Comparative Phylodynamics for Outbreak Source Attribution: Methodologies, Applications, and Future Directions

Abstract

This article provides a comprehensive comparison of phylodynamic methods for outbreak source attribution, tailored for researchers and public health professionals. It explores the foundational principles of phylodynamics, reviews and contrasts major methodological frameworks—from Bayesian phylogeography to scalable agent-based models—and addresses key challenges including computational scalability, sampling bias, and model misspecification. By synthesizing insights from recent SARS-CoV-2 studies and novel computational tools, it offers a validated guide for selecting and optimizing methods to accurately reconstruct transmission trees and identify outbreak origins, ultimately enhancing genomic surveillance and public health response.

Foundations of Phylodynamics: Core Principles for Source Attribution

Defining Phylodynamics and Its Role in Genomic Epidemiology

Phylodynamics is an interdisciplinary field that combines evolutionary biology with epidemiology to generate evidence about the spread and source of pathogens by exploiting the genomic signature left by ongoing evolution during transmission [1]. This approach allows researchers to corroborate findings from traditional epidemiological modeling and provides deeper insights where conventional case data over time and space may be insufficient [1]. The foundation of phylodynamics relies on "measurable evolution"—the phenomenon where pathogen molecular evolution occurs on the same timescale as transmission, making accumulated genetic diversity informative about the timing of transmission events [1].

During the SARS-CoV-2 pandemic, phylodynamics experienced more intense application than ever before, establishing it as a core component of coordinated outbreak responses [1] [2]. The field has made significant contributions to understanding the spread of various pathogens, including Ebola, Zika, and HIV, by capturing transmission dynamics in time and space that would otherwise remain inaccessible through traditional epidemiological analysis alone [1].

Core Concepts and Foundational Models

Theoretical Framework and Key Assumptions

Phylodynamic analysis requires pathogen genome sequences and their sampling times from infected hosts. Two key assumptions enable the inference of epidemiological parameters from genetic data:

• The hypothetical "true" phylogeny of the spreading pathogen mirrors the transmission network, with branching events closely corresponding to transmission events [1]
• The underlying pathogen population evolved according to a model that links epidemiological and phylogenetic dynamics [1]

In Bayesian phylogenetic frameworks, these models are implemented as "tree priors," providing an expression for the probability of a tree given parameters governing the epidemiological process that generated it [1]. The analysis requires phylogenetic trees with branch lengths corresponding to time units (chronograms), obtained by converting substitutions per site to time units using an evolutionary clock rate [1].

Foundational Models in Phylodynamics

Table 1: Foundational Phylodynamic Models and Their Characteristics

Model Type	Theoretical Basis	Key Parameters	Epidiological Applications
Coalescent	Population genetics - backwards-in-time process	Effective population size (Ne(t)), generation time (g)	Demographic history, population size through time [1]
Birth-Death	Epidemiological processes - forward-in-time process	Transmission rate (λ), removal rate (δ), sampling rate (ψ)	Reproductive number (R), growth rates, prevalence [1]
Multi-Type Birth-Death	Structured population expansion	Type-specific birth rates (λij,k), migration rates (mij,k), death rates (μi,k)	Migration patterns, spatial spread, between-population dynamics [3]

The coalescent model originated in population genetics and models how the ancestry of sampled populations relates to their demographic history [1]. Visualized as a genealogy of sampled individuals, internal nodes correspond to times when lineages coalesce into common ancestors, with time starting at the most recent sample and terminating at the most recent common ancestor [1].

The birth-death model takes a forward-in-time approach, modeling transmission (birth) and removal (death) events in an infected population [1]. This model more directly represents epidemiological processes, with parameters that can be directly linked to transmission rates, sampling rates, and recovery rates [1].

For structured populations, the multi-type birth-death model extends the basic framework to account for dynamics across different populations, geographic regions, or pathogen subtypes [3]. This model can quantify migration rates and type-specific parameters essential for understanding spatial spread and between-population dynamics [3].

Phylodynamics in Outbreak Source Attribution

Molecular Source Attribution Approaches

Source attribution refers to methods that reconstruct infectious disease transmission from a specific source, which could be a population, individual, or location [4]. Molecular source attribution uses pathogen molecular characteristics—most often genomic sequences—to reconstruct transmission events [4]. This approach has become increasingly powerful with advances in sequencing technology and computational methods.

Two primary approaches are used in molecular source attribution:

• Microbial subtyping: Infections are categorized into subtypes based on molecular varieties, with source attribution inferred from subtype similarity [4]
• Phylogenetic reconstruction: Genetic sequences are compared to reconstruct phylogenetic trees that approximate transmission history [4]

The resolution of source attribution depends on having sufficient genetic diversity to differentiate transmission pathways without defining so many subtypes that each individual appears unique [4]. Whole genome sequencing has significantly enhanced attribution precision, particularly for bacterial pathogens, by providing maximal discriminatory power [4].

Methodological Frameworks for Source Attribution

Table 2: Methodological Approaches for Geographical Source Inference

Method Class	Specific Methods	Key Features	Computational Considerations
Ancestral State Reconstruction	Discrete Trait Analysis (DTA)	Incorporates discrete metadata (e.g., travel history); relatively low computational demand [2] [5]	Less robust to uneven sampling; parameters difficult to interpret epidemiologically [2]
Structured Population Models	Structured Coalescent; Multi-Type Birth-Death	Accounts for variable sampling between regions; infers epidemiologically interpretable parameters [2] [5]	Computationally intensive; improved scalability with recent algorithmic advances [3]
Phylogeographic Models	Asymmetric migration models; BEAST phylogeography	Reconstructs spatial dispersal from phylogenetic tree topology [2] [5]	Limited scalability for large datasets (>600 sequences) without model optimizations [6]

Discrete Trait Analysis (DTA) assigns location states to nodes on a phylogeny and can incorporate travel history data in a straightforward manner [2]. However, it doesn't fully accommodate the interdependency of tree shape and migration rates and is sensitive to sampling biases [2].

Structured population models explicitly model migration events and rates at a population level, providing parameters that can be directly compared with epidemiological or mobility data [2]. These models are more robust to variable sampling between regions but are computationally intensive, though recent algorithmic improvements have enhanced their scalability [3].

Recent advances in multi-type birth-death models have addressed previous limitations in numerical stability and computational efficiency, enabling analysis of datasets containing several hundred genetic samples [3]. These improvements are particularly important for structured populations, where quantifying parameters for each subpopulation requires sufficient samples from each group [3].

Comparative Performance of Phylodynamic Methods

Experimental Evidence on Model Performance

Recent studies have systematically evaluated the performance of different phylodynamic methods under various conditions:

• Model misspecification robustness: Simple structured coalescent models can recover migration rates while adjusting for nonlinear epidemiological dynamics, with only small biases observed for sample sizes ≥1000 sequences [6]
• Genetic data requirements: Migration rates can be estimated using alignments equivalent to partial genes (e.g., HIV pol gene) or complete pathogen genomes, with higher migration rates estimated more accurately than lower rates [6]
• Computational scalability: Phylogeographic models in BEAST showed limited scalability for datasets of 600 or more sequences, though recent improvements in birth-death model implementation have dramatically increased analyzable sample sizes [6] [3]

A study evaluating HIV transmission dynamics found that even simplistic representations of complex epidemiological models could still estimate migration rates accurately, depending on the method and sample size used [6]. The research demonstrated that estimation of higher migration rates was more accurate than estimation of lower migration rates, highlighting method-specific sensitivities [6].

Advanced Methodological Innovations

PhyloTune, a recently developed method, accelerates phylogenetic updates using pretrained DNA language models [7]. This approach identifies the taxonomic unit of newly collected sequences and updates corresponding subtrees, significantly reducing computational time compared to complete tree reconstruction [7]. Experimental results demonstrated that:

• For smaller datasets (n=20-40 sequences), updated trees exhibited identical topologies to complete trees
• For larger datasets (n=60-100 sequences), minor topological discrepancies emerged (RF distances: 0.007-0.054)
• Computational time was relatively insensitive to total sequence numbers compared to exponential growth with complete tree reconstruction [7]

Multi-scale phylodynamic agent-based models represent another innovation, integrating within-host pathogen evolution with between-host transmission dynamics in heterogeneous populations [8]. These models can simulate feedback loops between public health interventions and pathogen evolution, capturing phenomena like the punctuated evolution observed in SARS-CoV-2 [8].

Essential Research Reagents and Tools

Table 3: Essential Research Reagent Solutions for Phylodynamic Analysis

Tool Category	Specific Tools	Primary Function	Application Context
Phylogenetic Software	BEAST2 (with bdmm package); RAxML-NG; PhyML	Bayesian phylogenetic inference; maximum likelihood tree estimation	Comprehensive phylodynamic analysis; tree topology estimation [1] [3] [9]
Sequence Alignment	MAFFT; BuddySuite	Multiple sequence alignment; genomic data processing	Preprocessing of genomic data for phylogenetic analysis [7]
Classification & Annotation	DNABERT; Kraken2; BLAST	Taxonomic classification; sequence similarity identification	Taxonomic unit identification; sequence annotation [7]
Language Models	PhyloTune	High-dimensional sequence representation; attention region identification	Efficient phylogenetic updates; informative region extraction [7]

The BEAST2 software platform with packages like bdmm (birth-death model migration) provides a comprehensive framework for phylodynamic analysis, enabling joint inference of tree topologies, phylodynamic parameters, molecular clock rates, and substitution models [3]. Recent algorithmic improvements to bdmm have dramatically increased the number of genetic samples that can be analyzed while improving numerical robustness and computational efficiency [3].

