal from the infectious pool; sampling rates (( \psi{i,k} )) representing the observation of infected individuals; and migration rates (( m{ij,k} )) representing type changes with ( m{ii,k} = 0 ) [25] [8]. Additionally, the model includes contemporaneous sampling probabilities (( \rho{i,k} )) at time points ( tk ), and removal probabilities (( r{i,k} )) that determine whether sampled individuals continue transmitting [8]. This comprehensive parameterization enables the model to capture complex epidemiological scenarios with structured populations and changing dynamics over time.

Key Model Variants and Their Applications

The MTBD framework has spawned several specialized models tailored to specific epidemiological contexts. The Birth-Death Exposed-Infectious (BDEI) model addresses pathogens with incubation periods by incorporating an exposed state between infection and becoming infectious, making it suitable for diseases like Ebola and SARS-CoV-2 [24]. The Multitype Birth-Death Skyline (BDSKY) model allows for piecewise-constant rate parameters through time, enabling the capture of changing epidemic dynamics in response to interventions or natural progression [25]. These models share a common mathematical foundation but differ in their state spaces and parameter constraints, making them adaptable to diverse public health scenarios.

The relationship between MTBD models in epidemiology and analogous models in macroevolution reveals important theoretical connections. The State Speciation and Extinction (SSE) model family, including BiSSE, MuSSE, and ClaSSE models, share mathematical similarities with MTBD models but differ in their sampling assumptions and biological interpretations [24]. While epidemiological models typically involve sampling through time, macroevolutionary models generally assume sampling only at present (extant species). Despite these differences, recent methodological advances have enabled cross-fertilization between these domains, with improvements in one area often benefiting the other.

Figure 1: Multi-Type Birth-Death Process Workflow. This diagram illustrates the stochastic process underlying MTBD models, showing how different events shape the complete transmission tree and ultimately produce the sampled phylogeny used for inference.

Computational Implementation: Inference Methods and Performance Considerations

Likelihood Computation and Numerical Challenges

The computational core of MTBD model inference involves calculating the probability density of the observed phylogenetic tree given the model parameters. This is achieved through the numerical integration of a system of differential equations known as master equations [24]. The likelihood computation employs a backward-time approach, evaluating probability densities ( g_{i,k}^e(t) ) that represent the probability that an individual of type ( i ) in time interval ( k ) at time ( t ) evolved as observed in the tree along edge ( e ) [25] [8]. The initial conditions for these equations depend on the type of node terminating the edge: serial sampling events, sampled ancestors, contemporaneous samples, or branching events [25].

A significant challenge in MTBD model inference has been numerical instability, particularly due to underflow issues when processing large trees [24] [8]. Early implementations struggled with datasets containing more than 250-500 sequences, limiting their applicability to the large genomic datasets increasingly generated during outbreaks [24] [8]. Recent advances have addressed these issues through mathematical reformulations that remove recursive dependencies between parent and child nodes, enabling parallel computation and improving numerical stability [24]. Additionally, techniques such as likelihood rescaling and careful management of extremely small probability values have extended the practical limits of these methods to trees with thousands of samples [8].

Inference Algorithms and Software Implementations

Two primary algorithmic approaches dominate MTBD model inference: maximum likelihood estimation and Bayesian methods. Maximum likelihood approaches aim to find parameter values that maximize the probability of the observed tree, often employing efficient equation resolution methods and optimization algorithms [24]. Bayesian methods implement MTBD models within Markov Chain Monte Carlo (MCMC) frameworks, enabling joint inference of trees and parameters while quantifying uncertainty through posterior distributions [25] [8]. Each approach offers distinct advantages: maximum likelihood methods typically provide faster computation, while Bayesian methods naturally incorporate prior knowledge and quantify estimation uncertainty.

Several software packages implement MTBD models with varying specializations and capabilities. The BEAST2 package bdmm provides a comprehensive Bayesian implementation of multitype birth-death models, allowing for co-estimation of phylogenies and model parameters [25] [8]. PyBDEI implements a maximum likelihood framework specifically for the BDEI model, employing parallelization strategies for efficient computation on large trees [24]. PhyloDeep offers an alternative approach using deep learning to bypass likelihood calculation entirely, though this requires extensive training on simulated trees [24]. These tools represent the current state-of-the-art in MTBD model inference, each with specific strengths for different analytical scenarios.

Performance Comparison: Experimental Evaluation of SBDM Implementations

Computational Performance Metrics

Recent methodological advances have dramatically improved the performance of MTBD model implementations, enabling analyses previously hampered by computational limitations. The table below summarizes key performance metrics for major implementations based on experimental evaluations reported in the literature:

Table 1: Performance Comparison of MTBD Model Implementations

Implementation	Inference Method	Maximum Tree Size	Computation Time	Key Advantages
PyBDEI [24]	Maximum Likelihood	10,000 samples	~2 minutes for 10,000 samples	High speed, parallel computation, numerical stability
bdmm (original) [8]	Bayesian (MCMC)	~250 samples	Hours to days for medium datasets	Joint tree and parameter estimation, uncertainty quantification
bdmm (improved) [8]	Bayesian (MCMC)	Several hundred samples	30-50% faster than original	Better handling of large datasets, multiple sampling events
PhyloDeep [24]	Deep Learning	No strict limit	Fast prediction (after training)	Bypasses numerical issues; requires extensive training data

The performance comparisons reveal distinct trade-offs between computational approaches. PyBDEI demonstrates remarkable efficiency, estimating parameters and confidence intervals for a 10,000-sample tree in approximately two minutes [24]. This represents orders of magnitude improvement over previous implementations and enables rapid analysis even during fast-moving outbreaks. The improved bdmm implementation offers more moderate gains, handling several hundred samples with better numerical stability and 30-50% faster computation compared to its predecessor [8]. PhyloDeep presents an alternative paradigm that avoids numerical instability entirely but requires computationally expensive training on millions of simulated trees spanning the expected parameter space [24].

Estimation Accuracy and Statistical Performance

Beyond computational efficiency, estimation accuracy represents a critical metric for evaluating MTBD model implementations. Experimental comparisons using simulated datasets have demonstrated that modern implementations not offer greater speed but also improved accuracy [24]. In side-by-side comparisons, PyBDEI showed superior accuracy in parameter estimation compared to both BEAST2 (using bdmm) and PhyloDeep, particularly for epidemiological parameters such as transmission rates and reproduction numbers [24]. This improvement likely stems from the numerical stability improvements that enable analysis of larger datasets, which in turn provide more information for parameter estimation.

The statistical performance of MTBD models also depends on model specification and dataset characteristics. Models with finer discretization schemes and larger state spaces tend to show artificially inflated support metrics (such as Kullback-Leibler divergence) with increasing dataset sizes, potentially misleading model selection efforts [6]. Interestingly, root state classification accuracy—a key metric in phylogeographic studies—tends to peak at intermediate sequence dataset sizes rather than increasing monotonically with data quantity [6]. These nuances highlight the importance of careful model specification and validation when applying MTBD models to empirical data.

Experimental Protocols: Methodologies for Model Evaluation

Simulation-Based Performance Assessment

The evaluation of MTBD model implementations relies heavily on simulation studies, where data is generated from known parameters and inference methods are assessed by their ability to recover these ground truths. A standard protocol involves: (1) specifying a complete set of model parameters including birth, death, sampling, and migration rates; (2) simulating phylogenetic trees under the MTBD process using these parameters; (3) performing inference on the simulated trees using the implementation being evaluated; and (4) comparing estimated parameters to their true values [24] [8]. This approach allows for controlled assessment of accuracy, precision, and computational efficiency across a range of epidemiological scenarios.

Performance metrics commonly reported in these studies include absolute error (difference between estimated and true parameter values), relative error (absolute error divided by true value), coverage probability (proportion of confidence or credibility intervals containing the true value), and computational time [24] [8]. For branching process parameters such as reproduction numbers, additional metrics like mean squared error and statistical power to detect changes in parameters over time may also be reported. These comprehensive assessments provide researchers with practical guidance for selecting appropriate methods based on their specific analytical needs and computational constraints.

Empirical Validation with Real-World Epidemics

Beyond simulation studies, MTBD models are validated through application to empirical datasets with known epidemiological history. A prominent example is the analysis of the 2014 Ebola epidemic in Sierra Leone using the PyBDEI implementation [24]. This validation followed a rigorous protocol: (1) collection of viral sequence data from public databases; (2) reconstruction of the phylogenetic tree using molecular clock methods; (3) parameter estimation under the BDEI model; and (4) comparison of estimated parameters to independent epidemiological observations. The successful application demonstrated both the computational feasibility of analyzing large datasets and the biological plausibility of the resulting parameter estimates [24].

Similar validation approaches have been applied to other pathogens, including influenza A virus using the bdmm implementation [8]. In these studies, researchers analyzed globally distributed H3N2 sequences to infer seasonal dynamics and migration patterns, with results compared to known influenza epidemiology such as the timing of seasonal peaks and dominant transmission routes [25] [8]. The consistent finding that the main migration path leads from tropical to northern regions aligns with independent epidemiological observations, providing external validation of the model inferences [25]. These real-world applications demonstrate the practical utility of MTBD models for addressing substantive questions in infectious disease dynamics.

Figure 2: Experimental Validation Workflows. The diagram illustrates the two complementary approaches for evaluating MTBD model implementations: simulation studies with known ground truth and empirical validation with real-world epidemiological data.

Computational Tools and Software Packages

Implementing MTBD models requires specialized software tools that can handle the complex likelihood calculations and numerical optimization procedures. The table below summarizes key resources available to researchers:

Table 2: Essential Computational Tools for MTBD Model Implementation

Tool/Resource	Function	Implementation Details	Application Context
BEAST2/bdmm [25] [8]	Bayesian inference of MTBD models	MCMC sampling with tree integration	General phylodynamic analysis with uncertainty quantification
PyBDEI [24]	Maximum likelihood estimation for BDEI model	Parallel ODE resolution, confidence intervals	Fast analysis of large trees for pathogens with incubation periods
PhyloDeep [24]	Likelihood-free inference via deep learning	Neural network trained on simulated trees	Applications where traditional inference fails due to numerical issues
TreeSim	Phylogenetic tree simulation under birth-death processes	R package with various model extensions	Simulation studies for method validation

Methodological Considerations and Best Practices

Successful implementation of MTBD models requires attention to several methodological considerations. Model specification should balance biological realism with parsimony, as overly complex models with unnecessary parameters can lead to identifiability issues and poor convergence [6] [8]. For Bayesian implementations, prior specification requires careful consideration, particularly for parameters with limited information in the data. For all implementations, diagnostic checks are essential—including assessment of convergence for MCMC methods and evaluation of numerical stability for likelihood-based approaches [24] [8].

Practical guidance from methodological studies suggests several best practices. When analyzing new datasets, researchers should begin with simplified models and gradually increase complexity while monitoring improvements in model fit [8]. Computational bottlenecks can often be addressed by leveraging parallelization strategies, particularly for the backward pass of likelihood calculations [24]. For applications focusing on origin estimation, researchers should be aware that root state classification accuracy typically peaks at intermediate dataset sizes rather than increasing monotonically with more data [6]. These evidence-based practices can significantly enhance the reliability and efficiency of MTBD model analyses.

The development of multi-type birth-death models represents a significant advancement in phylodynamics, providing a powerful framework for inferring transmission dynamics from pathogen genetic sequences. Recent improvements in computational implementations have dramatically expanded the scope of these methods, enabling applications to large datasets that were previously computationally prohibitive [24] [8]. The performance comparisons presented in this guide demonstrate that modern implementations offer not only greater speed but also improved accuracy and numerical stability, addressing key limitations that hampered earlier approaches.

Looking forward, several promising directions emerge for further development of MTBD models. Integration with additional data sources, such as incidence curves and contact patterns, could enhance parameter identifiability and epidemiological relevance [24]. Development of more efficient inference algorithms remains an active area of research, with potential benefits for real-time analysis during outbreaks [8]. Additionally, extending model flexibility to accommodate more complex population structures and between-type interactions would broaden the applicability of these methods to diverse public health scenarios. As genetic sequencing continues to play an increasingly central role in infectious disease surveillance, MTBD models will likely remain essential tools for translating these data into actionable epidemiological insights.

Understanding the transmission dynamics of pathogens like HIV across different risk groups is a cornerstone of effective public health intervention. Phylodynamics, which uses pathogen genetic sequences to infer epidemiological dynamics, provides two principal methodological frameworks for this task: the structured coalescent model and the multi-type birth-death model. While both can estimate migration rates between populations or risk groups, they operate under distinct assumptions that significantly impact their accuracy and appropriate application [26]. This guide provides a objective comparison of these approaches, focusing on their performance in uncovering HIV transmission dynamics among risk groups such as men who have sex with men (MSM), heterosexuals (Hetero), and injecting drug users (IDU). We summarize experimental data, provide detailed protocols, and offer practical guidance for researchers navigating these powerful but complex analytical tools.

Model Comparison: Structured Coalescent vs. Multi-Type Birth-Death

Quantitative Performance Comparison

A comprehensive simulation study compared the inferential outcomes of the structured coalescent model with constant population size and the multi-type birth-death model with a constant rate across various epidemic scenarios [26]. The table below summarizes the key performance metrics from this investigation.

Table 1: Performance comparison of structured phylodynamic models across epidemic scenarios

Epidemic Scenario	Model	Migration Rate Accuracy	Migration Rate Precision	Source Location Estimation
Epidemic Outbreaks	Multi-type Birth-Death	Superior	Not Specified	Comparable and Robust
Epidemic Outbreaks	Structured Coalescent	Less Accurate	Not Specified	Comparable and Robust
Endemic Diseases	Multi-type Birth-Death	Comparable	Less Precise	Comparable and Robust
Endemic Diseases	Structured Coalescent	Comparable	More Precise	Comparable and Robust

Key Findings and Recommendations

The research offers tangible modeling advice for infectious disease analysts [26]:

• For epidemic outbreaks or scenarios with varying population size, structured coalescent models with constant population size should be avoided as they can lead to inaccurate migration rate estimates. Instead, coalescent models accounting for varying population size or birth-death models should be favored.
• For endemic disease scenarios, either model can be used effectively, as both produce comparable coverage and accuracy of migration rates.
• Estimating the source location of a disease is robust across both models and different scenarios.

Experimental Protocols and Workflows

Core Phylodynamic Analysis Workflow

The foundational workflow for Discrete Trait Analysis (DTA) in HIV studies involves several standardized steps, from data collection to phylogenetic reconstruction and interpretation.

Figure 1: Core workflow for phylodynamic analysis of HIV transmission

Detailed Methodological Protocols

Sequence Data Curation and Preparation

Studies of HIV-1 CRF5501B, CRF5901B, and CRF07_BC provide representative protocols for data curation [27] [28] [29]. The standard workflow includes:

• Sequence Retrieval: Partial pol gene sequences are retrieved from the Los Alamos National Laboratory (LANL) HIV Sequence Database with known sampling year, geographic location, and risk group.
• Quality Control: Using RIP v3.0 from the LANL site for genotype confirmation and Hypermut v2.0 for analyzing APOBEC-induced hypermutation.
• Sequence Alignment: Performed with MAFFT v7.427 or Gene Cutter from LANL, followed by manual adjustment in BioEdit v7.2.5.
• Model Selection: Identification of the best-fitting nucleotide substitution model (typically GTR+Γ+I) using jModelTest v2.1.10 according to AIC, AICc, BIC, and DT methods.

Phylogenetic and Discrete Trait Analysis Implementation

The Bayesian phylogenetic approach is implemented in BEAST v1.8.2 or BEAST2 for DTA [27] [30] [8]:

• Tree Prior Selection: Using a Bayesian Skygrid coalescent tree prior or a birth-death skyline plot prior.
• Molecular Clock Model: Employing an uncorrelated lognormal relaxed-clock model.
• Discrete Trait Diffusion: Modeling geographic location and risk group as a diffusion process among discrete states using a non-reversible continuous-time Markov chain.
• MCMC Analysis: Running for 500 million steps with sampling every 50,000 steps, with convergence evaluated using Tracer v1.7.1 to ensure effective sample sizes (ESS) >200.
• Tree Summarization: Generating maximum clade credibility (MCC) trees using TreeAnnotator after discarding the first 10% as burn-in.

Transmission Network Analysis

HIV-TRACE (HIV TRAnsmission Cluster Engine) is used to infer transmission networks [27] [31]:

• Genetic Distance Calculation: Computing all pairwise distances using the TN93 substitution model.
• Cluster Identification: Considering sequences with genetic distance ≤0.02 substitutions/site as potentially linked, with multiple linkages combined into putative transmission clusters.
• Cluster Characterization: Analyzing the geographic and risk group composition of identified clusters.

Application to HIV Transmission Dynamics in China

Case Study Findings

Recent studies applying these methodologies to HIV-1 strains in China have revealed critical insights into transmission dynamics:

Table 2: Key findings from DTA studies of HIV-1 transmission in China

HIV-1 Strain	Origin	Major Transmission Hubs	Key Risk Group Interactions	Study Reference
CRF55_01B	Jan 2003, Guangdong, MSM	Guangdong Province, MSM	All sequences from unknown risk clustered within MSM groups	[27]
CRF59_01B	1992.83, Southeast China	Southeast China	26.67% of clusters included both MSM and heterosexuals	[28]
CRF07_BC	Oct 1992-Jul 1993, Yunnan, IDU	Yunnan (IDU), Guangdong (MSM)	Now accounts for >40% of infections in China, primarily among MSM	[29]

Advanced Birth-Death Skyline Analysis

The birth-death skyline plot represents a significant methodological advancement, enabling direct estimation of the effective reproductive number (R) through time [30]. This model is based on a forward-in-time process of transmission, death/recovery, and sampling, with parameters allowed to change in a piecewise fashion. Application to a UK HIV-1 subtype B dataset revealed temporal changes in R, showing a decline from approximately 1.87 at the origin of the cluster to below 1 around 1998, potentially reflecting the introduction and improvement of antiretroviral therapy [30].

The Scientist's Toolkit: Essential Research Reagents and Software

Table 3: Essential research reagents and computational tools for phylodynamic analysis

Tool/Reagent	Category	Primary Function	Application Example
LANL HIV Database	Data Resource	Curated repository of HIV sequences	Source of partial pol gene sequences for analysis [27] [29]
BEAST/BEAST2	Software Package	Bayesian evolutionary analysis	Phylogenetic reconstruction with DTA [27] [30] [8]
bdmm Package	Software Plugin	Multi-type birth-death model implementation	Phylodynamic inference in structured populations [8]
HIV-TRACE	Online Tool	Transmission cluster identification	Network analysis using genetic distance thresholds [27]
Tracer	Analysis Tool	MCMC diagnostics	Assessing convergence and effective sample sizes [27]

The choice between structured coalescent and multi-type birth-death models should be guided by the specific epidemiological context and research questions. For epidemic outbreaks with changing population sizes, the multi-type birth-death model provides more accurate estimates of migration rates between risk groups [26]. For endemic scenarios, either model is appropriate, with the structured coalescent potentially offering greater precision [26]. The Bayesian birth-death skyline plot offers particular advantage when estimating temporal changes in the effective reproductive number is a priority [30]. Recent algorithmic improvements to the bdmm package have dramatically increased the scalability of multi-type birth-death analyses, enabling robust phylodynamic inference of larger datasets [8]. By applying these sophisticated phylodynamic methods, researchers can continue to uncover critical insights into HIV transmission dynamics, ultimately guiding more effective and targeted public health interventions.

Influenza pandemics and seasonal epidemics present a persistent threat to global health, necessitating robust methods for the timely estimation of transmission dynamics. The effective reproductive number (Re), defined as the average number of secondary cases generated per typical infectious case in a non-fully susceptible population, serves as a crucial metric for assessing transmissibility and guiding public health interventions [32]. This case study objectively compares the application of Structured Birth-Death Models (SBDM) against alternative phylogenetic and statistical models for estimating Re during influenza outbreaks. Framed within a broader thesis on discrete trait analysis versus structured birth-death models research, this analysis provides experimental data, detailed methodologies, and performance comparisons tailored for researchers, scientists, and drug development professionals. The evaluation leverages empirical data from historical influenza outbreaks, including the 2009 H1N1 pandemic, to ground the comparison in real-world scenarios [33] [34].

Theoretical Background: Reproduction Numbers and Model Frameworks

Key Epidemiological Concepts

The basic reproduction number (R0) measures the transmission potential of a disease in a fully susceptible population. In contrast, the effective reproduction number (Re) reflects real-time transmissibility within a population with existing immunity, calculated as R = R0 * x, where x is the fraction of susceptible hosts [32]. An Re value greater than 1 indicates that an epidemic is growing, a value equal to 1 signifies an endemic state, and a value less than 1 suggests the outbreak is declining. The herd immunity threshold represents the proportion of the population that must be immune to prevent sustained transmission, occurring when Re is maintained below 1 [32].

Model Typologies in Phylogeographic Inference

Phylogeographic models reconstruct the spatial and temporal spread of pathogens using genetic sequence data. The choice between discrete and continuous trait models depends on the underlying assumptions of the migration process:

• Discrete Trait Models: Represent sample locations by grouping them into predefined categories (e.g., countries, regions). These are suitable when spread occurs through distinct jumps, as with modern human travel patterns. The Discrete Trait Model in BEAST Classic and structured coalescent approximations like MASCOT fall into this category [4].
• Continuous Trait Models: Represent sample locations as continuous coordinates (e.g., latitude-longitude pairs). These assume a random walk migration process and are more appropriate for land-dwelling organisms or historical viral spread [4].
• Structured Birth-Death Models (SBDM): A class of models that use birth-death processes to model transmission, speciation, and extinction events within structured populations (e.g., different geographic locations or host species). In BEAST2, the BDMM package implements SBDM, can distinguish different epidemic epochs, and is more appropriate than coalescent models when population sizes change rapidly [4].

Table 1: Overview of Model Classes for Phylogeographic Inference

Model Class	Key Features	Representative Packages	Best-Suited Context
Structured Birth-Death Models (SBDM)	Models transmission, speciation, extinction; handles structured populations; can incorporate different epochs.	BDMM	Rapidly changing populations; when a birth-death prior is more appropriate than a coalescent.
Discrete Trait Models	Groups locations into categories; uses Bayesian stochastic variable selection.	BEAST Classic (Discrete Trait Model)	Modern human travel (jump-based spread); many demes.
Structured Coalescent Models	Approximates population structure within the coalescent framework; can be informed by external data.	Multi Type Tree, MASCOT, SCOTTI	Population-level sampling; smaller number of demes.
Continuous Trait Models	Treats location as a continuous variable; models spread as a random walk.	GEO_SPHERE, BEAST Classic (Random Walk)	Land-based diffusion; hunter-gatherer societies; historical spread.

Comparative Analysis of Model Performance in Influenza Outbreaks

A systematic review of the literature provides a baseline for expected Re values across different influenza types and outbreaks, against which model performance can be contextualized.

Table 2: Historical Reproduction Number (R) Estimates for Seasonal, Pandemic, and Zoonotic Influenza [34]

Influenza Type/Strain	Median R Value (Interquartile Range - IQR)	Number of Studies (R Values)	Contextual Notes
1918 Pandemic (H1N1)	1.80 (IQR: 1.47-2.27)	24 studies (51 R values)	Higher transmissibility in confined settings (median R = 3.82).
1957 Pandemic (H2N2)	1.65 (IQR: 1.53-1.70)	6 studies (7 R values)	Based on the second wave of illnesses.
1968 Pandemic (H3N2)	1.80 (IQR: 1.56-1.85)	4 studies (7 R values)
2009 Pandemic (H1N1pdm)	1.46 (IQR: 1.30-1.70)	57 studies (78 R values)	Similar median R in first (1.46) and second (1.48) waves.
Seasonal Influenza	1.28 (IQR: 1.19-1.37)	24 studies (47 R values)	Represents typical inter-pandemic transmissibility.
Novel Influenza (e.g., H5N1)	Mostly <1 (4 of 6 R values)	4 studies (6 R values)	Limited human-to-human transmission.

Case Study: 2009 H1N1 Pandemic in Hong Kong

A detailed study from Hong Kong provides a direct comparison of Re estimation using different data sources and highlights the impact of control measures.

• Re Estimation and Temporal Dynamics: Prospective estimation of Re using case notifications and hospitalizations showed that Re declined from 1.4-1.5 at the start of the local epidemic to around 1.1-1.2 later in the summer of 2009 [33]. This decline coincided with school vacations and the implementation of mitigation strategies, suggesting that both seasonality and public health interventions can influence transmissibility.
• Model Robustness Across Data Sources: Estimates of Re based on case notifications and hospitalizations of confirmed H1N1 cases were found to be broadly consistent, indicating that hospitalization data alone can be a reliable surrogate for estimating transmissibility when case-based surveillance is inconsistent [33]. However, changes in hospitalization rates or clinical thresholds over short periods can pose challenges for estimation.
• Real-Time Monitoring Feasibility: The study demonstrated that real-time monitoring of Re is feasible and can provide useful information to public health authorities for situational awareness and calibration of mitigation strategies [33]. The real-time estimates were consistent with final estimates, with divergence occurring only in the last few days of each analysis period.

Performance Comparison of Modeling Approaches

Different models offer distinct advantages and face specific limitations in the context of Re estimation.

Table 3: Model Performance Comparison for Estimating Influenza Re

Model / Approach	Computational Efficiency	Key Strengths	Key Limitations
Structured Birth-Death Model (BDMM)	Moderate; requires strong priors for convergence [4].	Can model different epidemic epochs; appropriate for emerging outbreaks with changing dynamics.	Rate matrices cannot yet be informed by GLM; requires careful prior specification [4].
Discrete Trait Model (BEAST Classic)	Fast; suitable for many demes [4].	Fast inference; useful for capturing travel-mediated jump dispersal.	Does not incorporate the tree-generating process in the trait evolution model.
Structured Coalescent (MASCOT)	Good approximation; allows more demes than pure structured coalescent [4].	Migration rates can be informed by covariates (e.g., flight data, borders) via GLM.	An approximation; performance may vary with deme number and data structure.
Statistical (from Case Data)	High; suitable for real-time analysis [33].	Direct use of surveillance data; feasible for real-time public health decision-making.	Sensitive to reporting delays and changes in case ascertainment.
Continuous Phylogeographic (GEO_SPHERE)	High for continuous traits [4].	Integrates out internal node locations; accounts for Earth's curvature.	Assumes a random walk process, inappropriate for modern air travel.

Experimental Protocols and Methodologies

Protocol 1: Prospective Re Estimation from Surveillance Data

This protocol, as applied to the 2009 H1N1 data in Hong Kong, can be adapted for real-time monitoring [33].

• Data Collection: Gather individual patient data including illness-onset date, laboratory-confirmation date, hospital-admission date (if applicable), and an indicator for recent travel. Data should be curated in a centralized database (e.g., the "e-flu" database used in Hong Kong).
• Data Pre-processing:
  • Estimate the distribution of delays between illness onset and case notification.
  • Classify cases as imported or local. In the model, imported cases are incorporated as infectors but not infectees to prevent overestimation of local Re.
  • For cases with missing illness-onset dates, use multiple imputation techniques to handle missing data.
• Model Fitting and Re Calculation:
  • Extend methodologies like that of Cauchemez et al. to account for reporting delays and repeated importations.
  • Assume a distribution for the serial interval (time between symptom onset in infector and infectee). For influenza, a Weibull distribution with a mean of 3.2 days and standard deviation of 1.3 days is commonly used. Perform sensitivity analyses with alternative serial intervals (e.g., 2.6 and 3.6 days).
  • Estimate Re over time using statistical software such as R. The estimation can be performed on different data streams, including case notifications and hospitalizations, to compare results.
• Validation: Compare real-time estimates with final estimates calculated after the epidemic wave has passed to assess the accuracy and potential lag in prospective analysis.

Protocol 2: Phylogeographic Re Estimation using SBDM

This protocol outlines the steps for inferring spatial spread and Re using the BDMM package in BEAST2 [4].

• Sequence and Data Alignment:
  • Collect influenza virus sequences from public databases (e.g., GISAID) or original samples, ensuring they are annotated with precise collection dates and discrete location traits (e.g., country, region).
  • Perform multiple sequence alignment using tools like MAFFT or Clustal Omega.
• Model Setup in BEAST2:
  • Install the BDMM package.
  • Specify the discrete location trait for each sequence.
  • Define the model parameters, including the birth rate (speciation/transmission), death rate (extinction/recovery), and migration rates between locations.
  • If the outbreak spans distinct periods (e.g., pre- and post-intervention), define multiple epochs with different rate parameters.
  • Select a suitable clock model (e.g., relaxed log-normal clock) and a prior for the tree (the birth-death model itself will serve as the tree prior).
• Prior Specification:
  • Apply strong priors on at least one model parameter to aid convergence, as recommended for BDMM [4]. For example, informed priors on the clock rate or recovery rate can be set based on previous literature.
• MCMC Run and Diagnostics:
  • Run a Markov Chain Monte Carlo (MCMC) analysis for a sufficient number of steps to ensure effective sample sizes (ESS) for all parameters of interest are >200.
  • Use Tracer to assess convergence and log file statistics.
• Post-processing and Interpretation:
  • Use TreeAnnotator to generate a maximum clade credibility tree from the posterior tree distribution.
  • Visualize the spatiotemporal spread and estimate Re through time and across locations using the inferred parameters. Re can be derived from the model's birth and death rates, considering the population structure.

Diagram 1: SBDM Phylogeographic Workflow

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Research Reagents and Computational Tools

Item / Resource	Function / Application	Example / Note
BEAST2 Software Package	Open-source, cross-platform program for Bayesian phylogenetic analysis of molecular sequences.	Core platform for implementing SBDM (BDMM), discrete, and continuous trait models [4].
BDMM (Birth-Death Migration Model) Package	Specific BEAST2 package for implementing Structured Birth-Death Models.	Models transmission in structured populations; handles multiple epochs [4].
MASCOT Package	BEAST2 package for approximating the structured coalescent.	Allows migration rates to be informed by covariates (e.g., flight data) via GLM [4].
R Statistical Software	Environment for statistical computing and graphics.	Used for estimating Re from case data, statistical analysis, and visualization [33].
GISAID Database	Global initiative on sharing all influenza data; primary repository.	Source for annotated influenza virus sequences for phylogeographic analysis.
Tracer Tool	For analyzing the output of MCMC runs.	Assesses convergence (ESS values) and summarizes parameter estimates.
Serial Interval Distribution	Key parameter for estimating Re from case data.	Often modeled as a Weibull distribution (e.g., mean=3.2 days, sd=1.3 days for H1N1) [33].

This case study demonstrates that the choice of model for estimating the effective reproductive number (Re) of influenza is contingent on the research question, data availability, and the scale of outbreak investigation. Structured Birth-Death Models (SBDM), as implemented in the BDMM package, offer a powerful framework for inferring transmission dynamics within structured populations and across different epidemic epochs, making them particularly suitable for investigating complex, multi-wave outbreaks. However, they can be computationally demanding and require careful prior specification. For analyses focusing on discrete trait analysis, models like the Discrete Trait Model or MASCOT provide efficient and insightful alternatives, especially when the number of demes is large or when incorporating travel data into the model.

Prospective estimation of Re from case surveillance data remains a highly feasible and valuable tool for real-time public health response, providing robust estimates that can be cross-validated with hospitalization data. Ultimately, a pluralistic approach that leverages the strengths of both statistical models applied to surveillance data and phylogenetic models like SBDM applied to genetic data will provide the most comprehensive understanding of influenza transmission dynamics to inform control strategies and drug development.

Phylodynamics, the study of how evolutionary, ecological, and immunological processes shape phylogenetic trees, has become a cornerstone of modern molecular epidemiology and viral evolutionary studies. The core challenge in this field lies in selecting and applying the correct statistical models and software tools to convert genetic sequence data into meaningful biological insights. The structured birth-death model offers a powerful framework for inferring population dynamics and transmission flows across different types or locations, directly integrating population structure into the tree-generating process. In contrast, discrete trait analysis provides a more flexible, often computationally lighter, approach to reconstruct the history of trait evolution, such as geographic location or host species, along a phylogeny. The choice between these modeling paradigms can significantly impact the conclusions of a study, influencing estimates of transmission rates, reproductive numbers, and the history of spatial spread.

This guide provides a comparative analysis of a modern software toolkit centered on BEAST X, the latest version of the Bayesian Evolutionary Analysis by Sampling Trees platform. We objectively evaluate its performance and integration with the specialized bdmm (and its successor, BDMM-Prime) package for structured models, and the essential post-processing tool TreeAnnotator. Aimed at researchers and drug development professionals, this review synthesizes current experimental data and methodological protocols to inform tool selection for contemporary phylodynamic research, framed within the broader thesis of discrete trait analysis versus structured birth-death models.

The phylodynamic software ecosystem is built on core inference engines, extended via specialized packages, and supplemented with utilities for summarizing results.

Table 1: Core Software Toolkit for Phylodynamic Analysis

Tool Name	Primary Function	Key Strengths	Latest Version & Context
BEAST X	Bayesian phylogenetic & phylodynamic inference via MCMC and HMC sampling.	New Hamiltonian Monte Carlo (HMC) samplers; advanced clock & substitution models; scalable non-parametric tree priors [7].	2025 Release (v2.8+); Major update over BEAST 2.5 [35] [7].
BDMM-Prime	Phylodynamic inference under multi-type birth-death models with migration.	Epidemiological parameterization; type-dependent skyline parameters; efficient ancestral state sampling [36].	Fork of BDMM; integrated with BEAST 2.7 [36].
TreeAnnotator	Summarizes a posterior sample of trees to produce a maximum clade credibility tree.	Annotates nodes with mean heights and posterior supports; essential for visualizing posterior distributions [37] [9].	Standard component of the BEAST software package [37].

Key Advances in BEAST X

BEAST X introduces several thematic advances over its predecessors. Thematically, it focuses on state-of-science, high-dimensional models and new computational algorithms to accelerate inference [7].