DNA language models like DNABERT generate high-dimensional sequence representations that can be used for taxonomic classification and identification of phylogenetically informative regions [7]. These models leverage the transformer architecture with self-attention mechanisms to capture long-range dependencies in genomic sequences, similar to how language models process natural language [7].

Experimental Protocols and Workflows

Standard Phylodynamic Analysis Pipeline

The following workflow diagram illustrates a standard protocol for phylodynamic analysis of outbreak genomic data:

Standard Phylodynamic Analysis Workflow

This workflow was applied in an early SARS-CoV-2 analysis that used 86 genomes to estimate the TMRCA (Most Recent Common Ancestor) and growth rate parameters [9]. Key steps included:

• Sequence quality control: Removal of sequences with sequencing artefacts, insufficient information, or resequencing of the same sample [9]
• Dataset sub-setting: Inclusion of only a single representative genome from known epidemiologically-linked transmission clusters to meet coalescent model assumptions [9]
• Model selection: Comparison of constant size and exponential growth coalescent models, with exponential growth providing better fit to emerging pandemic data [9]
• Parameter estimation: Bayesian MCMC analysis to estimate evolutionary rates, TMRCA, and growth parameters with 95% credible intervals [9]

Source Attribution Experimental Protocol

For outbreak source attribution studies, the following specialized protocol is recommended:

• Genomic data collection: Collect whole genome sequences from outbreak isolates with comprehensive metadata including sampling dates and locations [4] [5]
• Genetic clustering: Apply clustering methods to group similar sequences, balancing resolution power with practical utility [4]
• Phylogeographic analysis: Implement structured birth-death models or discrete trait analysis to infer spatial transmission history [2] [5]
• Statistical validation: Assess confidence in source attribution through bootstrap support or posterior probability values [5]

A key consideration in source attribution is accounting for sampling bias, as uneven sampling across regions can strongly influence phylogeographic inferences [5]. Structured population models generally show better robustness to sampling heterogeneity compared to discrete trait analysis [2].

Phylodynamics has established itself as an essential component of genomic epidemiology, providing powerful methods for reconstructing transmission dynamics and identifying outbreak sources. The comparative analysis presented here demonstrates that while foundational models like the coalescent and birth-death processes provide the theoretical framework for phylodynamic inference, structured models offer enhanced capabilities for source attribution applications.

Method selection should be guided by specific research questions, data characteristics, and computational constraints. For rapid assessment of well-sampled outbreaks, discrete trait approaches provide efficient inference, while for complex transmission dynamics with uneven sampling, structured birth-death models offer more robust parameter estimation. Recent advances in algorithmic efficiency and multi-scale modeling continue to expand the boundaries of phylodynamic inference, promising even more powerful tools for future outbreak responses.

The integration of phylodynamic methods into public health practice represents a paradigm shift in outbreak epidemiology, enabling researchers to extract profound insights into pathogen spread from genetic sequences. As these methods continue to evolve and improve, they will undoubtedly play an increasingly central role in global infectious disease surveillance and control efforts.

Phylodynamics represents a powerful, integrative framework that combines phylogenetics, epidemiology, and population dynamics to uncover the transmission dynamics of infectious pathogens [10]. The core premise of phylodynamics is that epidemiological processes, such as transmission and population fluctuations, occur on timescales similar to the accumulation of evolutionary changes in pathogen genomes. This synergy leaves a distinct signature in the genetic data, allowing researchers to reconstruct key aspects of an outbreak's history from molecular sequences [10] [11]. Originally applied to rapidly evolving viruses, these methods are now instrumental for outbreak surveillance, enabling estimation of critical parameters like the effective reproduction number (Re), divergence times, and spatial spread patterns from sampled pathogen sequences [12] [13].

The field faces a fundamental challenge: extracting robust, biologically plausible inferences from complex and often limited genetic data [14] [13]. This article provides a comparative guide to modern phylodynamic methods, evaluating their performance, underlying models, and applicability for outbreak source attribution research.

Comparative Analysis of Phylodynamic Methods

Table 1: Comparison of Core Phylodynamic Methodologies

Method Category	Key Software/Approach	Underlying Model	Primary Applications	Key Strengths	Key Limitations
Coalescent-Based	BEAST (Bayesian Skyline Plot) [13] [15]	Coalescent	Estimating effective population size (Ne(t)) trajectory, demographic history [14] [10].	Models genetic diversity; conditions on known sampling times; well-established framework [14].	Sampling times are fixed inputs; indirect link to epidemiological parameters [14].
Birth-Death Based	BEAST2 (Birth-Death Model) [11] [15]	Birth-Death	Inferring transmission rates (λ), recovery rates (δ), reproductive number (R0), origin time (T) [14].	Directly models transmission and sampling as stochastic processes; jointly infers tree and sampling times [14].	Prior specification highly influences results with limited data; computationally intensive [14] [11].
Deep Learning / Simulation-Based	PhyloDeep [11]	Birth-Death variants (BD, BDEI, BDSS)	Fast parameter estimation and model selection from large phylogenies [11].	Extremely fast on large trees; avoids complex likelihood calculations; good accuracy [11].	Requires extensive training with simulated data; "black box" inference process [11].
Joint Inference Frameworks	EpiFusion [12]	Particle Filtering	Joint inference using both phylogenetic trees and case incidence data [12].	Combines strengths of different data types (genetic and epidemiological) for robustness [12].	Increased model complexity; requires multiple data streams [12].

Table 2: Performance Comparison on Simulated and Real Data

Method / Software	Computational Speed	Scalability to Large Trees (>1000 tips)	Accuracy on Simulated Data (vs. Known Truth)	Robustness to Prior Specification	Real-World Application Example
BEAST2 (Coalescent)	Slow [11]	Limited [11]	High when temporal signal is strong and priors are appropriate [14]	Low to Moderate (posteriors can be highly prior-dependent with limited data) [14] [13]	HIV epidemic dynamics in the UK [10]
BEAST2 (Birth-Death)	Slow [11]	Limited [11]	Can be biased if model misspecified or with reporting delays [14] [15]	Low (high sensitivity to tree prior choices in early outbreaks) [14]	Zika virus epidemic in the Americas [14]
PhyloDeep (FFNN-SS)	Fast (seconds to minutes) [11]	High [11]	Better than BEAST2 on complex models (BDEI, BDSS) [11]	High (trained on wide parameter ranges, less dependent on user priors) [11]	HIV superspreading dynamics in Zurich [11]
PhyloDeep (CBLV-CNN)	Fast (seconds to minutes) [11]	High [11]	State-of-the-art on tested models; outperforms BEAST2 and FFNN-SS [11]	High (same as FFNN-SS) [11]	HIV superspreading dynamics in Zurich [11]

Experimental Protocols in Phylodynamic Analysis

A robust phylodynamic analysis requires a carefully constructed pipeline to ensure reliable and biologically plausible inferences. The following workflow, adapted from foundational sources, outlines the critical steps and decision points [13] [11].

Detailed Methodological Considerations

• Sequence Preparation and Curation: The initial phase involves rigorous sequence collection, alignment, and curation. For reliable inference, the dataset must be temporally and spatially representative of the outbreak. Researchers must decide whether to analyze all available sequences or focus on specific monophyletic lineages, a choice that can significantly impact results [13]. Tools like the Recombination Detection Program (RDP) are often used to identify and remove recombinant sequences that violate phylogenetic assumptions [13].

• Evolutionary Model Selection: This critical step assesses the temporal signal in the data using tools like TempEst to determine if sampling dates can calibrate the molecular clock. Subsequently, statistical comparison (e.g., using Bayesian Information Criterion - BIC) selects the best-fitting nucleotide substitution model (e.g., HKY or GTR) [13]. A strict vs. relaxed molecular clock model is also chosen based on data characteristics [13].

• Tree Prior Selection and Robustness Testing: The choice between Coalescent and Birth-Death tree priors is fundamental. As demonstrated in Zika virus studies, estimates of the reproductive number and tree height can be highly sensitive to this choice, especially with limited data [14]. A robustness check, scanning different models and prior distributions, is mandatory. Only estimates robust to reasonable prior changes should be trusted for policy decisions [14] [13].

• Accounting for Real-World Biases: Modern extensions address common surveillance biases. For instance, reporting delays between sample collection and sequence deposition can severely bias real-time estimates of effective population size. New models incorporate reporting delay distributions to mitigate this effect, providing more reliable estimates closer to the present time [15]. Furthermore, preferential sampling models account for situations where sampling intensity is correlated with disease prevalence [15].