• Novel Substitution Models: It incorporates Markov-modulated models (MMMs), a class of mixture models that allow the substitution process to change across each branch and site independently. Furthermore, random-effects substitution models extend standard continuous-time Markov chain (CTMC) models to capture a wider variety of substitution dynamics, such as non-reversibility in SARS-CoV-2 [7].
• Advanced Clock Models: Enhancements include a more tractable shrinkage-based local clock model, a time-dependent evolutionary rate model, and a continuous random-effects clock model. These provide richer biological features to capture rate heterogeneities across the tree [7].
• Computational Efficiency: A primary innovation is the implementation of preorder tree traversal algorithms that, combined with postorder algorithms, enable linear-time evaluations of high-dimensional gradients. This, in turn, powers new Hamiltonian Monte Carlo (HMC) transition kernels, leading to substantial increases in effective sample size (ESS) per unit time compared to conventional Metropolis-Hastings samplers [7].

Table 2: Performance Comparison of Samplers in BEAST X for a SARS-CoV-2 Dataset

Model Component	Sampler Type	Effective Sample Size (ESS)/hour	Relative Speedup
Nonparametric Skygrid	Metropolis-Hastings	12.5	1.0x (Baseline)
Nonparametric Skygrid	HMC (BEAST X)	248.7	~20x [7]
Mixed-Effects Clock	Metropolis-Hastings	18.1	1.0x (Baseline)
Mixed-Effects Clock	HMC (BEAST X)	362.0	~20x [7]

BDMM-Prime for Structured Birth-Death Models

The BDMM-Prime package is a hard fork of the original BDMM package, designed for phylodynamic inference under multi-type birth-death models. It is particularly relevant for researchers comparing structured models to discrete trait analysis. Key enhancements include an improved BEAUti interface, automatic fall-back to analytical solutions for single-type analyses, and the use of stochastic mapping for efficient sampling of ancestral types [36]. Its model allows for type-dependent parameters, such as the effective reproductive number (Re) and sampling proportion, which can change in a piecewise-constant fashion through time (as skyline parameters), offering fine-grained insight into population dynamics [36].

Experimental Protocols and Application

To ground this comparison, we outline a standard experimental protocol for a phylodynamic analysis using this toolkit, which can be adapted for either a structured birth-death or a discrete trait analysis.

Protocol: Phylodynamic Analysis of Influenza H3N2 Data

This protocol is based on tutorials for bdmm and BDMM-Prime [9] [36].

1. Software Installation & Data Preparation

• Tools Required: Install BEAST X, the BDMM-Prime package (via BEAUti's package manager), and associated tools (TreeAnnotator, Tracer) [36].
• Data: Use an alignment, such as the 60-sequence H3N2 influenza HA dataset (h3n2_2deme.fna). Sequence headers should contain information for extracting dates and locations/traits (e.g., ID_Location_Date) [9] [36].

2. Configuring the Analysis in BEAUti

• Load Data & Template: In BEAUti, load the FASTA file. For a structured analysis, select the MultiTypeBirthDeath template.
• Tip Dates: In the "Tip Dates" panel, use the "Auto-configure" function to parse sampling dates from sequence names (e.g., using everything after last "_").
• Tip Locations/States: In the "Tip Locations" (or "Traits") panel, use the "Guess" function to parse the location/trait state (e.g., group 2 when splitting on "_") [9]. For a Discrete Trait Analysis (DTA), this would be configured as a discrete trait.
• Site Model: Select an appropriate substitution model (e.g., HKY) and account for site heterogeneity (e.g., 4 Gamma categories) [36].
• Clock Model: For a simple tutorial analysis, a Strict Clock with a known approximate rate (e.g., 0.005 subs/site/year for influenza) can be used. For publication, a relaxed clock is often preferred [36].
• Priors (Tree Model): This is the critical choice point.
  • For a Structured Birth-Death Model: In the Priors panel, select BDMMPrime as the tree prior. Expand its settings to use the "Epi Parameterization". Configure skyline parameters (Re, sampling proportion, become uninfectious rate) to be type-dependent and change over time as needed [36].
  • For a Discrete Trait Analysis: A different tree prior (e.g., Coalescent Bayesian Skyline) would be selected. The discrete trait (e.g., location) is then added as a separate trait evolution model, typically using a continuous-time Markov chain (CTMC) to model transitions between states [7].

3. Running the Analysis and Post-Processing

• Execute: Run the generated XML file in BEAST X. Monitor log files for convergence and adequate ESS values (>200).
• Summarize Output:
  • Use Tracer to analyze parameter estimates from the log file.
  • Use TreeAnnotator to generate a maximum clade credibility tree from the posterior tree set, discarding an appropriate burn-in.
  • Visualize the final, annotated tree in a viewer like IcyTree [37] [9].

The logical relationship and data flow between these components and the two main analytical paths can be visualized as follows:

Research Reagent Solutions for Phylodynamics

A successful phylodynamic study relies on a suite of computational "reagents" – the software, packages, and data that form the basis of the analysis.

Table 3: Essential Research Reagents for Phylodynamic Analysis

Reagent Solution	Function	Role in Analysis
BEAST X (Core Platform) [7]	Bayesian MCMC/HMC Inference Engine	Performs the core statistical sampling from the posterior distribution of trees and model parameters.
BDMM-Prime Package [36]	Structured Phylodynamic Model	Implements the multi-type birth-death model for inferring population dynamics and migration between types.
BEAUti Configuration Tool [37]	Graphical Analysis Setup	Generates the XML configuration file that defines the entire model, data, and prior settings for BEAST.
TreeAnnotator [37]	Tree Summary Utility	Summarizes the posterior sample of trees into a single target tree for visualization and interpretation.
Tracer [37]	Parameter & MCMC Diagnostic Tool	Assesses convergence and mixing of MCMC chains and summarizes posterior estimates of numerical parameters.
Discrete Trait CTMC Model [7]	Trait Evolution Model	Models the evolution of a discrete characteristic (e.g., location) along the branches of a phylogeny.

Discussion: Structured Models vs. Discrete Trait Analysis

The choice between a structured birth-death model (e.g., in BDMM-Prime) and a discrete trait analysis is fundamental and hinges on the research question and underlying biological assumptions.

• Structured Birth-Death Models (BDMM-Prime): These models explicitly assume that the population structure itself shapes the genealogy. The birth (transmission), death (recovery/removal), and migration rates are direct parameters of the tree-generating process. This makes them mechanistically rich and ideal for answering questions about within- and between-population dynamics, such as estimating type-specific effective reproductive numbers (Re) and migration rates [36]. The trade-off is that they are often more computationally demanding and make stronger assumptions about the underlying population dynamics.

• Discrete Trait Analysis (DTA): In a DTA, the genealogy is typically estimated first under a coalescent or birth-death model that does not account for the trait. The history of the discrete trait (e.g., geographic location) is then reconstructed upon the fixed tree using a CTMC model. This approach is more phenomenological and flexible, as it does not assume the trait directly influenced the tree's shape. It is well-suited for reconstructing ancestral states and testing for prior exposure to certain conditions [7]. However, it can be more sensitive to sampling biases and may not directly estimate key epidemiological parameters.

The experimental data shows that BEAST X's new HMC samplers significantly accelerate inference for complex models like the non-parametric skygrid and mixed-effects clocks, achieving up to 20x speedups in ESS per hour [7]. This performance enhancement benefits both modeling paradigms but is particularly impactful for the computationally intensive structured models. For researchers focused on obtaining the most accurate estimates of migration and population dynamics where the structured model assumptions hold, BDMM-Prime within the BEAST X framework is a powerful choice. For studies focused on trait history reconstruction on a shared background phylogeny, a discrete trait analysis leveraging BEAST X's new clock and substitution models may be more appropriate. Ultimately, the toolkit's power lies in its flexibility, allowing researchers to select and efficiently implement the model that best fits their specific hypothesis.

Overcoming Challenges: Bias, Computational Limits, and Model Selection

In phylogenetic studies of infectious diseases, understanding the geographic origin and spread of pathogens is a critical public health objective. Two primary methodological frameworks—Discrete Trait Analysis (DTA) and Structured Birth-Death Models (SBDM)—are commonly employed for these phylogeographic inferences. A significant challenge affecting both approaches is geographic sampling bias, where uneven surveillance across regions leads to incomplete or non-representative sequence data. This systematic review objectively compares how DTA and SBDM perform under such biased sampling conditions, synthesizing current evidence on their robustness, accuracy, and applicability for researchers and drug development professionals. The analysis is situated within the broader methodological debate concerning model-based approaches to reconstructing epidemiological dynamics from genetic data.

Performance Comparison Under Sampling Bias

The table below summarizes the core characteristics and documented performance of DTA and SBDM when faced with geographically incomplete data.

Table 1: Performance Comparison of DTA and SBDM Under Geographic Sampling Bias

Feature	Discrete Trait Analysis (DTA)	Structured Birth-Death Models (SBDM)
Core Methodology	Models trait evolution (e.g., location) along phylogeny branches using continuous-time Markov chains (CTMC) [7] [6].	Integrates population dynamics with phylogenetic trees, modeling birth, death, and sampling rates across subpopulations [7].
Handling of Sampling Bias	Highly sensitive; inferred root state and transition rates can be strongly biased towards locations with higher sampling intensity [6].	Better accounts for bias; can explicitly model and correct for preferential sampling as a function of time and location [7] [38].
Root State Classification Accuracy	Most accurate at intermediate data set sizes; accuracy decreases with larger state spaces and data sets due to model overconfidence [6].	Aims to provide more robust estimates of the origin by directly incorporating sampling effort into the model [7].
Key Performance Metric	Kullback-Leibler (KL) divergence increases with data set size and state space, but this does not correlate with higher root state accuracy [6].	Improved statistical fit and more realistic epidemiological parameter estimates when sampling heterogeneity is modeled [7].
Computational Considerations	Less computationally intensive than SBDM, but new HMC techniques in platforms like BEAST X are improving scalability [7] [6].	More parameter-rich and computationally demanding, but modern inference techniques (e.g., HMC) enhance feasibility [7].

Experimental Protocols and Evidence

The comparative performance outlined in Table 1 is supported by specific experimental investigations and model enhancements.

Evidence on DTA Performance from Simulation Studies

A key study evaluated DTA's performance in root state classification by analyzing simulated DNA datasets while progressively increasing (i) the number of sequences and (ii) the number of possible discrete trait values (state space) [6].

• Protocol: The researchers performed phylogeographic inference under a discrete trait model on these simulated datasets. They then measured the model's accuracy in identifying the true root state (geographic origin) and calculated the Kullback-Leibler (KL) divergence, a metric of model confidence [6].
• Findings: The study revealed a non-intuitive relationship: root state classification accuracy was highest at intermediate sequence data set sizes and decreased with larger datasets. Furthermore, while the KL divergence increased with both data set size and state space—suggesting greater model confidence—this metric was not a supported predictor of actual accuracy [6]. This indicates that relying on KL divergence for empirical data can be misleading, as it may inflate support for models with finer spatial discretizations and more data, even when the root state inference is incorrect.

Advancements in SBDM to Mitigate Bias

Recent developments in software like BEAST X have introduced more sophisticated SBDMs designed to address the limitations of DTA.

• Protocol for Enhanced SBDM: New episodic birth-death sampling models can explicitly account for preferential sequence sampling as a function of time and location [7] [38]. Furthermore, for discrete phylogeography, transition rates between locations can be parameterized as log-linear functions of environmental or epidemiological predictors (e.g., travel volume, population size). When predictor values are missing—a common issue in geographically incomplete data—BEAST X uses a novel Hamiltonian Monte Carlo (HMC) approach to integrate out this missing data within the Bayesian inference procedure [7].
• Findings: These modeling advances allow SBDMs to partially correct for geographic sampling bias. By incorporating direct measures of sampling effort or covariates that influence dispersal, SBDMs reduce the spurious signal of transition rates and root state locations being drawn towards well-sampled areas, leading to more reliable phylogeographic reconstructions [7].

Workflow and Methodological Diagrams

The following diagrams illustrate the core workflows for DTA and the more advanced SBDM, highlighting key steps where sampling bias enters and is handled.

Discrete Trait Analysis (DTA) Workflow

Diagram 1: The DTA workflow, showing how sampling bias is introduced early and propagates through the analysis, potentially biasing the inferred root state.

Structured Birth-Death Model (SBDM) Workflow

Diagram 2: The SBDM workflow, demonstrating the integration of external data to explicitly model and correct for sampling bias throughout the inference process.

Successful implementation of phylogeographic models, particularly for bias correction, relies on a suite of computational tools and data resources.

Table 2: Key Research Reagents and Solutions for Phylogeographic Analysis

Tool/Resource	Function	Relevance to Sampling Bias
BEAST X	A leading open-source software platform for Bayesian phylogenetic, phylogeographic, and phylodynamic inference [7].	Implements advanced SBDMs and HMC samplers to correct for preferential sampling and handle missing data in predictors [7].
Structured Birth-Death Models	A class of phylodynamic models within BEAST X that describe how populations grow, spread, and are sampled across structured locations [7] [38].	The core modeling framework for directly incorporating and correcting for heterogeneous sampling across geographic regions [7].
Hamiltonian Monte Carlo (HMC)	A modern Markov Chain Monte Carlo (MCMC) algorithm that uses gradient information for efficient sampling of high-dimensional model parameters [7].	Enables feasible inference under complex, bias-correcting SBDMs that were previously computationally infeasible [7].
Environmental & Epidemiological Predictors	Covariate data (e.g., travel flux, population density) used to explain transition rates between locations in a GLM framework [7].	Reduces reliance on sampling intensity alone to infer migration patterns, improving model realism and robustness to bias [7].
Sampling Effort Metadata	Data quantifying the intensity and distribution of pathogen surveillance efforts across different geographic regions.	Critical, often external, data required to parameterize sampling models within SBDMs and accurately correct for bias [7] [39].

Geographic sampling bias presents a fundamental challenge for phylogeographic inference. The evidence indicates that while Discrete Trait Analysis (DTA) is a accessible and widely used method, it is highly susceptible to this bias, which can lead to misleading conclusions about the geographic origin of an outbreak, especially with large but unevenly sampled datasets [6]. In contrast, Structured Birth-Death Models (SBDM) represent a more advanced framework that, through the explicit modeling of sampling heterogeneity and the integration of epidemiological predictors, offers a powerful approach to correct for this bias and achieve more reliable reconstructions of pathogen spread [7]. The choice between these models involves a trade-off between computational complexity and analytical robustness. For applied public health and drug development professionals seeking to understand outbreak origins for intervention strategies, the investment in employing bias-corrected SBDMs is strongly justified.

The explosion of genomic data presents unprecedented opportunities and challenges for evolutionary biology and drug development. Researchers tracing the spread of pathogens or analyzing trait evolution increasingly rely on complex phylogenetic models that integrate sequence data with discrete traits, such as geographic location or transmission status. Two predominant frameworks have emerged for this integration: discrete trait analysis (DTA) and structured birth-death models. As dataset sizes grow, the computational burdens of these methods diverge significantly, influencing their practical application in large-scale genomic studies. This guide objectively compares the performance and scalability of these approaches, providing experimental data and methodologies to inform model selection for contemporary genomic challenges.

Structured models explicitly incorporate population structure into the tree-generating process itself, either through the structured coalescent or multi-type birth-death processes. These models are often more biologically realistic but computationally intensive [12]. In contrast, DTA operates as a "trait evolution" model layered onto a pre-existing phylogenetic tree, making it computationally faster but potentially more sensitive to sampling biases [12] [6]. The choice between these methods increasingly hinges on their performance and feasibility with the large datasets generated by modern genomic surveillance.

Methodological Comparison: Discrete Trait Analysis vs. Structured Models

Core Computational Principles

Discrete Trait Analysis (DTA) models the evolution of a discrete characteristic, such as geographic location, along the branches of a fixed phylogenetic tree. It uses a continuous-time Markov chain (CTMC) to describe the rate of change between discrete states [12] [7]. Because it conditions on a single tree and does not model how the tree itself is generated, DTA is generally computationally efficient. However, this simplification can make its inferences vulnerable to biased sampling of sequences from different locations or populations [12] [6].

Structured Models (including the Structured Coalescent and Structured Birth-Death models) jointly infer the phylogenetic tree and the discrete trait history. Unlike DTA, they are tree-generating processes that account for the fact that lineages exist in different sub-populations (demes) and can only coalesce when they are in the same deme [12] [9]. This explicit modeling of population structure is more biologically realistic and can be less susceptible to certain sampling biases, but it requires calculating the probability of all possible migration histories for every lineage, a process that scales poorly with increasing numbers of demes or sequences [12].

Quantitative Performance Metrics

The table below summarizes key performance characteristics based on published evaluations and simulations.

Table 1: Performance Comparison of Phylogeographic Models

Feature	Discrete Trait Analysis (DTA)	Structured Birth-Death/Coalescent
Computational Speed	Faster; efficient for initial exploratory analysis on large datasets [12]	Slower; can become prohibitive for very large datasets or many demes [12]
Sampling Bias Sensitivity	High; can produce significantly biased results with uneven sampling across demes [12] [6]	Lower; more robust to uneven sampling, though not immune [12]
Biological Realism	Lower; models trait evolution conditional on a tree, not a tree-generating process [12]	Higher; explicitly models how population structure shapes the phylogeny [12]
Inference of Migration Dynamics	Can be inaccurate under biased sampling [12]	More accurate recovery of migration rates when population dynamics are modeled [12]
Root State Classification	Performance is highest at intermediate dataset sizes; accuracy can decrease with very large state spaces [6]	Improved accuracy by jointly modeling population and migration dynamics [12]

Experimental Protocols for Model Evaluation

Simulation-Based Benchmarking

To quantitatively assess the performance and biases of these models, researchers routinely employ simulation studies. A typical workflow is outlined below.

Figure 1: Workflow for simulation-based model benchmarking.

Step 1: Simulate Ground-Truth Data. Using a known model and parameters, researchers generate synthetic phylogenetic trees and sequence data. For spatiotemporal dynamics, this is often done with a structured coalescent or birth-death model in a simulator like MASTER [12]. Parameters include effective population sizes per deme, migration rates, and a molecular clock rate.

Step 2: Introduce Sampling Bias. To test model robustness, sequences are subsampled from the simulated dataset, often in a biased manner (e.g., disproportionately from one location) to mimic real-world surveillance inequalities [12] [6].

Step 3: Perform Inference. The subsampled dataset is analyzed using both DTA and structured models (e.g., in BEAST 2 or BEAST X). For structured models, this may involve newer approaches like MASCOT-Skyline, which infers non-parametric population size changes over time alongside migration [12].

Step 4: Compare to Ground Truth. The inferred parameters (e.g., migration rates, root state, population sizes) are compared to the known values from the simulation. Metrics include root state classification accuracy and the Kullback-Leibler (KL) divergence between true and inferred migration rates, though KL should not be used as a sole metric for empirical data accuracy [6].

Performance Evaluation on Empirical Data

Protocol: A common approach is to use a well-studied empirical dataset, such as global influenza A/H3N2 sequences or SARS-CoV-2 genomes, annotated with geographic and temporal data [12] [9] [7].

• Data Curation: Assemble a dataset of viral sequences with associated discrete traits (e.g., location).
• Model Fitting: Analyze the data using both DTA and structured models (e.g., bdmm or MASCOT in BEAST 2) [9].
• Convergence Assessment: Run Markov Chain Monte Carlo (MCMC) analyses for a sufficient number of steps, using tools like Tracer to ensure parameter effective sample sizes (ESS) exceed 200.
• Model Comparison: Compare models using marginal likelihood estimates (e.g., path sampling/stepping stone sampling) or information criteria to assess which model better explains the empirical data.
• Analysis of Inferred Dynamics: Compare the phylogeographic spread and ancestral state reconstructions from each method. Studies have shown that structured models can reveal different and potentially more accurate spatial dynamics compared to DTA, especially when sampling is unbalanced [12].

Computational Strategies for Scaling Analysis

Software and Algorithmic Innovations

Recent advances in Bayesian evolutionary analysis software directly address scalability hurdles. The next-generation platform BEAST X introduces several key innovations that benefit both DTA and structured models [7].

Table 2: Computational Advancements in BEAST X for Large Datasets

Innovation	Description	Impact on Scalability
Hamiltonian Monte Carlo (HMC)	A powerful MCMC sampler that uses gradient information to traverse high-dimensional parameter spaces more efficiently [7].	Dramatically increases effective sample size (ESS) per unit time for many parameters; shown to be up to 13x faster for clock models and 9x faster for skygrid models [7].
Linear-Time Gradient Algorithms	New preorder tree traversal algorithms calculate gradients for branch-specific parameters in time linear to the number of taxa (O(N)) [7].	Enables the application of HMC to very large trees, making advanced models feasible for big datasets.
Scalable Relaxed Clock Models	Shrinkage-based local clock and mixed-effects clock models that are more tractable and interpretable for large trees [7].	Improves inference of rate heterogeneity across large phylogenies without prohibitive computational cost.
Enhanced Structured Coalescent	Methods like MASCOT-Skyline that marginalize over migration histories and model population sizes non-parametrically [12].	Reduces computational burden of the structured coalescent, allowing joint inference of spatial and temporal dynamics.

Workflow for Efficient Large-Scale Inference

The diagram below illustrates a modern workflow designed to leverage these innovations for scaling phylogenetic analysis.

Figure 2: A scalable workflow for phylogeographic analysis.

• Start with DTA for Exploratory Analysis: Given its speed, use DTA on a large dataset to get an initial understanding of phylogeographic patterns and to check for obvious sampling biases [12].
• Leverage Advanced Samplers in BEAST X: For final, high-confidence inferences, use BEAST X with HMC transition kernels enabled. This is particularly beneficial for structured models, which have high-dimensional parameter spaces [7].
• Utilize Cloud Computing Resources: The computational demands of these analyses, especially with HMC, require scalable infrastructure. Cloud platforms (AWS, Google Cloud) provide the necessary on-demand resources for analyzing large datasets [40].
• Validate with Posterior Predictive Simulations: Check the model's adequacy by simulating new data under the inferred parameters and comparing key statistics (e.g., tree balance, trait distribution) to the empirical data [7].

The Scientist's Toolkit: Essential Research Reagents

The table below lists key software and data resources essential for conducting comparative analyses of phylogeographic models.

Table 3: Key Research Reagents for Phylogeographic Model Comparison

Tool / Resource	Type	Function in Analysis
BEAST 2 / BEAST X [12] [7]	Software Platform	Primary engine for Bayesian phylogenetic, phylogeographic, and phylodynamic inference. Supports both DTA and structured models.
MASTER [12]	Software Package	Plugin for simulating evolutionary processes under complex models (e.g., structured coalescent), used for benchmarking.
MASCOT & MASCOT-Skyline [12]	Software Package	BEAST 2 package for implementing the marginal structured coalescent, enabling joint inference of population and migration dynamics.
bdmm [9]	Software Package	BEAST 2 package for performing multi-type birth-death model inference.
Tracer [9]	Software Tool	For analyzing the output of MCMC runs, assessing convergence, and summarizing parameter estimates.
IcyTree [9]	Software Tool	For rapid visualization and annotation of phylogenetic trees.
Structured Genomic Datasets (e.g., H3N2, SARS-CoV-2) [12] [9]	Empirical Data	Annotated sequence datasets with discrete traits (e.g., location, host) used for empirical model testing and validation.

The strategic selection between discrete trait analysis and structured models is no longer solely a question of biological realism but is increasingly dictated by computational scalability. For large genomic datasets, a hybrid, iterative approach is often most effective: leveraging the speed of DTA for initial exploration and the accuracy of structured models, powered by modern computational innovations like HMC in BEAST X, for final inference. As genomic datasets continue to grow in size and complexity, the ongoing development and application of these scalable computational strategies will be critical for unlocking accurate insights into the spread of infectious diseases and the dynamics of evolution.

In the field of computational phylogenetics, researchers modeling pathogen spread face a fundamental challenge: how to parameterize models with sufficient complexity to capture real-world dynamics without introducing overfitting. This dilemma is particularly acute when choosing between two prominent classes of phylogeographic models—discrete trait analysis (DTA) and structured birth-death models (BDMM). Both approaches aim to reconstruct spatial transmission dynamics from genetic sequence data, but they differ significantly in their underlying assumptions, parameterization, and susceptibility to overfitting [4] [1]. The core challenge lies in ensuring parameter identifiability—the ability to uniquely determine parameter values from available data—while avoiding model overparameterization that can lead to biologically implausible conclusions [41] [42].

Parameter identifiability analysis provides crucial mathematical tools to address these challenges, distinguishing between structurally identifiable parameters (theoretically determinable from perfect data) and practically identifiable parameters (estimable with precision given real-world data limitations) [41] [42]. For researchers and drug development professionals applying these models to track outbreaks or design interventions, understanding these distinctions is essential for generating reliable, actionable results.

Theoretical Foundations: Discrete and Structured Models

Discrete Trait Analysis (DTA)

Discrete trait analysis operates by treating geographical locations as discrete states (e.g., countries or regions) and models transitions between these states as a probabilistic process along phylogenetic branches [4]. The primary advantage of DTA is its relatively low computational demand and straightforward incorporation of discrete metadata, such as travel histories [1]. However, a significant limitation is that DTA "evolves along the branches without taking the tree generating process in account, which can have a big effect on the reconstruction" [4]. This methodological characteristic makes DTA particularly susceptible to misinterpretation when sampling intensity varies spatially, as it does not explicitly account for variable sampling rates between regions [1].

Structured Birth-Death Models (BDMM)

Structured birth-death models represent an alternative approach that explicitly models migration events and rates at a population level [4]. These models incorporate the tree-generating process directly into the inference, potentially providing more accurate reconstructions of population dynamics [1]. The BDMM package implementation "can distinguish different epochs and allow for different rates in each of the epochs," adding temporal dimensionality to spatial inference [4]. The key advantage of structured models is their ability to "model variable sampling between regions," making them more robust to uneven sampling patterns that commonly occur in real-world surveillance data [1].

Table 1: Fundamental Characteristics of Phylogeographic Modeling Approaches

Feature	Discrete Trait Analysis (DTA)	Structured Birth-Death Models (BDMM)
State Representation	Discrete locations (countries, regions)	Discrete locations with population structure
Computational Demand	Relatively low	High, potentially computationally intensive
Tree Process Incorporation	Does not incorporate tree generating process	Explicitly models tree-generating process
Sampling Heterogeneity	Less robust to variable sampling	More robust to variable sampling between regions
Temporal Variation	Limited inherent temporal stratification	Built-in epoch modeling with different rates
Typical Applications	Historical reconstruction, early outbreak investigation	Contemporary outbreaks with complex dynamics

Methodological Workflow in Phylogeographic Analysis

The following diagram illustrates the core decision process and methodological relationships between discrete and structured modeling approaches in phylogeographic research:

Figure 1: Methodological Decision Framework for Phylogeographic Model Selection

Parameter Identifiability: Theoretical Framework and Assessment Methods

Parameter identifiability forms the mathematical foundation for determining whether model parameters can be uniquely estimated from available data. According to formal definitions, a model is considered "formally identifiable if two different parameter vectors lead to two different outputs" [41]. This concept is typically divided into structural identifiability (theoretical determinability from perfect data) and practical identifiability (actual estimability given real data constraints) [42].

Identifiability Assessment Methods

Several computational approaches have been developed to assess parameter identifiability in biological models:

• Differential Algebra for Identifiability of SYstems (DAISY): Performs structural identifiability analysis using differential algebra to provide exact answers about global or local identifiability, assuming output variables are known at all timepoints [41].
• Sensitivity Matrix Method (SMM): Analyzes the sensitivity matrix consisting of derivatives of model outputs with respect to parameters, with local unidentifiability formally characterized by a non-trivial null space [41].
• Fisher Information Matrix Method (FIMM): Computes the Fisher information matrix for given parameters and observation times, where local unidentifiability corresponds to zero curvature of the log-likelihood surface [41].
• Aliasing Method: Provides a continuous identifiability indicator (0-100%) that characterizes similarity between parameter derivatives, with high similarity suggesting unidentifiability [41].

Table 2: Computational Methods for Parameter Identifiability Analysis

Method	Identifiability Type	Indicator Type	Mixed Effects Support	Key Characteristics
DAISY	Structural (Global/Local)	Categorical	No	Provides exact answers via differential algebra
Sensitivity Matrix (SMM)	Practical (Local)	Categorical & Continuous	No	Analyzes derivatives of outputs w.r.t. parameters
Fisher Matrix (FIMM)	Practical (Local)	Categorical & Continuous	Yes	Computes curvature of log-likelihood surface
Aliasing	Practical (Local)	Continuous	No	Characterizes similarity between parameter derivatives

Research comparing these methods suggests that "FIMM provided the clearest and most useful answers" and was the only method capable of handling random-effects parameters, which are common in complex biological models [41].

Comparative Analysis: Discrete vs. Structured Models in Practice

Performance in Root State Classification

A critical application of phylogeographic models is identifying the geographic origin of outbreaks (root state classification). Simulation studies have revealed that "phylogeographic models tend to perform best at intermediate sequence data set sizes" rather than with very small or very large datasets [6]. This non-linear relationship between data quantity and model performance has important implications for study design.

Furthermore, studies have demonstrated that "a popular metric used for evaluation of phylogeographic models, the Kullback-Leibler (KL) divergence, both increases with discrete state space and data set sizes" [6]. This creates a potential pitfall where researchers might interpret higher KL values as indicating better model performance, when in reality this metric may reflect "artificially inflated support for models with finer discretization schemes and larger data set sizes" [6].

SARS-CoV-2 Pandemic Case Studies

The COVID-19 pandemic provided a real-world testing ground for these modeling approaches. Studies of international spread revealed that "earlier lineages were highly cosmopolitan, whereas later lineages tended to be continent-specific," reflecting the impact of travel restrictions [1]. Both DTA and structured models were deployed to track this spread, with each offering distinct advantages:

• DTA applications: Used to investigate efficacy of travel restrictions in Connecticut, USA, revealing that despite flight restrictions, community transmission of recently imported lineages occurred [1].
• Structured model applications: A multitype birth-death model applied to data from Taiwan showed "a decrease in Rt throughout the early pandemic even in the absence of substantially decreased mobility" [1].

Experimental Protocols for Model Comparison

To objectively compare discrete and structured models, researchers should implement the following experimental protocol:

• Data Simulation: Generate synthetic sequence datasets with known phylogenetic relationships and spatial parameters using tools like BEAST 2's sequence simulator [4].
• Model Implementation:
  • For DTA: Apply the discrete trait model in BEAST_CLASSIC package or the approximated structured coalescent in MASCOT for larger state spaces [4].
  • For BDMM: Implement the structured birth-death model in the BDMM package with appropriate epoch specification [4].
• Identifiability Assessment: Apply FIMM or SMM methods to assess practical identifiability of key parameters [41].
• Performance Metrics: Evaluate root state classification accuracy, parameter estimate bias, and computational requirements [6].

Table 3: Essential Research Tools for Phylogeographic Model Development

Tool/Resource	Function	Application Context
BEAST 2	Bayesian evolutionary analysis	Primary platform for phylogeographic inference
BEAST_CLASSIC	Discrete trait analysis	Implements discrete trait models for phylogeography
BDMM Package	Structured birth-death model	Implementation of birth-death model with migration
MASCOT Package	Approximated structured coalescent	Handles larger state spaces than pure structured coalescent
VisId Toolbox	Parameter identifiability analysis	MATLAB toolbox for practical identifiability assessment
DAISY Software	Structural identifiability analysis	Differential algebra-based identifiability testing
SCOTTI Package	Approximate structured coalescent	Alternative approximation for structured coalescent

Advanced Considerations: Addressing Non-Identifiability and Overfitting

When parameters are not identifiable, several advanced strategies can be employed:

Reparameterization Techniques

Non-identifiable problems can often be recast through "a simple re-parameterization of the likelihood function" which allows researchers to "re-cast the problem in terms of identifiable parameter combinations" [43]. This approach maintains model complexity while focusing inference on biologically meaningful parameter combinations that can be constrained by data.

Regularization Methods

Regularization techniques add penalty terms to the estimation process to constrain parameter values and reduce overfitting. The general form of regularized estimation is:

Figure 2: Regularization Framework for Preventing Overfitting in Parameter Estimation

This regularization approach has been shown to enable "calibrating medium and large scale biological models with moderate computation times" while avoiding overfitting [42].

Experimental Design Optimization

Identifiability analysis can directly inform experimental design by identifying "the subset of identifiable parameters and their interplay" before data collection [42]. This approach allows researchers to focus resources on measuring the observables most informative for estimating critical parameters, potentially reducing experimental costs while improving model reliability.

Based on comparative analysis of discrete trait analysis and structured birth-death models, researchers should consider the following evidence-based guidelines:

• Model Selection: Choose discrete trait analysis for exploratory analysis with many demes or when computational resources are limited. Select structured birth-death models for focused studies with few demes where sampling heterogeneity is a concern [4] [1].

• Identifiability Assessment: Implement Fisher Information Matrix Method (FIMM) analysis prior to parameter estimation to detect identifiability issues, as this method provides both categorical and continuous identifiability indicators and supports random effects [41].

• Performance Validation: Be cautious when interpreting Kullback-Leibler divergence metrics, particularly with large state spaces or datasets, as this metric may artificially inflate apparent model support [6].

• Regularization Implementation: Incorporate regularization techniques into estimation procedures, particularly for models with large parameter spaces, to balance model fit with complexity and reduce overfitting [42].

• Experimental Design: Use identifiability analysis to inform data collection strategies, focusing experimental resources on measurements that constrain biologically critical parameters.

The integration of rigorous identifiability assessment with appropriate model selection represents a best practice for researchers aiming to generate reliable phylogeographic inferences for public health decision-making and drug development applications.