The Scientist's Toolkit: Essential Research Reagents & Software

Table 3: Essential Tools for Phylodynamic Research

Tool Name	Category	Primary Function	Key Features
BEAST / BEAST2 [13] [10] [15]	Software Package	Bayesian evolutionary analysis by sampling trees.	Implements coalescent and birth-death models; integrates sequence evolution with demographic/epidemiological models; gold standard for many applications.
PhyloDeep [11]	Software Package	Fast parameter estimation and model selection using deep learning.	Uses neural networks on tree summaries (SS) or compact vector representations (CBLV); handles large trees efficiently.
EpiFusion [12]	Analysis Framework	Joint inference from phylogenetic and case incidence data.	Uses particle filtering; implemented in R and Java; improves estimation of the effective reproduction number.
Birth-Death Prior	Model	Tree prior for phylodynamic inference.	Models transmission (λ), becoming non-infectious (δ), sampling (ρ), and origin time; infers trees and sampling times jointly [14].
Coalescent Prior	Model	Tree prior for phylodynamic inference.	Models effective population size (Ne); conditions on fixed sampling times; infers population size changes from genetic data [14].
Bayesian Skyline Plot [15]	Model	Non-parametric estimation of population size.	Infers changes in effective population size (Ne(t)) over time; implemented in BEAST.
Compact Bijective Ladderized Vector (CBLV) [11]	Data Representation	A bijective, compact vector representation of a phylogenetic tree.	Preserves all tree information (topology & branch lengths); enables use of convolutional neural networks (CNN) for analysis.
Theophylline-d3	Theophylline-d3, MF:C7H8N4O2, MW:183.18 g/mol	Chemical Reagent	Bench Chemicals
Multi-target kinase inhibitor 2	Multi-target kinase inhibitor 2, MF:C20H14Cl2N6O, MW:425.3 g/mol	Chemical Reagent	Bench Chemicals

Conceptual Framework of Phylodynamic Models

The statistical foundation of Bayesian phylodynamics involves inferring the joint posterior distribution of the phylogenetic tree and model parameters given the sequence data and other relevant information [14]. This can be represented as: P(Tree, Parameters | Sequence Data, Other Data) ∝ P(Sequence Data | Tree, Parameters) × P(Tree, Other Data | Parameters) × P(Parameters) Here, the phylogenetic likelihood P(Sequence Data | Tree, Parameters) is determined by the evolutionary substitution model, while the phylodynamic likelihood P(Tree, Other Data | Parameters) is specified by the population dynamic model (e.g., Birth-Death or Coalescent) [14].

The diagram illustrates the core phylodynamic inference loop. The true, unobserved transmission process in the host population drives the evolution of the pathogen. Sampled sequences are used to infer a phylogenetic tree, which serves as the input for statistical models (e.g., Birth-Death or Coalescent). These models reverse-engineer the process to estimate the underlying transmission dynamics, such as the effective population size N_e(t) or the time-varying reproductive number R(t).

The Importance of Source Attribution in Public Health Response

Source attribution is a critical discipline in epidemiology that reconstructs the transmission of infectious diseases from specific sources—such as animal reservoirs, food products, or infected individuals—to humans [16] [17]. By quantifying the contributions of different sources to the human disease burden, it enables public health officials to prioritize interventions, measure their impact, and allocate resources efficiently [17] [18]. This guide compares the primary methodological approaches for source attribution, focusing on the growing role of phylodynamic methods which integrate phylogenetic analysis of pathogen genomes with models of disease dynamics.

Methodological Comparison of Source Attribution Approaches

Multiple methodologies exist for attributing the source of infections, each with distinct data requirements, applications, and strengths [17] [18]. The choice of method depends on the research question, the point in the farm-to-fork continuum one wishes to attribute, and, crucially, the availability and quality of data [18].

Table 1: Comparison of Major Source Attribution Methodologies

Methodology	Core Principle	Point of Attribution	Key Strengths	Key Limitations
Microbial Subtyping (e.g., Frequency-Matching Models) [19] [17]	Compares the distribution of pathogen subtypes (e.g., serotypes, sequence types) in human cases with their distribution in potential animal or food sources.	Primarily the point of production (animal reservoir) [19].	Well-established with a strong track record for pathogens like Salmonella; provides quantitative estimates of source contributions [17] [18].	Requires representative, strain-typed isolates from all major sources; subtypes must be stable across the farm-to-fork continuum [18].
Population Genetics Models (e.g., STRUCTURE) [18]	Uses genetic data to assess the genealogical history and evolutionary relationships among strains, assigning human cases to the genetically closest source.	Point of production (animal reservoir).	Can attribute cases even when perfect subtype matches are absent; accounts for pathogen evolution [18].	Requires high-resolution genetic data; the panel of potential sources must be complete to avoid misattribution [18].
Analysis of Outbreak Data [17]	Attributes cases based on the investigation of foodborne outbreaks where the source is identified.	Point of exposure (specific food vehicle).	Directly uses public health investigation data; no need for complex modeling.	Limited to outbreaks; results may not be representative of sporadic cases, which constitute the majority of illnesses [17].
Case-Control Studies of Sporadic Cases [18]	Compares the exposures of infected individuals (cases) with those of uninfected controls to identify risk factors.	Point of exposure (specific food, contact, etc.).	Identifies risk factors and specific exposure routes for sporadic cases.	Susceptible to recall and selection biases; cannot attribute cases to specific animal reservoirs directly [18].
Phylodynamic Methods [20] [21]	Reconstructs transmission trees and estimates epidemiological parameters by combining pathogen genome sequences with epidemiological and disease dynamic models.	Can infer transmission between individuals, populations, or locations.	Provides a unified framework for evolutionary and epidemiological inference; can identify direct transmission links and estimate key parameters like the reproductive number (R) [14] [21].	Computationally intensive; requires sequence data and can be sensitive to model specification and prior choices [14].
Quantitative Risk Assessment (QRA) [18]	A "bottom-up" approach that models the transmission pathway from source to human, incorporating data on contamination levels, food consumption, and dose-response.	Any point in the food chain (production, processing, consumption).	Can model the impact of interventions at different stages of the food production chain.	Data-intensive; requires detailed information on the entire farm-to-fork continuum [18].

Experimental Data: Phylodynamic Inference in Practice

Phylodynamic models are not just theoretical constructs; they are routinely applied to real-world outbreak data to infer transmission patterns. The following table summarizes results and protocols from key studies that employed phylodynamic methods for source attribution.

Table 2: Experimental Data from Phylodynamic Source Attribution Studies

Pathogen / Context	Core Objective	Method & Model Used	Key Findings & Quantitative Results
Mycobacterium tuberculosis in the Netherlands [20]	To determine Single Nucleotide Polymorphism (SNP) cut-offs for identifying probable transmission clusters using a phylodynamic model as a reference instead of contact tracing.	Model: phybreakData: 2,008 whole-genome sequences from TB patients (2015-2019).Protocol: Genetic clusters were first defined (≤20 SNP distance). phybreak was then run on each cluster to infer transmission events, which were used to assess the performance of various SNP cut-offs.	A SNP cut-off of 4 captured 98% of model-inferred transmission events. A cut-off beyond 12 SNPs effectively excluded transmission. The study demonstrated that phylodynamics provides a valuable alternative to often unreliable contact tracing for defining genetic thresholds [20].
Porcine Reproductive and Respiratory Syndrome Virus (PRRSV) in U.S. swine systems [21]	To infer the spread and population history of a specific PRRSV strain (RFLP 1-7-4) among five production systems.	Model: Coalescent and discrete-trait phylodynamic models in a Bayesian statistical framework.Data: 288 ORF5 gene sequences with metadata on farm system and type.Protocol: The best-fit nucleotide substitution model was selected. Models were used to infer demographic history and the ancestral system with root state posterior probability, and significant dispersal routes were identified using Bayes Factors (>6).	Identified the most likely ancestral production system (root state posterior probability = 0.95). Revealed that sow farms were central to viral spread within the systems. Showed that currently circulating viruses are evolving rapidly and have higher relative genetic diversity than earlier relatives [21].
Zika Virus epidemic in the Americas [14]	To assess how model choices (tree priors) influence the estimation of key parameters like the reproductive number (R) and tree height during an emerging epidemic.	Model: Comparison of Birth-Death and Coalescent tree priors in BEAST 2.Data: Zika virus genome sequences from Brazil and Florida, USA.Protocol: Analyses were run with different tree priors and prior distributions on parameters to test the robustness of estimates.	Parameter estimates were not robust for smaller, local epidemics (Brazil and Florida), highlighting that data may be uninformative early in an outbreak. Emphasizes the critical need for robustness checks by scanning models and priors; estimates can only be trusted if the posterior is robust to reasonable prior changes [14].

Detailed Experimental Protocol: SNP Cut-off Assessment forM. tuberculosis

The following workflow details the methodology from the tuberculosis study cited in Table 2 [20]:

• Data Preparation & Cluster Formation:

  • Sequencing & Alignment: Whole-genome sequences are obtained from bacterial isolates and aligned to a reference genome (e.g., H37Rv for TB).
  • SNP Calling & Filtering: A genotypes matrix is created from variant calls, filtering out sites in mobile genetic elements and those with low quality or high levels of missing data.
  • Transitive Clustering: Using a package like adegenet in R, sequences are clustered into genetic groups where each sequence is within a defined SNP distance (e.g., 20 SNPs) of at least one other sequence in the cluster. This large initial cut-off ensures all potentially linked cases are grouped.

• Phylodynamic Inference:

  • Model Application: For each genetic cluster of sufficient size, a phylodynamic model (e.g., phybreak) is run. This model uses the sequences and their collection dates to infer a posterior distribution of possible transmission trees.
  • Parameter Estimation: The model co-infers transmission events, the mutation rate of the pathogen, and other epidemiological parameters.

• Validation & Cut-off Assessment:

  • Reference Definition: The transmission events inferred by phybreak are used as the "reference standard" for determining which case pairs constitute a transmission link.
  • Performance Calculation: For a range of SNP cut-offs (e.g., 0 to 15), the proportion of model-inferred transmission pairs that fall at or below that cut-off is calculated. This identifies the SNP threshold that best captures true transmission events while minimizing false positives.

Workflow for Phylodynamic SNP Cut-off Assessment

The Scientist's Toolkit: Essential Research Reagents and Solutions

Successfully implementing phylodynamic and source attribution studies requires a suite of specialized tools, software, and data.

Table 3: Key Research Reagent Solutions for Phylodynamic Source Attribution

Tool / Resource	Type	Primary Function	Application Example
BEAST 2 [14]	Software Package	A cross-platform Bayesian evolutionary analysis software for inferring evolutionary history and population dynamics from genetic data.	Used to co-infer phylogenetic trees and epidemiological parameters using coalescent or birth-death tree priors [14].
EpiFusion [22]	Software Framework	A Java-based model for joint inference of outbreak characteristics using both phylogenetic trees and case incidence data via particle filtering.	Infers infection trajectories and the effective reproduction number (R~t~) by combining case data and a phylogenetic tree posterior [22].
phybreak [20]	R Package/Model	A phylodynamic method to infer transmission events from outbreak data (genomes and sampling times) without imputing many unobserved cases.	Used to infer transmission chains of M. tuberculosis in a low-incidence setting to validate SNP cut-offs [20].
Structured Coalescent Model [6]	Mathematical Model	A phylodynamic model that estimates migration rates between populations (e.g., geographic regions or host groups) while adjusting for epidemiological dynamics.	Applied to HIV sequence data to estimate migration rates between populations, showing scalability for large datasets (≥1000 sequences) [6].
Whole Genome Sequencing (WGS) [20]	Laboratory & Data	Provides the highest resolution data by sequencing the entire pathogen genome, enabling precise strain discrimination and detailed phylogenetic analysis.	The foundation for calling SNPs and building high-resolution phylogenies for M. tuberculosis transmission studies [20].
Reporting Delay Model [15]	Statistical Model	A method that incorporates the distribution of times between sample collection and sequence reporting to correct biases in real-time phylodynamic analyses.	Improves the accuracy of effective population size estimates for SARS-CoV-2 near the present time by accounting for missing data [15].
Dasatinib analog-1	Dasatinib analog-1, MF:C22H25ClFN7O2S, MW:506.0 g/mol	Chemical Reagent	Bench Chemicals
Bcl-2-IN-14	Bcl-2-IN-14, MF:C37H31N5O5S, MW:657.7 g/mol	Chemical Reagent	Bench Chemicals

Critical Considerations and Future Directions

While powerful, phylodynamic methods require careful implementation. A major consideration is model specification and robustness [14]. Choices regarding the tree prior (e.g., coalescent vs. birth-death) and parameter priors can significantly influence results, especially with limited or early-outbreak data. Researchers must perform robustness checks to ensure estimates are reliable [14]. Furthermore, model misspecification can introduce inductive bias, though for large sample sizes (e.g., ≥1000 sequences), this bias may be small [6].

The future of source attribution lies in data integration. Frameworks like EpiFusion, which jointly model phylogenetic trees and case incidence data, represent a move toward synthesizing all available data streams for a more complete and reliable picture of outbreak dynamics [22]. As whole-genome sequencing becomes standard, methods that leverage its full potential while accounting for real-world complexities like reporting delays will be indispensable for precise and timely public health response [15] [18].

The field of infectious disease dynamics has been transformed by the integration of two powerful data streams: classical epidemiological information and pathogen genomic sequences. This integration, formalized through phylodynamic methods, enables researchers to infer transmission patterns, identify outbreak sources, and reconstruct the evolutionary history of pathogens. Phylodynamics combines evolutionary models from molecular phylogenetics with epidemiological models from population dynamics to create a unified framework for analyzing infectious disease spread [23]. This approach has been applied to diverse pathogens including HIV, influenza, Mycobacterium tuberculosis, and SARS-CoV-2, providing crucial insights for public health interventions [6] [20] [23].

The core premise of this framework is that pathogen genomes accumulate mutations over time, and the relationships between these genetic sequences contain valuable information about the timing and spread of infections. When combined with epidemiological data such as symptom onset dates, contact networks, and geographic locations, these molecular sequences enable powerful inferences about transmission dynamics that neither data type could provide alone. This article compares the leading phylodynamic methods and provides a conceptual framework for their application in outbreak source attribution research.

Comparative Analysis of Phylodynamic Methods

Methodological Approaches and Applications

Table 1: Comparison of Phylodynamic Methods and Applications

Method	Primary Application	Data Requirements	Key Outputs	Strengths	Limitations
Structured Coalescent Models	Estimating migration rates between populations [6]	Genetic sequences, population structure	Migration rates, effective population sizes	Can adjust for nonlinear epidemiological dynamics [6]	Potential inductive bias with model misspecification [6]
Agent-Based Models (PhASE TraCE)	Multi-scale pandemic modeling with rapid variant emergence [8]	Genomic surveillance, demographic, mobility data	Transmission chains, variant emergence patterns, intervention impacts	Captures feedback between evolution, interventions, and behavior [8]	Computational intensity with large populations [8]
Bayesian Birth-Death Models	Cluster-based transmission rate estimation [24]	Time-stamped sequences, epidemiological priors	Transmission rates, reproductive numbers, cluster influence	Quantifies uncertainty in parameter estimates [24]	Influence varies with cluster size and rate heterogeneity [24]
Phylogenetic Network Methods	Lateral spread inference in outbreaks [25]	Whole genomes, epidemiological contact data	Genetic networks, transmission links, diffusion routes	Integrates multiple transmission drivers simultaneously [25]	Dependent on quality of epidemiological metadata [25]
Transmission Tree Inference (phybreak)	SNP cut-off determination for transmission clusters [20]	WGS data, serial interval distributions	Transmission probabilities, SNP thresholds	Provides biological reference without contact tracing [20]	Assumes uniform generation time distributions [20]

Performance and Operational Characteristics

Table 2: Performance Metrics of Computational Tools

Tool/Method	Computational Efficiency	Scalability	Statistical Power	Implementation Requirements
HyPhy	30 minutes for 1,776 sequences [26]	High	61.4% sequences clustered [26]	Patristic distance ≤2% threshold [26]
MEGA	324 hours for 1,776 sequences [26]	Moderate	33.7% sequences clustered [26]	Patristic distance ≤1.5% threshold [26]
BEAST Phylogeography	Not scalable for ≥600 sequences [6]	Low	Accurate migration rates with simple models [6]	Complex model specification [6]
Exponential Random Graph Models (ERGM)	Moderate	High for network inference	Identifies significant transmission drivers [25]	Genetic networks, epidemiological covariates [25]
Agent-Based Models	Variable with population size	Scalable with computational resources	Replicates complex multi-scale dynamics [8]	High-resolution genomic and mobility data [8]

Experimental Protocols for Phylodynamic Analysis

Protocol 1: Molecular Transmission Cluster Analysis

Objective: Identify recent transmission clusters using genetic sequence data to guide public health interventions.

Methodology:

• Sequence Preparation: Align viral sequences (e.g., HIV pol gene) using MAFFT v7 or equivalent tool [25] [26].
• Phylogenetic Reconstruction: Build maximum likelihood trees using IQ-TREE v1.6.6 with 1000 bootstrap replicates [25].
• Genetic Distance Calculation: Compute patristic distances (evolutionary distance along tree branches) between all sequence pairs [26].
• Cluster Identification: Apply distance thresholds (≤1.5% for MEGA, ≤2% for HyPhy) to define transmission clusters [26].
• Validation: Compare cluster composition with epidemiological data to validate transmission links.

Key Parameters:

• Genetic distance thresholds optimized for specific pathogens and epidemiological contexts
• Bootstrap support >70% for cluster stability
• Temporal signal assessment through root-to-tip regression

Protocol 2: Integrated Genomic-Epidemiological Network Analysis

Objective: Identify factors driving viral spread during outbreaks using combined genomic and epidemiological data.

Methodology:

• Genetic Network Construction: Generate median-joining networks from concatenated gene segments using NETWORK 10.2.0.0 [25].
• Epidemiological Covariate Preparation:
  • Calculate geographic distances between cases
  • Determine risk window overlaps (periods of potential infectiousness and susceptibility)
  • Record production system relationships (same owners, poultry companies) [25]
• Model Implementation: Apply Exponential Random Graph Models (ERGM) to assess the effect of covariates on genetic link probability [25].
• Interpretation: Identify significant drivers (e.g., same poultry company Est. = 0.548, risk windows overlap Est. = 0.339) [25].

Key Parameters:

• Genetic difference threshold for link definition
• Confidence intervals for covariate effect estimates
• Model goodness-of-fit assessment

Protocol 3: Transmission SNP Threshold Determination

Objective: Establish evidence-based SNP cut-offs for defining transmission clusters using phylodynamic inference.