In Bayesian phylogenetic analysis, estimating evolutionary timelines requires models that describe how molecular substitution rates vary across lineages. The choice between strict clock and relaxed clock models is a fundamental decision, carrying significant implications for the accuracy of inferred divergence times and, consequently, for evolutionary and epidemiological conclusions. The strict clock model assumes a constant substitution rate across all branches of a phylogenetic tree, while relaxed clock models allow rate variation among lineages. In the broader context of phylogeographic model selection—particularly the debate between discrete trait analysis (DTA) and structured birth-death models—the clock assumption acts as a critical component influencing the overall robustness of inference. This guide objectively compares the performance, applicability, and accuracy of strict versus relaxed clock models to inform researchers, scientists, and drug development professionals in selecting the appropriate model for their data.

Theoretical Foundations of Clock Models

The Strict Clock Model

The strict clock model represents the simplest approach to modeling molecular evolution, operating on the assumption that the evolutionary rate is constant across all lineages. Mathematically, this means every branch in the phylogenetic tree is associated with the same substitution rate, requiring only a single parameter (μC) to represent the overall clock rate. Its simplicity offers computational efficiency and reduced parameter space, making it particularly suitable for data with low expected rate variation, such as closely related populations or recently emerged pathogens. However, its fundamental assumption is often biologically unrealistic, as empirical data frequently reveal substantial rate heterogeneity across lineages due to variations in generation time, metabolic rates, and other life-history traits.

The Relaxed Clock Model

In contrast, relaxed clock models accommodate rate variation by allowing each branch in the tree to have its own substitution rate. These models belong to two primary categories: uncorrelated relaxed clocks, where branch rates are independently and identically distributed, and correlated relaxed clocks, which assume autocorrelation between rates on adjacent branches. A common implementation draws branch-specific rates from a log-normal distribution, with a prior mean of 1 to maintain identifiability with the overall clock rate parameter. While biologically more realistic, this model's increased parameterization (one rate per branch) demands greater computational effort and risks wider credible intervals on parameter estimates. The model's flexibility, however, allows it to capture heterogeneities that a strict clock would miss, potentially preventing biased inferences.

Performance Comparison: Accuracy and Precision

The core trade-off between strict and relaxed clock models involves a balance between precision (narrower confidence intervals) and accuracy (closeness to true values). Simulation studies and empirical analyses reveal that each model excels under specific conditions, with data characteristics—particularly the degree of rate variation—determining the optimal choice.

The table below summarizes key performance metrics for strict and relaxed clock models under varying levels of rate heterogeneity, as established through simulation studies:

Table 1: Performance of strict and relaxed clock models under different levels of rate variation (σ)

Simulated Rate Variation (σ)	Clock Model	Coverage Probability*	Relative Posterior Interval Width	Suitability
Low (σ ≤ 0.1)	Strict	High (>95%)	Narrow	Superior
	Relaxed (Uncorrelated)	High (>95%)	Wider	Appropriate
Moderate (σ = 0.1 - 0.2)	Strict	Declining Significantly	Narrow but Biased	Inappropriate
	Relaxed (Uncorrelated)	High (>95%)	Wider	Superior
High (σ > 0.2)	Strict	Very Low	Narrow but Highly Biased	Inappropriate
	Relaxed (Uncorrelated)	High (>95%)	Widest	Superior

*Coverage Probability: The proportion of analyses where the true ages of all nodes on the tree are recovered within the posterior credibility intervals.

Experimental data demonstrates that the strict clock performs optimally only when rate variation is genuinely low (σ ≤ 0.1). Under these conditions, its constrained parameterization yields precise, accurate estimates with the narrowest posterior intervals. However, its performance degrades rapidly as rate heterogeneity increases. When the standard deviation of log rate (σ) exceeds 0.1, strict clock analyses show a marked decline in their ability to recover true node ages, despite maintaining deceptively narrow confidence intervals. This results in inaccurate, overly precise estimates that can mislead research conclusions [44].

Conversely, the uncorrelated relaxed clock model maintains high accuracy across all levels of rate variation, correctly capturing true divergence times even when heterogeneity is high. This robustness comes at the cost of reduced precision, as evidenced by significantly wider posterior intervals. This is a direct consequence of the model accounting for greater uncertainty in branch-specific rates. The correlated relaxed clock model often shows performance intermediate between the strict and uncorrelated relaxed models, but can struggle with high rate variation, sometimes behaving similarly to the inadequate strict clock under extreme conditions [44].

Experimental Protocols and Methodologies

Benchmarking the performance of clock models typically relies on coalescent simulations and analysis of empirical datasets with known evolutionary histories. The standard workflow involves generating sequence data under a known phylogeny with controlled levels of rate variation, then comparing the ability of different models to recover the true node ages and tree parameters.

Data Simulation Protocol

• Tree Generation: Simulate a rooted, time-measured phylogenetic tree under a coalescent or birth-death process. Studies often use shallow (e.g., Miocene or later) root ages to reflect realistic scenarios for recent radiations or emerging pathogens.
• Rate Assignment: Assign branch-specific substitution rates according to a predefined distribution. The independent rates (uncorrelated) model is commonly used, drawing log rates from a normal distribution with a mean of -0.5σ² (ensuring a mean rate of 1) and a standard deviation of σ. Varying σ (e.g., from 0.01 to 2.0) creates datasets with different levels of rate heterogeneity.
• Sequence Evolution: Evolve genetic sequences along the branches of the simulated tree using a nucleotide substitution model (e.g., Jukes-Cantor, HKY, or GTR), with the expected number of substitutions for a branch calculated as branch_length * branch_rate.

Inference and Comparison Protocol

• Bayesian MCMC Analysis: Analyze the simulated datasets using Bayesian software (e.g., BEAST 2, MCMCTREE). For each dataset, perform inference under both strict and relaxed (uncorrelated and/or correlated) clock models.
• Convergence Assessment: Ensure Markov Chain Monte Carlo (MCMC) chains have converged by checking effective sample sizes (ESS > 200) for all key parameters.
• Performance Evaluation:
  • Accuracy: Measure the difference between the posterior mean estimate of node ages and the true, simulated ages.
  • Precision: Compare the widths of the 95% highest posterior density (HPD) intervals for node ages.
  • Coverage: Calculate the proportion of analyses in which the true node age falls within the estimated 95% HPD interval.

The following workflow diagram illustrates the core steps in this benchmarking process:

The Clock Model in the Context of Phylogeographic Inference

The selection of a clock model is deeply intertwined with the choice of a phylogeographic framework—the core of the thesis contrasting Discrete Trait Analysis (DTA) with structured birth-death models.

In DTA, geographic locations are treated as a discrete trait that evolves along the branches of the tree, analogous to a nucleotide substitution model. This approach often relies on a strict clock assumption for the trait's evolution, inherently separating the process of migration from the coalescent process generating the tree. This separation, combined with strong assumptions (e.g., that sample sizes per location are proportional to population sizes), can make DTA sensitive to sampling biases and lead to unreliable estimates of migration routes and root state (origin), despite its computational speed [6] [2].

Structured models (e.g., the structured coalescent or multi-type birth-death model), by contrast, explicitly jointly model the genetic and migration processes. These models naturally incorporate a relaxed clock for sequence evolution, as the branch rates and times are co-estimated within a cohesive population-genetic framework. This integrated approach is generally more biologically realistic and robust to sampling bias, providing more reliable inferences of migration dynamics and outbreak origins, albeit at a higher computational cost [4] [2].

Therefore, a researcher prioritizing the computational speed of DTA for an initial exploratory analysis must be acutely aware that its reliability can be compromised if the underlying data violate the strict clock assumption for sequence evolution. For definitive conclusions, especially in applied public health contexts, a structured model paired with a relaxed clock is often the more prudent and accurate choice.

The Scientist's Toolkit: Essential Research Reagents and Software

Successful and accurate phylogenetic dating requires a suite of specialized software tools and an understanding of their associated analytical components.

Table 2: Key software and analytical components for clock model selection

Tool / Component	Type	Primary Function	Relevance to Clock Models
BEAST 2 [45] [4]	Software Package	Bayesian evolutionary analysis using MCMC.	Primary platform for implementing and comparing strict and relaxed clock models.
BEAUti 2 [9]	Software Tool	Graphical utility for configuring BEAST 2 XML files.	Facilitates easy setup of clock models, tree priors, and substitution models.
Tracer [9]	Software Tool	Analyses MCMC output logs.	Assesses convergence (ESS) and compares model fit (e.g., via Bayes factors).
ORC Package [45]	BEAST 2 Package	Implements optimised operators for relaxed clocks.	Dramatically improves MCMC efficiency for relaxed clock parameter estimation.
Strict Clock [44]	Model	Assumes a single, constant substitution rate.	The simpler model to use when rate variation is confirmed to be low.
Uncorrelated Relaxed Clock [45] [44]	Model	Models branch rates as independent draws from a distribution (e.g., log-normal).	The robust default for data with moderate-to-high rate variation.
Substitution Model (e.g., HKY, GTR) [9]	Model	Describes the process of nucleotide substitution.	A core model component that works in conjunction with the chosen clock model.
Tree Prior (e.g., Coalescent, Birth-Death) [9]	Model	Provides a prior distribution on the tree topology and node heights.	Works jointly with the clock model to estimate evolutionary timescales.

The choice between strict and relaxed clock models is not one-size-fits-all but should be guided by data properties and research goals. The relaxed uncorrelated clock model is generally the more robust and safer choice, protecting against severe inaccuracies that can arise from unmodeled rate variation. However, the strict clock is superior for data with minimal rate heterogeneity, yielding the most precise estimates.

To make an informed decision, researchers should:

• Perform a Likelihood Ratio Test (LRT): This classic test compares a clock-constrained tree to an unconstrained one. However, be aware that it has low power to detect low levels of rate variation (σ = 0.01-0.1) [44].
• Examine the Posterior of σ²: In a relaxed clock analysis, a posterior estimate of σ² (the variance of log rates) close to zero supports clock-like evolution, while a larger value indicates significant heterogeneity.
• Consider the Biological Context: Shallow phylogenies of recently diverged taxa or fast-evolving pathogens are more likely to conform to a strict clock than deep phylogenies spanning diverse lineages with varying life histories.

The following decision chart synthesizes these considerations into a practical workflow for researchers:

Ultimately, the impact of clock model selection cascades into downstream conclusions, especially in phylogeography. An erroneous strict clock assumption within a DTA can lead to misplaced confidence in an incorrect geographic origin, with tangible consequences for public health interventions. Therefore, benchmarking model performance and justifying the chosen clock model are not mere methodological formalities but essential steps toward reliable scientific inference.

In the field of phylogeography, which aims to reconstruct the migration history and spread of pathogens using genetic data, selecting an appropriate model is paramount for drawing accurate and reliable conclusions. The core of this choice often hinges on understanding the data requirements and performance characteristics of different modeling frameworks. This guide provides a comparative analysis of two primary approaches: Discrete Trait Analysis (DTA) and models based on the Structured Coalescent and Structured Birth-Death processes. We focus specifically on how these models perform under varying data conditions, particularly the number of genetic sequences and the number of discrete geographic or population traits (states). The findings are contextualized within a broader thesis on phylogeographic inference, highlighting critical trade-offs between model accuracy, computational feasibility, and biological realism for researchers and drug development professionals [2].

Discrete Trait Analysis (DTA) models the movement of lineages between locations as if the location were a discrete trait evolving analogously to a genetic substitution [4] [2]. This approach is computationally efficient and can handle a large number of demes. However, it operates under assumptions that can be unrealistic for population migration, such as treating subpopulation sizes as drifting over time and being sensitive to sampling biases [2].

In contrast, the Structured Coalescent (SC) model explicitly accounts for the effects of migration on the shape and branch lengths of the genealogy [2]. It assumes stable subpopulation sizes over time and constant migration rates, providing a more principled foundation rooted in population genetics. Its primary limitation has been computational expense, which becomes prohibitive with a large number of subpopulations [4] [2]. The Structured Birth-Death (SBD) model offers an alternative to the coalescent for situations where a birth-death process is a more appropriate tree prior [4].

Table 1: Core Conceptual Differences Between Phylogeographic Models

Feature	Discrete Trait Analysis (DTA)	Structured Coalescent/Birth-Death
Theoretical Basis	Analogy to trait evolution/mutation	Principled population genetics model
Computational Speed	Fast	Slow (exact), Moderate (approximations)
Handling of Sampling Bias	Highly sensitive; inferences can be misled	More robust; explicitly models population sizes
Biological Plausibility	Lower; ignores population size and coalescent process	Higher; integrates migration with demographic process
Typical Software	BEAST (BEAST_CLASSIC package)	BEAST2 (MultiTypeTree, BASTA, BDMM packages)

Quantitative Comparison: Data Requirements and Model Performance

Impact of Number of Sequences and Trait States

A critical study evaluated Bayesian phylogeographic models, specifically examining how the number of sequences and discrete trait states influence the accuracy of inferring the root state (the geographic origin of an outbreak) [6]. The key findings are summarized below.

Table 2: Impact of Data Parameters on Discrete Trait Model Performance

Data Parameter	Impact on Model Performance	Key Finding
Number of Sequences	Non-linear relationship with root state classification accuracy.	Performance peaks at intermediate sequence data set sizes; extremely large datasets do not necessarily improve and can sometimes reduce accuracy [6].
Number of Trait States	Increases the Kullback-Leibler (KL) divergence.	The KL divergence, a metric of model fit, increases with both the discrete state space and data set sizes. This can lead to artificially inflated support for models with finer discretization, which may not reflect true accuracy [6].
KL Divergence	Poor predictor of root state accuracy.	Logistic regression modeling showed that KL divergence is not supported as a predictor of model accuracy, limiting its utility for assessing performance on empirical data [6].

Comparative Performance in Pathogen Outbreak Scenarios

The choice of model can lead to dramatically different conclusions in real-world scenarios. A landmark analysis of Ebola virus genomes illustrates this stark contrast. The structured coalescent analysis correctly inferred that successive human Ebola outbreaks were seeded by a large, unsampled non-human reservoir population. In contrast, the Discrete Trait Analysis implausibly concluded that undetected human-to-human transmission persisted over four decades, a finding at odds with epidemiological knowledge [2]. This highlights that DTA can be extremely unreliable and sensitive to biased sampling, which is common in outbreak sequencing.

Experimental Protocols and Workflows

Protocol 1: Simulating Data to Evaluate Root State Classification

This protocol is derived from studies that assess the performance of phylogeographic models through simulation [6].

• Simulation Setup: Simulate multiple genetic sequence datasets using a known phylogeny and a defined migration model between discrete traits (e.g., geographic locations). The true root state is known by design.
• Parameter Variation: Systematically vary two key parameters across simulations:
  • The total number of genetic sequences in the dataset.
  • The total number of possible discrete trait values (state space).
• Phylogeographic Inference: Perform phylogeographic inference (e.g., using DTA) on each simulated dataset to estimate the root state.
• Accuracy Assessment: Compare the estimated root state against the known, simulated root state for each run. Calculate the classification accuracy.
• Model Evaluation: Model the relationship between the number of sequences, trait states, and classification accuracy using statistical methods like logistic regression. Evaluate common metrics like Kullback-Leibler divergence for their power to predict accuracy.

Protocol 2: Comparing DTA and Structured Models on Empirical Data

This protocol outlines the steps for a comparative analysis on real sequence data, as performed in studies like [2].

• Data Curation: Compile a dataset of genetic sequences annotated with discrete trait information (e.g., sampling location). Document any known sampling biases.
• Model Implementation:
  • Run Discrete Trait Analysis using software like the BEAST_CLASSIC package in BEAST2 [4].
  • Run a Structured Coalescent Approximation (e.g., BASTA, MASCOT) or a Structured Birth-Death Model (e.g., BDMM) in BEAST2 [4].
• Inference of Key Parameters: For each model, estimate:
  • The root state (location of the outbreak origin).
  • Migration rates between discrete locations.
• Result Validation and Comparison: Compare the inferred root state and migration history from both models against known epidemiological data. Assess the biological plausibility of each model's conclusions.

Workflow Diagram

The following diagram illustrates the logical workflow for selecting and applying a phylogeographic model based on data characteristics and research goals.

The Scientist's Toolkit: Essential Research Reagents and Software

Table 3: Key Software and Analytical Tools for Phylogeographic Research

Tool Name	Type	Primary Function	Key Consideration
BEAST2 [4]	Software Platform	Bayesian evolutionary analysis sampling trees; core platform for many phylogeographic packages.	The central framework for running most modern phylogeographic analyses.
BEAST_CLASSIC [4]	Software Package	Performs Discrete Trait Analysis (DTA) and continuous random walk models.	Computationally efficient but may produce biased results with biased sampling [2].
BASTA [2]	Software Package	A Bayesian structured coalescent approximation.	Offers a good balance of accuracy and computational efficiency for structured models [2].
MASCOT [4]	Software Package	Approximates the structured coalescent and allows migration rates to be informed by covariates (e.g., flight data).	Useful for testing hypotheses about factors influencing migration.
MultiTypeTree (MTT) [2]	Software Package	Implements the exact structured coalescent.	Highly accurate but computationally prohibitive with more than ~4 demes [4] [2].
BDMM [4]	Software Package	Implements the structured birth-death model.	More appropriate than coalescent models when a birth-death process is a better representation of population dynamics.