Methodology:

• Genetic Clustering: Perform transitive clustering with initial 20-SNP threshold to define candidate transmission clusters [20].
• Transmission Inference: Apply phybreak method to infer transmission events within clusters using WGS data and serial interval distributions [20].
• SNP Distance Analysis: Calculate proportion of inferred transmission events below various SNP cut-offs (e.g., 3-12 SNPs) [20].
• Threshold Selection: Identify optimal SNP cut-off that captures majority of transmission events (e.g., 4 SNPs captured 98% of transmissions in TB study) while minimizing false positives [20].

Key Parameters:

• Mutation rate calibration for specific pathogens
• Serial interval distribution parameters
• Sampling density considerations

Conceptual Framework and Workflows

Integrated Phylodynamic Analysis Framework

Method Selection Workflow

Table 3: Essential Research Reagents and Computational Tools

Category	Specific Tool/Resource	Primary Function	Application Context
Sequence Alignment	MAFFT v7 [25]	Multiple sequence alignment	Preprocessing of genomic data for phylogenetic analysis
Phylogenetic Reconstruction	IQ-TREE v1.6.6 [25]	Maximum likelihood tree building	Inferring evolutionary relationships from genetic sequences
Phylodynamic Inference	BEAST [6]	Bayesian evolutionary analysis	Estimating evolutionary parameters and population dynamics
Transmission Cluster Analysis	HyPhy [26]	Hypothesis testing using phylogenetics	Identifying molecular transmission clusters
Transmission Tree Inference	phybreak [20]	Transmission network reconstruction	Inferring who-infected-whom from genomic data
Network Analysis	NETWORK 10.2.0.0 [25]	Median-joining network construction	Visualizing genetic relationships between closely related sequences
Statistical Analysis	R Software [25]	Data analysis and visualization	Implementing ERGM and other statistical models
Molecular Evolution	MEGA [26]	Molecular evolutionary genetics analysis	Comparative analysis of genetic sequences

Discussion and Future Directions

The integration of epidemiological and evolutionary data represents a paradigm shift in outbreak investigation and source attribution. Our comparative analysis demonstrates that method selection should be guided by specific research questions, data availability, and computational resources. For rapid assessment of transmission clusters, HyPhy offers significant advantages in computational efficiency and clustering sensitivity compared to MEGA [26]. For complex outbreaks with heterogeneous transmission, Bayesian birth-death models provide robust inference but require careful consideration of cluster influence and sample size effects [24].

A critical insight from our analysis is that model misspecification can introduce inductive biases, particularly when simple models are applied to complex transmission systems [6]. However, structured coalescent models can still recover accurate migration rates despite some simplification of epidemiological dynamics [6]. For large-scale pandemic modeling with rapid variant emergence, multi-scale agent-based approaches like PhASE TraCE offer the unique advantage of capturing feedback between evolutionary dynamics, intervention policies, and human behavior [8].

Future methodological development should address several key challenges: improving computational scalability for large genomic datasets, developing standardized approaches for integrating heterogeneous data sources, and creating robust methods for real-time phylodynamic inference during ongoing outbreaks. As sequencing technologies continue to advance and genomic surveillance becomes more routine, the conceptual framework presented here will serve as a foundation for the next generation of phylodynamic tools in public health practice.

A Landscape of Phylodynamic Methods: From Theory to Real-World Application

Bayesian phylogeographic models have emerged as a powerful statistical framework for reconstructing the spatiotemporal spread and evolution of pathogens. These methods combine molecular sequence data with epidemiological, geographic, and temporal information to infer patterns of pathogen dispersal across landscapes. The foundational principle underpinning these approaches is that evolutionary relationships inferred from genetic sequences, when calibrated in time, contain valuable information about the demographic history and spatial dynamics of pathogen populations [10]. In the context of infectious disease outbreaks, this enables researchers to address critical questions about origin estimation, the number of independent introductions, rates of spread between locations, and the impact of interventions.

The field represents a synthesis of evolutionary biology, epidemiology, and spatial statistics. Phylogeography specifically models how discrete or continuous traits, such as geographic location, evolve along the branches of a time-scaled phylogenetic tree [27] [10]. When these models are applied within a Bayesian statistical framework, they naturally quantify uncertainty in parameter estimates—including tree topology, divergence times, and evolutionary rates—providing a posterior distribution of possible scenarios consistent with the observed data [28]. The integration of such models with epidemic birth-dedeath processes has given rise to the subfield of phylodynamics, which aims to understand the interaction of evolutionary and ecological processes shaping pathogen populations [10].

For outbreak source attribution research, a key quantity of interest is often the root state of the inferred phylogeny, which represents the geographic origin of the sampled outbreak [27]. The performance of different models in accurately identifying this root state, and the factors affecting this performance, forms a critical basis for comparison. The following sections provide a comparative analysis of leading software packages, their underlying models, performance characteristics, and experimental protocols for evaluating their accuracy in outbreak source attribution.

Comparative Analysis of Software Platforms

Multiple software platforms implement Bayesian phylogeographic inference, each with distinct strengths, model offerings, and computational characteristics. The table below provides a structured comparison of three prominent tools.

Table 1: Comparison of Bayesian Phylogeographic Software Platforms

Software	Core Strengths	Primary Phylogeographic Models	Key Innovations	Performance & Scalability
BEAST X [29]	Integrated Bayesian inference, rich model library, active community development.	Discrete-trait CTMC, Relaxed Random Walk (RRW), Structured Birth-Death.	Hamiltonian Monte Carlo (HMC) samplers; models for sampling bias; missing data integration.	HMC samplers provide ~5x faster convergence for skygrid models; efficient for large datasets [29].
MTML-msBayes [30]	Hierarchical Approximate Bayesian Computation (HABC) for multi-taxon, multi-locus comparative phylogeography.	Multi-taxon coalescent model with divergence and migration.	Hyper-parameters to quantify variability in divergence times across taxon-pairs.	Computationally efficient for complex multi-taxon models via ABC, but accuracy depends on summary statistics.
EpiFusion [22]	Joint inference from phylogenetic trees and case incidence data via particle filtering.	Particle MCMC integrating incidence data and tree(s).	"Single process model, dual observation model" particle filter.	Fits force of infection via particle filter; other parameters via MCMC; suitable for outbreak-scale analysis.

BEAST X represents the state-of-the-art, introducing significant advances in flexibility and scalability. Its novel shrinkage-based local clock model offers a more tractable and interpretable alternative to the classic random local clock, while new preorder tree traversal algorithms enable linear-time gradient evaluations for high-dimensional parameters [29]. This computational efficiency allows BEAST X to handle the large genomic datasets now common in pathogen research.

MTML-msBayes serves a specific niche in comparative phylogeography. Instead of focusing on a single pathogen, it uses Hierarchical Approximate Bayesian Computation (HABC) to infer patterns of divergence and gene flow across multiple codistributed species or populations (taxon-pairs) [30]. This is particularly useful for identifying common biogeographic histories.

EpiFusion takes a different approach by formally integrating two key data sources: phylogenetic trees and case incidence data. Its particle filtering framework is designed to infer the effective reproduction number (R_t) and infection trajectories by evaluating simulated outbreaks against both types of data [22]. This joint inference can provide a more robust understanding of outbreak characteristics.

Performance in Source Attribution and Key Findings

Evaluating the performance of Bayesian phylogeographic models, particularly their accuracy in root state classification (source attribution), is crucial for applied public health. Simulation studies have revealed how model performance is influenced by data set characteristics.

Table 2: Key Factors Influencing Root State Classification Accuracy

Factor	Impact on Root State Classification	Supporting Evidence
Data Set Size	Performance is highest at intermediate sequence data set sizes; very small datasets lack signal, while very large datasets can introduce complex model fit challenges [27].	Simulation studies measuring classification accuracy across a range of dataset sizes (10s to 1000s of sequences) [27].
Discrete State Space Size	As the number of possible discrete locations (state space) increases, the difficulty of the classification task also increases, requiring more data for the same level of accuracy [27].	Logistic regression modeling of accuracy against state space size (e.g., from 2 to 56 discrete states) [27].
Sampling Bias & Metadata Uncertainty	Models are sensitive to geographic sampling bias. Missing or uncertain location metadata for sequences can significantly impact inference if not properly accounted for [27] [29].	Development of the Uncertain Trait Model (UTM) to incorporate sampling probability mass functions (PMFs) for tips with missing data [27].
Model Parameterization	Incorporating prior epidemiological information and using advanced spatial models (e.g., RRW) can improve accuracy by better reflecting realistic spread processes [29].	Comparison of discrete trait analysis (DTA) vs. structured birth-death models; continuous phylogeography with biased priors [29] [2].

A critical insight from systematic evaluations is that a common model evaluation metric, the Kullback-Leibler (KL) divergence, tends to increase with both larger state spaces and larger data set sizes. However, statistical modeling has shown that KL divergence is not a reliable predictor of root state classification accuracy [27]. This indicates that relying solely on KL divergence for model selection can be misleading, potentially favoring models with artificially inflated support.

The Uncertain Trait Model (UTM) provides a coherent method for incorporating sequences with missing or uncertain location metadata. Instead of discarding such sequences, UTM allows the researcher to specify a prior probability mass function over possible states. Studies show that an "informed" UTM prior (where most mass is on the correct trait) can improve inference, while a "misspecified" prior can harm it, highlighting the importance of careful prior specification [27].

Experimental Protocols for Model Validation

Robust validation of phylogeographic models relies on simulation-based approaches where the "true" history is known, allowing for direct assessment of inference accuracy. The following workflow outlines a standard protocol for such performance evaluation.

Well-Calibrated Simulation Study

A well-calibrated simulation study tests whether the software implementation can accurately recover known parameters across repeated analyses [28]. The specific protocol is as follows:

• Parameter and Tree Simulation: From a specified Bayesian model (e.g., containing an HKY substitution model, uncorrelated lognormal relaxed clock, and a Yule tree prior), randomly draw 100 or more independent sets of parameters (e.g., base frequencies, transition-transversion ratio, birth rate) and corresponding time-scaled phylogenetic trees [28].
• Sequence Alignment Simulation: For each parameter set and tree, simulate a nucleotide sequence alignment using a phylogenetic continuous-time Markov chain. This represents the evolutionary process along the branches of the tree.
• Phylogeographic Inference: Using only the simulated sequences and their sampling times (and associated discrete traits for phylogeography), perform full Bayesian phylogeographic inference with the software and model being tested.
• Validation: Compare the posterior estimates of parameters and the root state to the true values used in the simulation. A well-calibrated model should contain the true value within the 95% Highest Posterior Density (HPD) interval approximately 95% of the time [28].

Performance Benchmarking

To compare computational efficiency and sampling performance between different software or operators, the following protocol is used:

• Data Sets: Run analyses on a range of datasets, from small to large (e.g., 50 to 500 taxa).
• MCMC Execution: Perform MCMC sampling for a fixed number of steps or until convergence is achieved.
• Metrics Calculation: Calculate the Effective Sample Size (ESS) for each key parameter, which estimates the number of independent samples from the posterior. Also record the total computer running time.
• Efficiency Comparison: The primary metric for comparison is the ESS per hour. An operator or software that yields a higher ESS per hour for the same level of accuracy is considered more efficient [28]. For example, the "Constant Distance" operator developed for BEAST 2 was shown to improve overall mixing efficiency by up to half an order of magnitude for large data sets [28].

The Researcher's Toolkit

Implementing Bayesian phylogeographic analyses requires a suite of software tools and research reagents. The table below details essential components for a standard workflow.

Table 3: Essential Research Reagents and Software for Bayesian Phylogeography

Tool Category	Specific Examples	Function in Analysis
Primary Inference Software	BEAST X [29], BEAST 2 [28], MTML-msBayes [30], EpiFusion [22]	Core platforms for performing Bayesian MCMC or ABC inference of phylogenetic trees, evolutionary parameters, and trait diffusion.
High-Performance Computing Library	BEAGLE [28] [29]	A software library that uses parallel processing (CPUs/GPUs) to drastically accelerate likelihood calculations, which are the computational bottleneck in phylogenetic inference.
Result Analysis & Visualization	Tracer [28], FigTree, R packages (ape, ggtree, EpiFusionUtilities [22])	Used to diagnose MCMC convergence (ESS), summarize posterior distributions of parameters and trees, and visualize phylogenetic trees with annotated traits.
Sequence Data & Management	GenBank, GISAID, PANGOLIN	Public repositories for obtaining sequence data with metadata. Tools for lineage assignment and preliminary analysis.
Uncertain Trait Pipelines	Geographic Location Resolution Pipelines [27]	Bioinformatic tools that output probability mass functions (PMFs) for the location of infected host (LOIH) for sequences with missing metadata, for use with the Uncertain Trait Model.
Adam-20-S	Adam-20-S, MF:C17H21FN2O4S, MW:368.4 g/mol	Chemical Reagent
Cox-2-IN-34	Cox-2-IN-34, MF:C13H11NO4, MW:245.23 g/mol	Chemical Reagent

The relationship between these tools in a standard phylogeographic analysis is visualized below.

Bayesian phylogeographic models are indispensable tools for outbreak source attribution research. The choice of software and model—whether it is the comprehensive and scalable BEAST X, the comparative MTML-msBayes, or the data-integrating EpiFusion—should be guided by the specific research question, the nature and scale of the data, and computational constraints. Performance validation studies consistently show that accuracy is highest at intermediate data set sizes and is challenged by large discrete state spaces and sampling bias. The adoption of modern techniques like the Uncertain Trait Model, Hamiltonian Monte Carlo samplers, and models that correct for sampling bias is critical for generating robust, actionable insights for public health intervention. As the field evolves, continued rigorous evaluation of model performance under realistic conditions will ensure these powerful methods remain reliable guides for understanding and combating infectious disease outbreaks.

Phylodynamics integrates epidemiological and genetic data to reconstruct the transmission dynamics of infectious diseases, a capability that is crucial for effective outbreak management. Two mathematical frameworks form the cornerstone of modern phylodynamic inference: the birth-death model and the coalescent model. Each provides a distinct method for relating the phylogenetic tree of pathogen samples to the underlying population processes. The birth-death model is a forward-time process that describes the transmission (birth) and removal (death) of infected individuals, from which the observed phylogeny is a subsample. In contrast, the coalescent model is a backward-time process that starts with the sampled individuals and traces their lineages backward in time until they merge (coalesce) into common ancestors. While often used interchangeably in some studies, these models operate under fundamentally different assumptions about population dynamics and sampling, which directly impacts their performance in estimating key epidemiological parameters such as migration rates, growth rates, and the basic reproductive ratio ( [31] [32]). The choice between these models is not merely academic; it significantly influences the accuracy and reliability of the epidemiological insights gained, particularly in outbreak source attribution research. This guide provides a structured, data-driven comparison to inform this critical methodological choice.

Performance Comparison: Key Quantitative Findings

Direct comparative studies reveal that the performance of birth-death and coalescent models is highly dependent on the epidemiological context, specifically whether the disease is in an early epidemic growth phase or a stable endemic state.

Table 1: Model Performance Across Epidemiological Scenarios

Epidemiological Scenario	Performance Metric	Birth-Death Model	Coalescent Model	Key Findings
Epidemic Outbreak	Accuracy of Migration Rate	Superior (accurate across migration rates) [32]	Less accurate [32]	Birth-death model better accounts for population dynamics [32].
	Coverage of Growth Rate (HPD Interval)	Higher Coverage (2-13% error rate) [31]	Lower Coverage (31-75% error rate) [31]	Coalescent's deterministic population assumption is problematic in early outbreaks [31].
Endemic Disease	Accuracy & Precision of Migration Rate	Comparable Accuracy [32]	Comparable Accuracy, Higher Precision [32]	Both models perform well; coalescent may yield more precise estimates [32].
Source Location Identification	Accuracy	Comparable [32]	Comparable [32]	Both models similarly estimate the source of the disease [32].

Performance in Epidemic Outbreaks

For epidemic outbreaks characterized by exponential growth, the birth-death model demonstrates a clear advantage. A simulation study found it exhibits a superior ability to retrieve accurate migration rates regardless of the actual migration rate, whereas the structured coalescent model with a constant population size can lead to inaccurate estimates [32]. Furthermore, when estimating the epidemic growth rate from phylogenetic trees simulated under a birth-death process, the birth-death model achieved a much higher coverage probability, meaning the true parameter value was contained within the 95% highest posterior density (HPD) interval far more often (87-98% of the time) compared to the coalescent model with exponential growth (25-69% of the time) [31]. This superior performance is attributed to the birth-death model's inherent ability to account for the stochastic population fluctuations that are pronounced in the early phase of an outbreak. The coalescent model, which often assumes a deterministically changing population size, struggles to capture this early stochasticity [31].

Performance in Endemic Situations

In contrast, for endemic scenarios where the infected population size is relatively stable, the performance gap between the two models narrows significantly. A comparative investigation demonstrated that both models produce comparable coverage and accuracy for estimating migration rates in this context [32]. Interestingly, the same study noted that the coalescent model can even generate more precise estimates (tighter confidence intervals) than the birth-death model for endemic diseases [32]. This makes the coalescent a valid and potentially preferable option for studying the spread of pathogens in stable, endemic settings.

Experimental Protocols for Model Comparison

The quantitative findings summarized above are derived from rigorous simulation studies. The following outlines the standard protocol employed by researchers to objectively compare the performance of birth-death and coalescent models.

Protocol 1: Assessing Migration Rate Estimation

This protocol is designed to evaluate the models' ability to infer pathogen spread between subpopulations [32].

• Tree Simulation: Generate a large number of phylogenetic trees using a known, simulated multi-type birth-death process. This process defines the "true" history with known parameters, including specified migration rates between demes and a pre-defined population growth dynamic (e.g., epidemic or endemic).
• Parameter Inference: For each simulated tree, analyze it using both the structured coalescent model (e.g., with constant population size) and the multitype birth-death model (e.g., with a constant rate) in a Bayesian statistical framework like BEAST2.
• Output Analysis: For each analysis, record the estimated migration rate parameters and their associated measures of statistical uncertainty.
• Performance Calculation:
  • Accuracy: Calculate the difference between the median estimated migration rate and the known true value used in the simulation.
  • Precision: Calculate the width of the 95% HPD interval for the estimates.
  • Coverage: Determine the percentage of analyses in which the true parameter value falls within the estimated 95% HPD interval.