Head-to-Head Comparison: Evaluating Model Performance and Accuracy

In the field of computational biology and genetics, researchers often rely on sophisticated statistical models to decipher complex relationships from biological data. Two distinct classes of models employed for different but sometimes overlapping purposes are Discrete Trait Analysis (DTA) and Structured Birth-Death Models (SBDM). While both approaches can analyze trait evolution across species or populations, they stem from different theoretical frameworks and are optimized for different research questions. DTA primarily focuses on identifying associations between genetic markers and observable traits, particularly when those traits are categorical in nature [46]. In contrast, SBDM represents a class of phylodynamic models that reconstruct population dynamics, including speciation, extinction, and migration rates, from phylogenetic trees derived from genetic sequence data [8].

The fundamental theoretical divergence between these approaches lies in their core objectives. DTA methods, particularly genome-wide association studies (GWAS) for discrete traits, aim to connect phenotypic variation back to its underlying genetic causes [47] [46]. These models are fundamentally designed to identify statistical associations between specific genetic polymorphisms and traits of interest. On the other hand, SBDM operates under a birth-death process framework, which is a continuous-time Markov process that models how lineages speciate (birth) and go extinct (death) over evolutionary time [48]. These models are particularly powerful for quantifying past population dynamics and migration patterns in structured populations [8].

Table 1: Fundamental Characteristics of DTA and SBDM

Characteristic	Discrete Trait Analysis (DTA)	Structured Birth-Death Models (SBDM)
Primary Focus	Identify genotype-phenotype associations	Reconstruct population history and dynamics
Theoretical Basis	Statistical association tests	Birth-death stochastic processes
Data Input	Genotype and phenotype data	Genetic sequences and/or phylogenetic trees
Trait Handling	Direct analysis of categorical traits	Inference of traits affecting diversification
Evolutionary Model	Often model-free or simple evolutionary assumptions	Explicit evolutionary models (e.g., Brownian motion)

Methodological Approaches and Experimental Protocols

Discrete Trait Analysis Framework

The methodological approach for Discrete Trait Analysis, particularly in genome-wide association studies, involves specific protocols for analyzing discrete or categorical traits. The standard workflow begins with collecting genotype and phenotype data across many individuals, followed by applying statistical tests to identify significant associations between genetic markers and the traits of interest [46]. For discrete traits, researchers commonly use chi-square tests for contingency tables or logistic regression models for initial genome-wide scans [46]. These methods evaluate whether the distribution of genetic variants differs significantly between groups with different trait states.

More advanced DTA methods incorporate sophisticated statistical frameworks to enhance power and address limitations. Multi-marker methods analyze combinations of SNPs simultaneously, using approaches such as sliding windows with principal component analysis, penalized orthogonal-components regression, wavelet-based transformations, and Bayesian variable selection methods [46]. These advanced techniques help overcome the limitations of single-marker analyses by combining information across multiple genetic variants within genomic regions. Additionally, careful control for population stratification is essential in DTA protocols, as cryptic relatedness or population structure can generate spurious associations [46]. Methods such as genomic control, structured association, and principal component analysis are routinely incorporated into DTA workflows to address these confounding factors.

Structured Birth-Death Model Framework

The experimental protocol for Structured Birth-Death Models involves a fundamentally different approach centered around phylogenetic trees and population dynamics. The multi-type birth-death model with sampling is implemented in software packages such as the BEAST 2 package bdmm, which enables quantification of past population dynamics in structured populations based on phylogenetic trees [8]. The model calculates the probability density of a phylogenetic tree given population dynamic parameters through numerical integration of systems of differential equations [8].

The methodological workflow for SBDM begins with genetic sequence data collection, from which phylogenetic trees are inferred. These trees then serve as input for the birth-death model, which estimates parameters including birth rates (speciation), death rates (extinction), migration rates between subpopulations, and sampling rates through time [8]. Recent algorithmic improvements have dramatically increased the scalability of these models, allowing analysis of larger datasets containing several hundred genetic sequences while improving numerical robustness and computational efficiency [8]. The model has been extended to allow for more complex scenarios, including homochronous sampling events at multiple time points and more flexible migration rate specifications with piecewise-constant changes through time [8].

SBDM Analysis Workflow: From genetic data to population insights

Performance Comparison and Experimental Data

Analytical Strengths and Applications

Discrete Trait Analysis excels in its ability to directly connect specific genetic variants to observable discrete traits, providing clear biological interpretations for disease associations and morphological characteristics [46]. GWAS methods, a primary implementation of DTA, have successfully identified numerous genetic loci associated with complex diseases, typically with relative risks ranging from 1.2 to 1.5 for significant associations [46]. The strength of DTA lies in its straightforward framework that can be applied across large datasets with hundreds of thousands of genetic markers, offering a comprehensive view of genetic contributions to traits.

DTA methods are particularly powerful when analyzing self-fertilizing organisms like Arabidopsis thaliana, where inbred lines allow repeated phenotyping of genetically identical individuals [47]. This approach has successfully identified genetic loci underlying traits including glucosinolate levels, shade avoidance, heavy metal and salt tolerance, flowering time, and other life history traits [47]. The ability to work with existing genotype data for thousands of accessions makes DTA highly efficient for initial trait mapping studies.

Structured Birth-Death Models demonstrate unique strengths in reconstructing historical population dynamics and quantifying evolutionary processes. SBDM can estimate critical parameters such as migration rates between subpopulations, speciation and extinction rates, and how these parameters change over time [8]. These models have been successfully applied to infectious disease dynamics, helping to understand the spread of pathogens across geographic regions and between different host populations [8].

A key advantage of SBDM is their ability to incorporate population structure explicitly, allowing researchers to test hypotheses about how different subpopulations contribute to overall evolutionary dynamics. Recent improvements in SBDM algorithms have enabled analyses of larger datasets with improved precision, particularly for structured models with many inferred parameters [8]. When applied to Influenza A virus sequences, these improved models successfully revealed global migration patterns and seasonal dynamics, demonstrating their utility for understanding pathogen spread [8].

Limitations and Methodological Constraints

Discrete Trait Analysis faces several important limitations that affect its performance and applicability. The power of GWAS is highly dependent on effect size and allele frequency, with rare variants and small effect sizes presenting significant challenges for detection [47]. This limitation is particularly problematic for traits controlled by many rare variants, each with large effects, or many common variants with only small phenotypic effects [47]. Synthetic associations, where non-causative markers show stronger association with a phenotype than true causative variants due to genetic heterogeneity, can also generate false positive signals [47].

Additionally, DTA methods are sensitive to sample size and population composition. While some traits with simple architecture can be analyzed with fewer than 100 accessions, more complex polygenic traits require larger sample sizes [47]. The selection of mapping populations involves trade-offs between maximizing genetic diversity and introducing genetic heterogeneity, which can weaken correlations between phenotypes and specific variants [47].

Structured Birth-Death Models confront different sets of limitations, particularly regarding computational complexity and data requirements. Early implementations of multi-type birth-death models were numerically limited to analyzing trees with approximately 250 genetic samples due to numerical instability issues [8]. Although recent algorithmic improvements have partially addressed these limitations, computational demands remain substantial compared to simpler analytical approaches.

SBDM also face methodological constraints related to model specification and parameter identifiability. Complex models with many parameters require significant amounts of data from each subpopulation for reliable estimation [8]. Model misspecification, such as incorrect assumptions about rate constancy through time or improper structure assignment, can lead to biased estimates of evolutionary parameters. The computational intensity of these methods also limits the exploration of model space and comprehensive sensitivity analyses.

Table 2: Performance Comparison of DTA and SBDM

Performance Metric	Discrete Trait Analysis (DTA)	Structured Birth-Death Models (SBDM)
Sample Size Requirements	Varies by trait architecture; 100s to 1000s of individuals	Recently improved from ~250 to 500+ sequences
Computational Efficiency	Relatively fast; genome-wide scans in minutes on standard PCs	Computationally intensive; requires specialized software
Handling of Rare Variants	Limited power for rare variants	Can incorporate rare variants through tree structure
Population Structure Control	Requires explicit correction methods	Explicitly models population structure
Trait Architecture Insight	Better for simple architectures with large effect loci	Better for complex evolutionary dynamics

Implementation Requirements and Research Reagents

Successful implementation of both DTA and SBDM requires specific computational resources, software tools, and data quality standards. The research reagent solutions differ significantly between these approaches due to their distinct methodological foundations.

Discrete Trait Analysis relies on genotype-phenotype datasets with careful quality control procedures. Essential components include high-density SNP arrays or sequencing data, precise phenotype assessment protocols, and statistical packages capable of handling large-scale association testing. For discrete traits, specialized methods such as the FP test that exploits information in 2×3 contingency tables about inbreeding in addition to standard association tests have shown improved performance over traditional approaches [46]. Implementation typically requires software such as PLINK, R, or specialized packages that incorporate population structure control methods including principal components analysis or mixed models.

Structured Birth-Death Models require phylogenetic trees as fundamental input, either estimated from genetic sequence data or obtained from existing resources. The BEAST 2 package bdmm represents a primary implementation platform for these models, providing Bayesian framework for joint inference of phylogenetic trees and population dynamic parameters [8]. Recent improvements in bdmm have expanded its capabilities to allow homochronous sampling at multiple time points, more flexible migration rate specifications, and improved numerical stability for larger datasets [8]. Implementation typically requires substantial computational resources, particularly for Bayesian MCMC analyses that jointly estimate tree topologies and population parameters.

Table 3: Essential Research Reagents and Computational Tools

Resource Type	Discrete Trait Analysis (DTA)	Structured Birth-Death Models (SBDM)
Primary Software	PLINK, R, specialized GWAS packages	BEAST2 with bdmm package, phytools, ape
Data Requirements	Genotype data, precise phenotype measurements	Genetic sequences, sampling times, population labels
Key Statistical Methods	Logistic regression, chi-square tests, mixed models	Markov chain Monte Carlo, numerical integration
Computational Demands	Moderate; standard workstations sufficient	High; often requires cluster computing
Specialized Indices	Bayes factors, attributable risk measures	Bayesian posterior probabilities, Bayes factors

DTA Analysis Workflow: From raw data to genetic associations

Discrete Trait Analysis and Structured Birth-Death Models represent complementary approaches in the computational biology toolkit, each with distinct strengths tailored to different research questions. DTA provides powerful, relatively straightforward methodology for connecting genetic variation to discrete phenotypic traits, with particular utility for initial mapping studies and identifying candidate loci for further validation [47] [46]. In contrast, SBDM offers sophisticated framework for reconstructing historical population dynamics and evolutionary processes, with unique capabilities for modeling structured populations and temporal rate changes [8] [48].

The choice between these methodologies should be guided by specific research objectives, data resources, and analytical requirements. For researchers focused on identifying genetic variants underlying discrete traits in well-characterized populations, DTA provides efficient and powerful approach. For investigations aimed at understanding evolutionary dynamics, population history, and structured processes, SBDM offers unparalleled insights despite greater computational demands. Future methodological developments will likely continue to address current limitations in both approaches, particularly regarding rare variant detection for DTA and computational efficiency for SBDM, further expanding their utility for biological research.

Inferring the geographic origin, or root state, of pathogen outbreaks and species dispersal is a central challenge in molecular epidemiology and evolutionary biology. The accuracy of these inferences has profound implications for understanding spread dynamics and informing public health and conservation strategies. This critical task is typically accomplished using phylogenetic models that reconstruct geographic history from genetic sequence data. The methodological landscape is dominated by two principal paradigms: * Discrete Trait Analysis (DTA), which models location evolution on a fixed phylogenetic tree, and *structured models (including structured coalescent and birth-death approaches), which jointly infer the tree and location history [4]. Framed within broader thesis research comparing these approaches, this guide objectively evaluates their performance using current experimental data, revealing a nuanced picture where no single method universally outperforms others, but rather their accuracy is highly dependent on specific dataset characteristics and modeling assumptions.

Phylogeographic models form the computational engine for root state classification. These models are broadly categorized into discrete and continuous approaches, with discrete models being the primary focus for geographic origin inference when locations are grouped into distinct regions or demes.

• Discrete Trait Analysis (DTA): Implemented in the BEAST Classic package, DTA treats geographic location as a discrete trait that evolves along the branches of a phylogenetic tree according to a continuous-time Markov chain (CTMC) process [4]. A key limitation is that it models trait evolution conditional on a pre-existing tree without accounting for how the tree-generating process itself might be influenced by population structure [4]. Its popularity stems from relative computational speed and ease of use, particularly when analyzing many demes [12] [4].

• Structured Coalescent Models: These models, such as the Multi Type Tree (MTT) and the Marginal Approximation of the Structured Coalescent (MASCOT), explicitly model how lineages coalesce within and migrate between sub-populations through time [12] [4]. They are considered more biologically realistic as they jointly model the genetic and spatial processes. The pure structured coalescent (MTT) becomes computationally intractable with more than 3-4 demes, while approximations like MASCOT can handle larger numbers of demes and incorporate external data (e.g., flight passenger numbers) through Generalized Linear Models (GLM) to inform migration rates [4].

• Structured Birth-Death Models (BDMM): As an alternative to coalescent approaches, structured birth-death models implemented in packages like BDMM model lineage birth (transmission), death (recovery), and sampling across different sub-populations [4]. These can be more appropriate for certain epidemiological contexts but may require strong priors for convergence.

• Recent Advancements in BEAST X: The latest version of the BEAST software introduces significant improvements for phylogeographic inference, including more scalable computation and novel approaches to handle sampling bias—a known issue in discrete trait analysis [7]. It features Hamiltonian Monte Carlo (HMC) sampling techniques that enable faster and more efficient inference under complex models, such as those with environmental predictors of migration [7].

The table below summarizes the core characteristics of these primary model types.

Table 1: Key Phylogeographic Models for Root State Classification

Model Type	Representative Software/Package	Core Principle	Key Advantages	Key Limitations
Discrete Trait Analysis (DTA)	BEAST Classic [4]	Trait evolution on a fixed tree via CTMC	Computational speed; handles many demes [12]	Ignores tree-generating process; sensitive to sampling bias [12] [7]
Structured Coalescent	Multi Type Tree (MTT) [4]	Joint inference of tree and location via coalescent	Biologically realistic; accounts for population structure	Computationally intense (>4 demes intractable) [4]
Structured Coalescent (Approximate)	MASCOT [12] [4]	Approximates the structured coalescent	Handles more demes; GLM for migration rates [4]	Still more computationally demanding than DTA
Structured Birth-Death	BDMM [4]	Joint inference via birth-death process	Epidemiologically meaningful parameters	May require strong priors for convergence [4]

Quantitative Performance Comparison

Recent simulation studies and benchmarking efforts have provided quantitative data on the performance of these competing approaches under controlled conditions.

Accuracy in Ancestral State Reconstruction

A critical evaluation of MASCOT-Skyline versus DTA using SARS-CoV-2 sequences and Susceptible-Infected-Recovered (SIR) simulations demonstrated that the choice of model significantly impacts root state inference. The study concluded that modeling spatial and temporal dynamics jointly is crucial, even if the researcher is primarily interested in only one of these aspects [12]. Failure to do so can lead to biased estimates of migration rates and ancestral locations. Specifically, DTA models were found to be particularly susceptible to biased results under certain sampling schemes, a weakness that structured models like MASCOT-Skyline aim to mitigate [12] [7].

Robustness to Sampling Bias

Sampling bias is a pervasive challenge in phylogeography. A key finding from recent research is that Discrete Trait Analysis through CTMC is highly sensitive to geographic sampling bias [7]. While structured coalescent models offer some improvement, they do not completely account for this bias. Novel modeling strategies in BEAST X, which integrate out missing predictor data using Hamiltonian Monte Carlo, represent a significant step forward in addressing this issue and improving the robustness of root state inference [7].

Computational Performance

Computational requirements often dictate methodological choice in practice. DTA remains the fastest option, especially for analyses involving a large number of demes [4]. The pure structured coalescent (MTT) is at the other end of the spectrum, becoming computationally intractable beyond a handful of demes. Approximate methods like MASCOT and SCOTTI offer a middle ground, enabling analyses with ten or more demes [4]. The advent of more efficient inference algorithms in BEAST X, such as HMC, is narrowing the performance gap by enabling faster sampling of high-dimensional posterior distributions for complex structured models [7].

Table 2: Experimental Performance Comparison of Phylogeographic Models

Performance Metric	Discrete Trait Analysis (DTA)	Structured Coalescent (MASCOT)	Structured Birth-Death (BDMM)
Root State Accuracy (under biased sampling)	Low to Moderate (Biased) [12] [7]	High (Less Biased) [12]	Varies (Data lacking)
Migration Rate Estimation	Can be biased [12]	More accurate [12]	Varies (Data lacking)
Computational Speed	Fastest [12] [4]	Moderate [4]	Slow [4]
Scalability (Number of Demes)	High (Many demes feasible) [4]	Moderate (~10 demes) [4]	Low

Experimental Protocols for Benchmarking

To ensure the reproducibility of phylogeographic comparisons, the following section outlines standard experimental protocols derived from recent studies.