Protocol 2: Assessing Growth Rate Estimation

This protocol specifically tests the models' performance in inferring the rate of epidemic spread [31].

• Tree Simulation under Contrasting Models:
  • Simulate one set of phylogenetic trees under a constant-rate birth-death model.
  • Simulate a second set of trees under a coalescent model with a deterministically, exponentially growing infected population.
• Cross-Inference: Analyze each set of simulated trees using both the birth-death and coalescent inference models. This creates a scenario where the data-generating process and the inference model are sometimes mismatched.
• Performance Calculation: The key metric is, again, the coverage probability of the growth rate parameter. This tests the robustness of each inference model when its underlying assumptions are violated by the data.

Figure 1: A generalized workflow for a phylodynamic model comparison study, illustrating the process of simulation, inference, and evaluation.

The Scientist's Toolkit: Essential Research Reagents & Solutions

Successful implementation of the phylodynamic models discussed requires a suite of specialized software tools and computational resources.

Table 2: Essential Research Reagents for Phylodynamic Analysis

Tool / Resource	Function	Relevance to Models
BEAST 2	A cross-platform software for Bayesian evolutionary analysis of molecular sequences using MCMC.	Primary framework for implementing both birth-death and coalescent model inference [3] [31].
bdmm Package	A BEAST 2 package for multi-type birth-death inference.	Enables phylodynamic analysis under the birth-death model for structured populations; improved to handle larger datasets (>250 samples) [3].
ModelFinder (IQ-TREE)	A fast model selection method for accurate phylogenetic estimates.	Used to select the best-fitting nucleotide/amino acid substitution model and model of rate heterogeneity, which is a critical step prior to phylodynamic inference [33].
High-Performance Computing (HPC) Cluster	A network of computers for computationally intensive tasks.	Essential for running complex Bayesian MCMC analyses, which are computationally demanding and time-consuming.
Structured Sequence Data with Metadata	Pathogen genetic sequences annotated with data such as sampling date and location.	The fundamental input data for phylodynamic analysis. Rich metadata is crucial for meaningful structured model analysis.
SARS-CoV-2-IN-53	SARS-CoV-2-IN-53, MF:C23H18F2N2O4S, MW:456.5 g/mol	Chemical Reagent
Icmt-IN-27	ICMT-IN-27|ICMT Inhibitor	ICMT-IN-27 is a potent ICMT inhibitor (IC50=0.1 µM) for cancer research. For Research Use Only. Not for human use.

The choice between birth-death and coalescent models is not one of absolute superiority but of contextual fitness. The experimental data consistently demonstrates that the birth-death model is the more robust and reliable choice for analyzing epidemic outbreaks, where its explicit modeling of transmission and removal events allows it to naturally accommodate the stochastic population dynamics of an emerging pathogen. For endemic diseases or stable populations, both models are viable, with the coalescent sometimes offering advantages in computational efficiency and estimation precision. When the primary research goal is outbreak source attribution, both models perform equally well in identifying the source location [32]. Therefore, researchers should base their model selection on the specific epidemiological context of their study, the key parameters of interest, and the nature of the available data. As the field progresses, the development of more complex models that integrate genomic and ecological data will further refine our ability to reconstruct and forecast pathogen spread.

The field of phylodynamics, which unifies epidemiological processes with pathogen evolutionary dynamics, has become indispensable for modern outbreak response. It enables researchers to infer critical variables such as transmission trees, reproductive numbers, and migration patterns. However, the exponential growth of pathogen genomic data—exemplified by millions of SARS-CoV-2 sequences—has exposed significant computational bottlenecks in traditional methods [34] [35]. These bottlenecks hinder real-time analysis during public health emergencies.

Two innovative frameworks, SPRTA (Subtree Pruning and Regrafting-based Tree Assessment) and ScITree (Scalable Bayesian inference of Transmission tree), have emerged to address this challenge. Each tackles a distinct yet complementary aspect of phylodynamic inference. SPRTA revolutionizes the assessment of confidence in massive phylogenetic trees, while ScITree enables scalable, accurate reconstruction of transmission trees. This guide provides a detailed, objective comparison of their performance, methodologies, and applicability for researchers and drug development professionals engaged in outbreak source attribution.

SPRTA: Scalable Phylogenetic Confidence Assessment

SPRTA (Subtree Pruning and Regrafting-based Tree Assessment) introduces a paradigm shift in measuring confidence for phylogenetic trees inferred from millions of genomes. Traditional methods, like Felsenstein's bootstrap, require computationally prohibitive data resampling and are poorly suited to pandemic-scale datasets [34] [36].

• Core Innovation: SPRTA shifts from a "topological focus" (evaluating confidence in clades) to a "mutational focus." It assesses the probability that a specific branch in a tree correctly represents the evolutionary origin of a lineage or variant [34].
• Mechanism: For each branch in a given tree, SPRTA systematically explores alternative evolutionary histories by performing Subtree Pruning and Regrafting (SPR) moves. These moves represent plausible alternative placements of a lineage within the tree. The support for the original branch is then calculated as a function of the likelihoods of the original tree and these alternative topologies [34].
• Integration: It is embedded within widely used phylogenetic software, including MAPLE and IQ-TREE, making it accessible for genomic epidemiological studies [36].

ScITree: Scalable Transmission Tree Inference

ScITree addresses a different computational bottleneck: the inference of transmission trees (who-infected-whom) from epidemiological and genomic data. While the previous Bayesian mechanistic model by Lau et al. was highly accurate, it faced major scalability issues due to its explicit, nucleotide-level modeling of mutations [35].

• Core Innovation: ScITree overcomes this by incorporating the infinite sites assumption to model the evolutionary process. Instead of imputing every single nucleotide, it models genetic mutations between sequences through time as a Poisson process, dramatically reducing the computational parameter space [35].
• Mechanism: It is a fully Bayesian mechanistic model that uses an exact likelihood to integrate spatio-temporal SEIR (Susceptible-Exposed-Infectious-Removed) models with observed genomic and epidemiological data. Its efficient data-augmentation Markov Chain Monte Carlo (MCMC) algorithm enables inference of the complete transmission tree and key epidemiological parameters [35].
• Implementation: ScITree is available as an open-source R package, facilitating adoption by the public health and research communities [35].

Performance Comparison and Experimental Data

The following tables summarize key performance metrics and characteristics of SPRTA and ScITree based on published benchmarks and simulations.

Table 1: Key Performance and Benchmarking Data

Metric	SPRTA	ScITree	Context & Notes
Computational Demand	>2 orders of magnitude reduction in runtime/memory vs. bootstrap methods [34]	Linear scaling with outbreak size; significant improvement over Lau method's exponential scaling [35]	Benchmark against Felsenstein's bootstrap, aLRT, etc. for SPRTA; benchmark against predecessor for ScITree.
Scalability Demonstrated On	Dataset of >2 million SARS-CoV-2 genomes [34] [36]	Simulated outbreaks; real-world FMD outbreak (UK, 2001) [35]	Demonstrates practical application at pandemic scale.
Inference Accuracy	N/A (Assesses confidence, not tree topology)	Comparable to the highly accurate Lau method [35]	Accuracy measured by transmission tree reconstruction in simulations.
Primary Output	Confidence scores for phylogenetic branches	Transmission tree, epidemiological parameters (e.g., reproductive number) [35]	Outputs are complementary for a full phylodynamic analysis.
Handling of Uncertainty	Identifies plausible alternative evolutionary origins for lineages [34]	Full Bayesian framework providing posterior distributions for all inferred parameters [35]	Both provide robust uncertainty quantification.

Table 2: Methodological Comparison and Application Scope

Aspect	SPRTA	ScITree
Primary Goal	Assess confidence in phylogenetic trees	Infer transmission trees and epidemiological dynamics
Core Method	Likelihood comparison via Subtree Pruning & Regrafting (SPR) moves	Bayesian MCMC with infinite sites mutation model
Epidemiological Model	Not directly integrated	Integrated spatio-temporal SEIR model
Ideal Use Case	Evaluating reliability of large-scale phylogenies for tracking variant emergence	Reconstructing fine-grained transmission dynamics and superspreading events
Key Advantage	Interpretability and speed for massive trees	Scalability without sacrificing mechanistic accuracy

Detailed Experimental Protocols

SPRTA Workflow and Benchmarking

The validation of SPRTA involved a rigorous benchmarking process against established methods.

• Experimental Setup: Researchers compared SPRTA against several branch support methods, including Felsenstein’s bootstrap, local bootstrap probability (LBP), and approximate likelihood ratio test (aLRT), using simulated SARS-CoV-2-like genome datasets where the true evolutionary history was known [34].
• Workflow:
  • Input: A multiple sequence alignment and an inferred rooted phylogenetic tree.
  • SPR Operation: For each branch b in the tree, the algorithm generates alternative tree topologies by performing SPR moves. This involves pruning the subtree S_b and regrafting it onto other parts of the tree to create hypothetical alternative origins for that lineage.
  • Likelihood Calculation: The likelihood of the original tree and each alternative topology is computed.
  • Support Score Calculation: The SPRTA support for branch b is calculated as the likelihood of the original tree divided by the sum of the likelihoods of all considered alternative topologies [34]. This represents the approximate probability that the branch correctly represents the evolutionary origin of the lineage.
• Outcome Measurement: The primary metrics were computational runtime/memory usage and the accuracy of the mutational history assessment, confirming SPRTA's efficiency and reliability [34].