Simulation-Based Validation

Simulation studies are the gold standard for evaluating inference accuracy, as the true evolutionary history is known.

• Workflow: The typical protocol involves 1) Simulating pathogen spread using a known model (e.g., structured coalescent in MASTER [12] or Wright-Fisher coalescent in msprime [49]) with predefined population sizes, migration rates, and demographic history; 2) Generating sequence alignments from the simulated trees; 3) Performing phylogenetic inference using the methods under comparison (DTA, structured coalescent, etc.) in a platform like BEAST 2 or BEAST X [12] [7]; and 4) Comparing the inferred parameters (root state, migration rates, population sizes) to the known ground truth.
• Key Metrics: Common accuracy metrics include the Relative Root-Mean-Squared Error (RRMSE) for continuous parameters [49], the proportion of correct root state assignments, and topological distances (e.g., triplet distance) between inferred and true trees [50].

Phylogeographic Inference with Empirical Data

When applying these methods to real sequence data, a standardized workflow ensures robust results.

• Data Curation: Publicly available sequences (e.g., from GenBank or BV-BRC) are collected, followed by rigorous quality control. This includes filtering sequences with unclear metadata, removing those with excessive ambiguous nucleotides, and reducing sequence redundancy to avoid over-representation of specific lineages [51].
• Temporal Signal Assessment: The correlation between sampling date and genetic divergence (root-to-tip distance) is analyzed using tools like TempEst to identify sequences with sufficient temporal signal for accurate molecular dating [51].
• Genomic Region Selection (for RNA viruses): For resource-constrained settings or specific viruses like flaviviruses, studies show that targeting highly variable genomic regions (~2700 nt) can recapitulate phylogenies from whole-genome data, providing an efficient alternative for surveillance [51].
• Model Fitting and Comparison: Multiple models (DTA, structured coalescent) are fitted using Bayesian software (BEAST 2/X). Bayesian model selection techniques, such as comparing marginal likelihoods, can then be used to identify the model that best explains the data [7].

The Scientist's Toolkit

Successful phylogeographic analysis relies on a suite of software tools and reagents. The table below details essential solutions for conducting root state classification experiments.

Table 3: Key Research Reagent Solutions for Phylogeographic Analysis

Item Name	Category	Function Description
BEAST 2 / BEAST X	Software Platform	The leading open-source software for Bayesian phylogenetic, phylogeographic, and phylodynamic inference, hosting many of the models discussed [12] [7].
MASCOT Package	Software Plugin	Implements an approximate structured coalescent model within BEAST 2, enabling joint inference of population dynamics and migration for more demes than the pure model [12] [4].
BDMM Package	Software Plugin	Implements a structured birth-death model in BEAST 2, suitable for epidemiological analyses where birth-death assumptions are appropriate [4].
BEAST Classic Package	Software Plugin	Provides the implementation of the Discrete Trait Analysis (DTA) model for phylogeography [4].
MASTER	Software Tool	A software package for simulating evolutionary processes under detailed phylogenetic models, used for simulation-based validation [12].
msprime	Software Tool	A library for simulating ancestral recombination graphs (ARGs) and population genetic data, widely used for benchmarking [49].
TempEst	Software Tool	A utility for assessing the temporal signal in sequence data by investigating the relationship between sampling date and genetic divergence [51].
Phased Whole-Genome Sequences	Data Input	The primary input data for methods like SINGER and other ARG-inference tools; high-quality, phased data is crucial for accurate inference [50].

The accurate classification of the geographic root state from genetic data remains a challenging but essential task. This comparison guide demonstrates that the choice between Discrete Trait Analysis and structured models like the structured coalescent or birth-death is not trivial. DTA offers speed and practicality for analyses with many demes but can produce biased inferences under non-uniform sampling or complex population dynamics. In contrast, structured models provide a more biologically realistic framework that jointly infers tree and location history, generally leading to more accurate root state and migration rate estimates, albeit at a higher computational cost [12] [4].

The prevailing recommendation from recent research is to model spatial and temporal dynamics jointly where feasible, as implemented in methods like MASCOT-Skyline [12]. The ongoing development of more scalable and robust inference algorithms in platforms like BEAST X, which directly address issues like sampling bias, is steadily making these more complex models accessible for larger and more realistic datasets [7]. Ultimately, researchers should base their model selection on a careful consideration of their specific research question, the number of demes, available computational resources, and potential sampling biases, ideally using simulation studies to validate their chosen approach.

In the field of Bayesian phylodynamics, accurately reconstructing the spatial and temporal dynamics of infectious diseases from pathogen genomic data is a cornerstone of epidemiological research. The choice of phylogeographic model—specifically, between the widely used Discrete Trait Analysis (DTA) and the more computationally intensive structured models (like the structured coalescent or structured birth-death)—is critical. This guide provides an objective comparison of these model classes by focusing on two essential quantitative metrics for evaluating model performance and inference reliability: the Kullback-Leibler (KL) divergence and the Effective Sample Size (ESS). We synthesize findings from recent methodological research and software advancements to help researchers and drug development professionals select the most appropriate tool for their analyses.

Theoretical Background and Key Metrics

The Role of KL Divergence and ESS in Model Evaluation

In Bayesian inference, we rely on numerical algorithms, primarily Markov Chain Monte Carlo (MCMC), to approximate complex posterior distributions. The quality of this approximation must be measured rigorously.

• Kullback-Leibler (KL) Divergence ((D{KL})): This metric quantifies the difference between two probability distributions. In model evaluation, it can be used to measure the information loss when an approximate model is used instead of the true model. A lower (D{KL}) indicates a closer approximation. However, recent research highlights that in phylogeography, (D_{KL}) can artificially inflate with larger data sets and state spaces, making it a potentially misleading indicator of model accuracy on its own [6].
• Effective Sample Size (ESS): In MCMC analysis, ESS quantifies the number of effectively independent samples in a correlated sequence. A high ESS indicates that the chains have mixed well and explored the parameter space sufficiently, leading to reliable parameter estimates. Conversely, a low ESS suggests poor mixing and potentially biased inferences. ESS is thus a fundamental diagnostic for the practical identifiability of model parameters and computational efficiency [7].

Phylogeographic Model Classes

• Discrete Trait Analysis (DTA): This approach models the evolution of a discrete trait (e.g., geographic location) along the branches of a fixed phylogenetic tree. It is a neutral trait evolution model and is not itself a tree-generating process [12].
• Structured Models: These models, which include the structured coalescent and structured birth-death, explicitly model the population structure and lineage history within the tree-generating process itself. They co-estimate the phylogenetic tree alongside the spatial dynamics, accounting for how population structure influences the coalescence of lineages [12].

Performance Comparison: Discrete Trait Analysis vs. Structured Models

Empirical and simulation studies have consistently revealed performance trade-offs between these model classes, particularly concerning sampling bias and the accuracy of inferred parameters.

The table below summarizes the core characteristics and performance of each model class based on current research.

Table 1: Objective Comparison of Phylogeographic Model Classes

Feature	Discrete Trait Analysis (DTA)	Structured Coalescent/Birth-Death
Core Methodology	Models trait evolution on a fixed tree [12]	Co-estimates tree and spatial dynamics within a population genetic framework [12]
Handling of Sampling Bias	Highly sensitive to biased sampling; can lead to significantly biased estimates of migration rates and ancestral states [12] [7]	More robust to sampling biases; explicitly models population sizes to correct for uneven sampling [12]
Inferred Parameters	Migration rates, ancestral state probabilities (e.g., root state)	Migration rates, effective population sizes ((N_e)) per location, ancestral states
Computational Speed	Generally faster and easier to use [12]	Computationally intensive, but advancements like MASCOT improve efficiency [12]
Key Performance Limitation	Root state classification accuracy is highest at intermediate data set sizes and does not consistently improve with more data or traits [6]. KL divergence can be a misleading performance metric [6].	Assumption of constant population sizes can bias inference; newer models (e.g., MASCOT-Skyline) relax this by modeling (N_e) over time [12].
Recommended Use Case	Preliminary, exploratory analysis with well-sampled data.	Inference for publication or public health decision-making, especially with known or suspected sampling bias.

Quantitative Data on Model Performance

• Root State Accuracy: One study evaluated root state classification accuracy—a key public health question—using simulated data. It found that phylogeographic models, including DTA, perform best at intermediate sequence data set sizes. Surprisingly, accuracy does not necessarily improve with larger data sets or more discrete traits, highlighting a critical limitation of these methods for determining outbreak origins [6].
• ESS Performance in Software: Advancements in computational inference, such as the implementation of Hamiltonian Monte Carlo (HMC) samplers in BEAST X, have led to substantial increases in ESS per unit time compared to traditional Metropolis-Hastings samplers. This improvement is crucial for achieving reliable convergence in complex structured models [7].

Experimental Protocols for Performance Evaluation

To ensure the validity of phylogeographic inferences, researchers should employ the following experimental protocols, which are derived from the methodologies used in the cited studies.

Simulation-Based Model Validation

This protocol is used to test a model's ability to recover known "true" parameters under controlled conditions.

• Simulate Ground Truth Data: Use a simulation tool like MASTER [12] to generate phylogenetic trees under a known model with predefined parameters (e.g., migration rates, effective population size trajectories).
• Generate Sequence Data: Evolve genetic sequences along the simulated tree using a substitution model.
• Perform Inference: Analyze the simulated sequence data using the model to be tested (e.g., DTA or a structured model in BEAST 2 or BEAST X).
• Compare Estimates to Truth: Quantify performance by calculating the deviation of the inferred parameters (e.g., migration rates, root state) from the known simulated values. Metrics like Mean Squared Error (MSE) are commonly used.
• Vary Conditions: Repeat the process under different scenarios, such as biased sampling schemes, varying numbers of sequences, and different numbers of discrete locations, to assess model robustness [12] [6].

Benchmarking Computational Efficiency

This protocol assesses the practical feasibility of running a model.

• Define Dataset: Select a benchmark dataset (e.g., a large viral genome alignment with associated location traits).
• Configure MCMC: Run identical analyses using DTA and a structured model, ensuring the MCMC chain length and sampling frequency are the same.
• Monitor Diagnostics: After the run, calculate the ESS for all key parameters (e.g., migration rates, clock rates, tree likelihood).
• Calculate Efficiency: Determine the "ESS per hour," which measures how quickly a model produces independent samples. Models with higher ESS per hour are more computationally efficient [7].

The logical relationship between model choice, potential biases, and the evaluation workflow is summarized in the following diagram:

Essential Research Reagents and Tools

Successful phylogeographic analysis relies on a suite of software tools and computational resources. The table below details key solutions for building a robust research pipeline.

Table 2: Key Research Reagent Solutions for Phylogeographic Analysis

Tool / Resource	Function / Description	Relevance to KL Divergence & ESS
BEAST 2 / BEAST X [12] [7]	A primary software platform for Bayesian evolutionary analysis. BEAST X is the latest version with enhanced performance.	Provides the environment for running DTA and structured models. Its advanced HMC samplers in BEAST X significantly improve ESS for many parameters [7].
MASCOT (BEAST 2 Package) [12]	Implements the Marginal Approximation of the Structured COalescenT, a method that efficiently infers migration rates and population sizes.	The MASCOT-Skyline extension allows joint inference of spatial and temporal dynamics, mitigating bias and improving parameter reliability (effective sample size) [12].
MASTER [12]	A package for simulating stochastic evolutionary processes (birth-death and coalescent models) under complex scenarios.	Essential for the simulation-based validation protocol, allowing researchers to test model identifiability and quantify bias.
Tracer	A common companion tool for BEAST to analyze MCMC output.	Directly calculates ESS and other convergence diagnostics for all model parameters, crucial for assessing the quality of an inference run.
High-Performance Computing (HPC) Cluster	A computer cluster designed for heavy computational tasks.	Running structured models, especially on large datasets, is computationally intensive and often requires HPC resources to achieve convergence in a practical timeframe.

The choice between Discrete Trait Analysis and structured models involves a direct trade-off between computational speed and statistical robustness. While DTA offers a fast and accessible entry into phylogeography, its sensitivity to sampling bias can compromise the accuracy of its conclusions. Structured models, particularly modern implementations like MASCOT-Skyline in BEAST 2 and those accelerated by HMC in BEAST X, provide a more rigorous framework that accounts for population structure and sampling heterogeneity, leading to more reliable inferences for critical public health applications.

When quantifying model support, researchers must move beyond a single metric. ESS remains a non-negotiable diagnostic for MCMC reliability. In contrast, the utility of KL divergence is context-dependent; it should not be used as a sole indicator of model accuracy, especially when comparing across different discretization schemes or data set sizes [6]. A robust analytical workflow should prioritize structured models where feasible, leverage simulation studies to validate methods, and rigorously monitor ESS to ensure that inferences are both statistically and computationally sound.

Evaluating evolutionary models via simulation is a cornerstone of computational biology, providing critical insights into model accuracy, robustness, and applicability before their deployment on empirical data. For researchers investigating pathogen spread, molecular adaptation, and trait evolution, selecting a model with validated performance for a specific biological scenario is paramount. This guide objectively compares the performance of contemporary phylogenetic models—focusing on discrete trait analysis and structured birth-death frameworks—under simulated conditions with known evolutionary parameters. We synthesize current experimental data to aid researchers and drug development professionals in choosing the optimal model for their research objectives, framed within the broader thesis of discrete trait analysis versus structured birth-death models.

Comparative Performance Metrics of Evolutionary Models

The performance of evolutionary models is typically quantified using metrics such as statistical power, accuracy in parameter estimation, and computational efficiency when applied to simulated data where the true evolutionary history and parameters are known.

Table 1: Performance Metrics of Phylogenetic Signal Detection Methods in Simulation Studies

Model/Index	Trait Type Handled	Key Performance Findings (from Simulations)	Reference
M Statistic	Continuous, Discrete, & Multiple Trait Combinations	Performs well, not inferior to existing methods; effectively handles continuous variables, discrete variables, and multiple trait combinations. [52]
Blomberg's K / Pagel's λ	Continuous	Established baseline for continuous trait performance comparison. [52]
D Statistic	Binary Discrete	Applicable only to binary traits evolving under a Brownian motion threshold model. [52]
δ Statistic	Discrete	Theoretically applicable to any discrete trait without a specific requirement for the number of states. [52]

Table 2: Performance of Advanced Evolutionary Inference Frameworks

Model/Framework	Primary Application	Key Performance Findings (from Simulations & Applications)	Reference
BEAST X	Bayesian phylogenetic, phylogeographic, and phylodynamic inference	Achieves substantial increases in Effective Sample Size (ESS) per unit time compared to conventional Metropolis-Hastings samplers; scalable for large trees and state spaces. [7]
Polyepoch Clock Model	Estimating time-varying evolutionary rates	Through simulation, successfully recovers true timescales and rates under different evolutionary scenarios; captures strong time-varying patterns in empirical virus data. [53]
ProteinEvolver2 (Forecasting)	Forecasting protein evolution	Shows acceptable errors in predicting folding stability of forecasted protein variants; sequence prediction errors are larger. Feasible in evolutionary scenarios with measurable selection. [15] [14]
Structured Birth-Death with SCS Models	Forecasting protein evolution	Unifies evolutionary history simulation with molecular evolution, addressing biological incoherence of traditional two-step methods. [15] [14]

Experimental Protocols for Model Evaluation

Evaluating Phylogenetic Signal Detection with the M Statistic

The novel M statistic was evaluated against established indices (Abouheif's C mean, Moran's I, Blomberg's K, Pagel's λ, D statistic, δ statistic) using simulated data across different sample sizes [52].

• Simulation Protocol:
  • Data Generation: Simulated trait data (continuous and discrete) and phylogenies were generated under various evolutionary models, including Brownian motion and Markov models.
  • Signal Detection: The power and Type I error rate of each statistic were assessed by testing its ability to detect a known phylogenetic signal (or its absence) in the simulated data.
  • Trait Combination Testing: The M statistic was further tested on combinations of multiple traits, leveraging Gower's distance to calculate a composite dissimilarity matrix from mixed data types [52].
• Performance Analysis: The performance was compared based on the accuracy and statistical power of each index across the different simulation scenarios. The M statistic demonstrated robust performance across continuous variables, discrete variables, and multiple trait combinations, providing a unified method for phylogenetic signal detection [52].

Assessing Forecasting Accuracy for Protein Evolution

The ProteinEvolver2 framework, which integrates birth-death population models with structurally constrained substitution (SCS) models, was evaluated for its forecasting accuracy [15] [14].

• Simulation & Validation Protocol:
  • Forward Simulation: The method starts with a root protein sequence and structure. A forward-in-time birth-death process is simulated, where the fitness of a protein variant (calculated from its folding stability) determines its birth and death rates [15].
  • Integration with SCS Models: Protein evolution along the branches of the emerging phylogeny is simulated using SCS models, which incorporate selection on protein folding stability [15] [14].
  • Prediction Error Calculation: The framework was applied to monitored viral protein data. Forecasted protein variants were compared to subsequently observed empirical data. The error was quantified for both predicted folding stability (free energy, ΔG) and the corresponding amino acid sequences [15].
• Performance Outcome: The method showed acceptable errors in predicting the folding stability of future protein variants. As expected, the errors were larger for predicting the exact amino acid sequences, highlighting the challenge of precise sequence-level forecasting despite feasible stability prediction [15] [14].

Benchmarking Time-Varying Rate Estimation with the Polyepoch Clock

The polyepoch clock model, an inhomogeneous continuous-time Markov chain (ICTMC) that models evolutionary rate as a flexible, piecewise-constant function of time, was assessed via simulation [53].

• Simulation Protocol:
  • Scenario Design: Data were simulated under two distinct evolutionary scenarios with known, time-varying rate trajectories.
  • Inference: The polyepoch clock model was applied to the simulated data to infer the phylogeny and the rate-vs-time function.
  • Recovery Accuracy: The estimated rate trajectory and divergence times were compared against the known, simulated "ground truth" [53].
• Performance Outcome: The model demonstrated a strong ability to recover the true timescales and evolutionary rates across the tested scenarios. This validated its utility before application to empirical datasets from West Nile virus, Dengue virus, and influenza A/H3N2, where it successfully identified strong time-varying patterns [53].

Model Workflows and Logical Relationships

The following diagram illustrates the core workflow for simulating and evaluating evolutionary models, highlighting the key differences between traditional and integrated forecasting approaches.

Table 3: Essential Software and Statistical Tools for Evolutionary Model Simulation

Tool/Resource	Type	Function in Simulation Studies
phylosignalDB R Package	Software Package	Facilitates calculation of the M statistic for phylogenetic signal detection in continuous, discrete, and multiple traits. [52]
BEAST X	Software Platform	Enables Bayesian phylogenetic, phylogeographic, and phylodynamic inference under complex models, leveraging HMC for efficiency. [7]
ProteinEvolver2	Software Framework	Implements the integrated birth-death and SCS model for forecasting protein evolution. [15] [14]
Gower's Distance	Statistical Metric	Converts various types of traits (continuous, discrete) into a unified dissimilarity matrix, enabling the analysis of mixed data. [52]
Hamiltonian Monte Carlo (HMC)	Computational Algorithm	A Markov chain Monte Carlo method that uses gradients for efficient sampling of high-dimensional posteriors, implemented in BEAST X. [7] [53]
Effective Sample Size (ESS)	Performance Metric	Measures the efficiency of an MCMC sampler; higher ESS per unit time indicates better performance. [7]
Structurally Constrained Substitution (SCS) Models	Evolutionary Model	Substitution models that incorporate protein structure to inform evolutionary constraints, often leading to more accurate inferences. [15] [54]

This guide provides an objective comparison between Discrete Trait Analysis (DTA) and Structured Birth-Death (SBD) models to help researchers select the appropriate phylodynamic method for their work.

Core Principles and Methodological Comparison

Understanding the fundamental differences in how these models operate is the first step in selection.

Discrete Trait Analysis (DTA) operates as a neutral trait evolution model. It infers the history of a discrete trait, such as geographic location, along the branches of a pre-existing phylogenetic tree. Crucially, it is not a tree-generating process itself and does not model the population dynamics that shape the tree [12]. It typically uses Continuous-Time Markov Chain (CTMC) models to describe the rates of transition between discrete states [55] [7].

Structured Birth-Death (SBD) Models are tree-generating processes that jointly model the population dynamics and the phylogenetic tree. They describe how lineages multiply (birth), go extinct (death), and are sampled through time, with rates that can depend on the discrete type of an individual (e.g., location, host type) [56] [9]. These models directly infer the parameters that govern the epidemic process itself.

The table below summarizes their core methodological distinctions.

Feature	Discrete Trait Analysis (DTA)	Structured Birth-Death (SBD) Models
Core Principle	Trait evolution model on a fixed tree	Tree-generating population dynamic model
Treatment of Population Dynamics	Not explicitly modeled	Explicitly models birth, death, and sampling rates per type
Inference Target	History of trait changes & transition rates	Population parameters (e.g., transmission, migration rates) and type history
Computational Demand	Generally faster	More computationally intensive [56]

Performance and Limitations in Practice

The theoretical differences lead to distinct performance characteristics, especially regarding a critical issue in real-world data analysis: sampling bias.

• Sensitivity to Sampling Bias: A key limitation of DTA with CTMC models is their sensitivity to unbalanced sampling across discrete states [55] [7]. If sequences are disproportionately sampled from one location, the model can infer artificially high transition rates to that location. The Adjusted Bayes Factor (BFadj) has been developed to mitigate this by incorporating sample counts per location, reducing false-positive support for transitions (Type I errors) at the cost of increasing false negatives (Type II errors) [55].
• Robustness of Structured Models: Structured coalescent approaches (a category within SBD methods) are considered less subject to these sampling biases [12] [7]. However, they historically assumed constant population sizes, which can itself bias results. Newer methods like MASCOT-Skyline address this by jointly inferring time-varying population sizes and migration rates, improving accuracy [12].
• Root State Inference: When inferring the root state (e.g., the geographic origin of an outbreak), DTA models show variable performance. Accuracy tends to be highest at intermediate sequence dataset sizes and can be inflated with a larger number of discrete traits, making a popular metric, the Kullback-Leibler divergence, an unreliable performance indicator [6].

The following workflow diagram illustrates the core analytical processes of each model, highlighting where key biases can be introduced.

Guidelines for Model Selection

Your choice should be guided by your research question, data structure, and computational resources.

When to Choose Discrete Trait Analysis (DTA)

• Primary Goal: To perform ancestral state reconstruction and infer the history of transitions between discrete states on a phylogeny.
• Data Context: The phylogenetic tree is the primary object of interest, and you are adding a layer of trait history.
• Resource Constraints: When computational time is a significant limiting factor.
• Important Consideration: If using DTA, be cautious of unbalanced sampling and consider using bias-correction methods like the adjusted Bayes Factor (BFadj) [55].

When to Choose Structured Birth-Death Models

• Primary Goal: To infer epidemiologically relevant parameters such as type-specific effective population sizes, prevalence, migration/transmission rates, and effective reproduction numbers [12] [56].
• Data Context: The research question requires jointly understanding the population dynamics and the spatial/trait-based spread.
• Key Application: Inferring multi-type population trajectories, such as host-specific case numbers through time and the timing of spillover events, which provide information about the entire population, including unsampled individuals [56].
• Critical Scenarios: When sampling is known or suspected to be highly unbalanced across different states, and robustness to this bias is required [12] [55].

Experimental Protocols & Data Presentation

To ensure reproducible and high-quality inference, follow these established experimental protocols.

Protocol for a Structured Birth-Death Model Analysis

This protocol uses the BEAST2 software platform with the bdmm package, a standard for this type of analysis [9].

• Software Installation: Install BEAST2, BEAUti2, and the necessary packages (bdmm, MultiTypeTree). The bdmm package can be installed via BEAUti's package manager [9].
• Template Configuration: In BEAUti, load the appropriate template by selecting File > Template > MultiTypeBirthDeath [9].
• Data Input: Load the sequence alignment in FASTA format. Annotate the sequences by parsing sampling dates (e.g., "after last" underscore) and discrete traits/locations (e.g., "group 2" from underscore-delimited names) from the sequence headers using BEAUti's "Tip Dates" and "Tip Locations" panels [9].
• Model Specification: Define the substitution model (e.g., JC69, HKY), clock model (e.g., Strict or Relaxed Clock), and the SBD tree prior. Configure the priors for birth, death, and migration rates [9].
• MCMC Execution: Run the analysis in BEAST2 with a sufficient chain length to achieve convergence.
• Output Analysis: Use Tracer to assess parameter convergence (effective sample size > 200) and TreeAnnotator to generate a maximum clade credibility tree for visualization [9].

Quantitative Comparison of Model Biases

The following table summarizes findings from simulation studies that highlight the performance differences under specific conditions.

Condition	Discrete Trait Analysis (DTA)	Structured Birth-Death/Skyline Models
Unbalanced Sampling	Higher type I error for transitions to oversampled locations [55]. BFadj reduces type I but increases type II error [55].	More robust; less subject to sampling biases [12].
Non-Constant Population Sizes	Can lead to biased reconstruction of migration dynamics [12].	Methods like MASCOT-Skyline jointly infer time-varying population sizes and migration, reducing bias [12].
Root State Inference	Accuracy is highest at intermediate dataset sizes; common support metrics (KL) can be misleading [6].	Infers origin as part of the cohesive population dynamic model.

The Scientist's Toolkit: Essential Research Reagents

A successful phylodynamic analysis relies on a suite of software tools and computational resources.

Tool / Resource	Function	Relevance to Models
BEAST2 / BEAST X [7]	Core software platform for Bayesian evolutionary analysis.	Essential for both DTA and SBD. BEAST X introduces new, more scalable inference techniques [7].
BEAUti2 [9]	Graphical utility for generating BEAST2 configuration files (XML).	Used to set up analyses for both model types.
bdmm & MASCOT Packages [12] [9]	Implements structured birth-death and structured coalescent models.	Essential for running SBD analyses.
Tracer [9]	Diagnoses MCMC convergence and summarizes parameter estimates.	Critical for post-analysis diagnostics for both models.
TreeAnnotator [9]	Summarizes a set of posterior trees into a single consensus tree.	Used for final tree visualization for both models.
High-Performance Computing (HPC)	Provides necessary CPU power for complex calculations.	Particularly critical for the computationally intensive SBD models.

Conclusion

Discrete Trait Analysis and Structured Birth-Death Models are powerful, complementary tools in the phylodynamics toolkit. DTA excels at reconstructing ancestral states and visualizing trait history across phylogenies, while SBDM provides a more robust framework for directly quantifying population dynamic parameters like effective reproductive numbers (Re) and sampling rates. The choice between them depends on the research question, data quality, and computational resources. Future directions point towards model integration, improved handling of sampling bias, and leveraging advancements in Bayesian software like BEAST X for more scalable and accurate inference. For biomedical researchers, mastering both approaches is crucial for unraveling transmission patterns, assessing public health interventions, and preparing for future pathogen threats.

References

• 1. Phylogenetic and phylodynamic approaches to ... [https://www.nature.com/articles/s41576-022-00483-8]
• 2. A Bayesian Structured Coalescent Approximation - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC4534465/]
• 3. Discrete Trait Phylogenetic Association - EEOB Macroevolution [https://eeob-macroevolution.github.io/Practicals/Phylo_Assoc_Discrete/Phylo_Assoc_Discrete_Tutorial.html]
• 4. Phylogeography in BEAST 2 [https://www.beast2.org/2022/03/01/phylogeography.html]
• 5. Phylogenetic Tools for Comparative Biology: A few useful demos on... [https://blog.phytools.org/2024/04/a-few-useful-demos-on-ancestral-state.html]
• 6. Evaluating Bayesian phylogeographic models with discrete ... [https://www.sciencedirect.com/science/article/pii/S1567134820303324]
• 7. BEAST X for Bayesian phylogenetic, phylogeographic and ... [https://www.nature.com/articles/s41592-025-02751-x]
• 8. Robust Phylodynamic Analysis of Genetic Sequencing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9413058/]
• 9. Structured birth-death model [https://taming-the-beast.org/tutorials/Structured-birth-death-model/]
• 10. 10.2: The Birth-Death Model [https://bio.libretexts.org/Bookshelves/Evolutionary_Developmental_Biology/Phylogenetic_Comparative_Methods_(Harmon)/10%3A_Introduction_to_Birth-Death_Models/10.02%3A_The_Birth-Death_Model]
• 11. Chapter 10: Introduction to birth-death models [https://lukejharmon.github.io/pcm/chapter10_birthdeath/]
• 12. MASCOT-Skyline integrates population and migration ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12500135/]
• 13. Robust phylodynamic inference and model specification for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12344737/]
• 14. Forecasting protein evolution by integrating birth-death ... [https://elifesciences.org/articles/106365]
• 15. Forecasting protein evolution by integrating birth-death ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12459951/]
• 16. How the individual human mobility spatio-temporally ... [https://www.nature.com/articles/s41598-020-68230-9]
• 17. Mixing patterns and the spread of pandemics [https://www.nature.com/articles/s43588-022-00312-2]
• 18. Social mixing patterns relevant to infectious diseases spread ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9176177/]
• 19. Likelihood of single continuous trait in Beast2? [https://groups.google.com/g/beast-users/c/uz4cCrTyNi4]
• 20. Introduction to BEAST2 [https://taming-the-beast.org/tutorials/Introduction-to-BEAST2/]
• 21. Divergence Dating Tutorial with BEAST 2.2.x [https://beast2-dev.github.io/beast-docs/beast2/DivergenceDating/DivergenceDatingTutorial.html]
• 22. Robust regression rescues poor phylogenetic decisions [https://pmc.ncbi.nlm.nih.gov/articles/PMC12532477/]
• 23. Improved fixed tree and tree set support [https://www.beast2.org/2023/09/01/improvde-fixed-tree-and-treeset-support.html]
• 24. Fast and Accurate Maximum-Likelihood Estimation of Multi ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10924745/]
• 25. Phylodynamics with Migration [https://pmc.ncbi.nlm.nih.gov/articles/PMC4948704/]
• 26. Estimating pathogen spread using structured coalescent ... [https://www.sciencedirect.com/science/article/pii/S1755436524000562]
• 27. Tracing the transmission dynamics of HIV-1 CRF55_01B [https://www.nature.com/articles/s41598-020-61870-x]
• 28. Molecular cluster, transmission characteristics, origin and ... [https://www.sciencedirect.com/science/article/pii/S0001706X2400278X]
• 29. The emergence and transmission dynamics of HIV-1 ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8946679/]
• 30. Birth–death skyline plot reveals temporal changes of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3538216/]
• 31. Defining HIV-1 transmission clusters based on sequence data [https://pmc.ncbi.nlm.nih.gov/articles/PMC5482559/]
• 32. Epidemic theory (effective & basic reproduction numbers ... [https://www.healthknowledge.org.uk/public-health-textbook/research-methods/1a-epidemiology/epidemic-theory]
• 33. The effective reproduction number of pandemic influenza [https://pmc.ncbi.nlm.nih.gov/articles/PMC3084966/]
• 34. Estimates of the reproduction number for seasonal, pandemic ... [https://bmcinfectdis.biomedcentral.com/articles/10.1186/1471-2334-14-480]
• 35. BEAST 2.5: An advanced software platform for Bayesian ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6472827/]
• 36. BDMM-Prime - GitHub Pages [https://tgvaughan.github.io/BDMM-Prime/]
• 37. BDMM-Prime for Macroevolution [https://taming-the-beast.org/tutorials/BDMM-Prime-for-macroevolution/]
• 38. The EDM Association Global Data Management ... [https://edmcouncil.org/announcement/the-edm-association-global-data-management-benchmark-2025-survey-is-live/]
• 39. Data quality issues in data used in species distribution ... [https://www.sciencedirect.com/science/article/pii/S1574954125003875]
• 40. 2025 and Beyond: The Future of Genomic Data Analysis ... [https://blog.crownbio.com/2025-and-beyond-the-future-of-genomic-data-analysis-and-innovations-in-genomics-services]
• 41. Navigating the landscape of parameter identifiability methods [https://pmc.ncbi.nlm.nih.gov/articles/PMC11247124/]
• 42. Parameter identifiability analysis and visualization in large ... [https://bmcsystbiol.biomedcentral.com/articles/10.1186/s12918-017-0428-y]
• 43. Parameter identifiability, parameter estimation and model ... [https://arxiv.org/html/2405.08177v2]
• 44. Rate variation and estimation of divergence times using strict ... [https://bmcecolevol.biomedcentral.com/articles/10.1186/1471-2148-11-271]
• 45. Revisiting the phylogenetic relaxed clock model - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7880504/]
• 46. Genome-wide Association Studies for Discrete Traits - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC2920891/]
• 47. The advantages and limitations of trait analysis with GWAS [https://plantmethods.biomedcentral.com/articles/10.1186/1746-4811-9-29]
• 48. 10: Introduction to Birth-Death Models [https://bio.libretexts.org/Bookshelves/Evolutionary_Developmental_Biology/Phylogenetic_Comparative_Methods_(Harmon)/10%3A_Introduction_to_Birth-Death_Models]
• 49. GHIST 2024: The First Genomic History Inference ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12596078/]
• 50. Robust and accurate Bayesian inference of genome-wide ... [https://www.nature.com/articles/s41588-025-02317-9]
• 51. Inferring Geographic Spread of Flaviviruses Through Analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12567808/]
• 52. A Unified Method for Detecting Phylogenetic Signals in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11925719/]
• 53. Inhomogeneous continuous-time Markov chains to infer ... [https://arxiv.org/html/2510.11982v1]
• 54. Trends in substitution models of protein evolution for ... [https://www.sciencedirect.com/science/article/pii/S1055790325001903]
• 55. Assessing the Performance of an Adjusted Bayes Factor [https://pubmed.ncbi.nlm.nih.gov/41056469/]
• 56. Bayesian Phylodynamic Inference of Multitype Population ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12166523/]