ScITree Validation and Simulation Study

The performance of ScITree was assessed through comprehensive simulations and a real-data application.

• Experimental Setup: The method was tested on multiple simulated outbreak datasets of varying sizes. Its performance was compared against the predecessor (Lau method) in terms of accuracy and computational time. It was also applied to a documented Foot-and-Mouth Disease (FMD) outbreak [35].
• Workflow:
  • Model Definition: A stochastic, continuous-time spatio-temporal SEIR model is specified, defining the rates at which individuals transition between compartments based on spatial proximity and other parameters.
  • Data Integration: The model incorporates observed data: infection times, removal times, genetic sequences, and individual locations.
  • Inference via MCMC: The algorithm employs a data-augmentation MCMC to explore the posterior distribution of the unobserved transmission tree, model parameters, and the mutational history under the infinite sites model.
  • Output: The result is a posterior distribution over all unknown quantities, including the who-infected-whom network and key parameters like the reproductive number [35].
• Outcome Measurement: Accuracy was measured by the coverage of the true transmission tree and parameter estimates in simulations. Computational efficiency was measured by the runtime relative to outbreak size, demonstrating linear scaling [35].

Visualizing the Core Workflows

The diagrams below illustrate the fundamental logical processes underlying the SPRTA and ScITree frameworks.

Figure 1: SPRTA Branch Confidence Assessment

Figure 2: ScITree Phylodynamic Inference Process

Successful implementation of these frameworks relies on a suite of software tools and data resources.

Table 3: Key Research Reagents and Solutions

Tool/Resource	Function	Relevance
MAPLE [36]	Software for efficiently constructing massive phylogenetic trees from millions of genomes.	Provides the initial phylogenetic tree required for SPRTA analysis.
IQ-TREE [36]	A widely-used software package for phylogenetic inference by maximum likelihood.	Another platform where the SPRTA method is available for user-friendly application.
phydynR [37]	An R package for phylodynamic analysis using structured coalescent and birth-death models.	An alternative tool for phylodynamic inference, mentioned in comparative studies.
ScITree R Package [35]	The official implementation of the ScITree model for Bayesian transmission tree inference.	The essential software reagent for deploying the ScITree framework.
Structured Coalescent Models [37]	A class of population genetic models used to infer population sizes and migration rates from genealogies.	Provides the theoretical foundation for many phylodynamic methods, a context for ScITree's advances.
Multiple Sequence Alignment [34]	The fundamental input data structure representing homologous nucleotides across all sampled pathogen sequences.	A mandatory input for both phylogenetic tree building (for SPRTA) and direct analysis in ScITree.

SPRTA and ScITree represent significant, complementary leaps forward in scalable phylodynamic inference. SPRTA is the specialized tool for researchers who need to quickly vet the reliability of phylogenetic relationships in trees built from millions of sequences, offering unprecedented speed and interpretability for tracking variant evolution. ScITree is the comprehensive solution for epidemiologists aiming to reconstruct precise transmission networks and estimate key parameters with mechanistic rigor, achieving scalability without compromising the accuracy of its full-Bayesian framework.

For outbreak source attribution research, the choice between—or sequential use of—these frameworks depends on the core scientific question. If the goal is to understand the broad evolutionary origins and confidence of a pathogen's lineage on a global scale, SPRTA is indispensable. If the objective is to pinpoint individual transmission links and superspreading events within a local outbreak, ScITree provides the necessary granularity. Together, they equip the scientific community with robust, scalable tools to enhance pandemic preparedness and response.

The reconstruction of outbreak transmission chains, known as source attribution, is a cornerstone of effective public health response. Molecular source attribution methods that utilize pathogen genetic sequence data have become increasingly prevalent. This guide provides a comparative analysis of two dominant computational paradigms in this field: phylogenetic clustering methods and model-based source attribution, with a specific focus on the emerging integration of these methods into multi-scale agent-based models (ABMs). We evaluate their performance, data requirements, and applicability for researchers and public health professionals, highlighting how the synthesis of these approaches addresses critical gaps in modeling complex, evolving outbreaks.

Source attribution refers to a category of epidemiological methods with the objective of reconstructing the transmission of an infectious disease from a specific source, which could be a population, an individual, a location, or a specific event [38]. In practice, it is a problem of statistical inference because transmission events are rarely observed directly. Molecular source attribution uses the molecular characteristics of the pathogen—most often its nucleic acid genome—to reconstruct these transmission events [38].

The increasing affordability of whole-genome sequencing (WGS) has provided an unprecedented volume of high-resolution data for tracking pathogen spread. WGS represents the maximal extent of multi-locus typing, covering all possible loci in the genome, which significantly enhances the power to distinguish between even closely related lineages and provides a solid foundation for source attribution [38]. Two primary methodological frameworks have been developed to interpret this genetic data for epidemiological inference:

• Cluster-Based Methods: These methods identify groups of individuals (clusters) with closely related pathogen sequences, implying they are part of a recent transmission chain. The analysis often involves regressing cluster membership or size on patient covariates to identify transmission risk factors [39].
• Model-Based Source Attribution (SA) Methods: These methods go beyond simple clustering by estimating the probability that a given sampled case was the source of infection for another case (the infector probability). This approach uses a phylogenetic framework that can incorporate additional epidemiological data, such as incidence, prevalence, and time since infection, to weight the possible transmission links between cases [39].

The choice between these methods involves critical trade-offs between computational cost, statistical robustness, and the ability to model complex, multi-scale population dynamics, which we will explore in the following sections.

Comparative Performance Analysis of Phylodynamic Methods

A critical simulation study compared the ability of phylogenetic clustering and source attribution methods to identify patient attributes as transmission risk factors [39]. The study modeled HIV epidemics among men who have sex with men and generated phylogenies comparable to those from real-world surveillance data.

Table 1: Performance Comparison of Clustering vs. Source Attribution Methods

Feature	Phylogenetic Clustering Methods	Model-Based Source Attribution (SA)
Core Principle	Identifies groups with closely related pathogen sequences [39].	Estimates infector probabilities between each pair of individuals from a time-scaled phylogeny [39].
Key Output	Cluster membership (binary) or cluster size [39].	Out-degree ((di = \sum{j \neq i} W_{ij})), the estimated number of transmissions originating from a patient [39].
Statistical Robustness	Can show misleading associations with covariates correlated with time since infection (e.g., CD4 count, age) [39].	Can account for time since infection, reducing spurious associations [39].
Sensitivity & Error Rates	Usually has higher error rates and lower sensitivity for identifying true transmission risk factors [39].	Generally lower error rates and higher sensitivity than clustering methods [39].
Handling of Uncertainty	Relies on arbitrary genetic distance thresholds, neglecting informative links above the threshold [39].	Probabilistic framework naturally incorporates uncertainty; no need for arbitrary thresholds [39].
Computational Tractability	Computationally cheap once a phylogeny is estimated; applicable to very large datasets (tens of thousands of patients) [39].	More computationally intensive; scalability can be a challenge for very large sample sizes [6].

The findings indicate that while clustering methods are computationally efficient and easily implemented, they can produce misleading associations. A key weakness is that any patient covariate correlated with time since infection (e.g., CD4 count, viral load, age) is likely to be found associated with clustering, regardless of its actual role in transmission [39]. The model-based SA approach, by explicitly modeling the probability of transmission between pairs, generally achieves lower error rates and higher sensitivity, providing more robust estimates of transmission risk factors [39].

The Emergence of Multi-Scale Phylodynamic Agent-Based Models

A frontier in computational epidemiology is the development of multi-scale models that integrate within-host pathogen evolution with between-host transmission dynamics across entire populations. A prime example is the PhASE TraCE framework (Phylodynamic Agent-based Simulator of Epidemic Transmission, Control, and Evolution) [8] [40].

Model Architecture and Workflow

PhASE TraCE is a stochastic agent-based model (ABM) of pandemic spread coupled with a phylodynamic model of within-host pathogen evolution [8] [40]. It builds upon validated large-scale pandemic simulators and is designed to simulate feedback loops between public health interventions, population behavior, and pathogen evolution [40]. The following diagram illustrates the core multi-scale workflow of such a framework.

This multi-scale integration allows the framework to replicate key pandemic features, as required by its core capabilities [8] [40]:

• Modeling Pandemic Patterns: Reproducing recurrent incidence waves and transitions to endemicity.
• Examining Pathogen Fitness: Tracing how mutations accumulate and increase transmissibility (basic reproductive number (R_0)).
• Detecting Variants of Concern: Identifying the emergence of new variants via changes in genomic diversity and their alignment with incidence peaks.

Experimental Protocol for Model Validation

The validation of a phylodynamic ABM like PhASE TraCE against real-world data follows a rigorous protocol. A case study using SARS-CoV-2 genomic surveillance data from 2020-2024 typically involves the following steps [8] [40]:

• Data Integration: Contemporary genomic, demographic, and human mobility data are integrated into the model.
• Parameterization: The model is initialized with parameters describing the ancestral pathogen's transmissibility and the initial host population structure.
• Simulation Execution: The coupled ABM and phylodynamic model are run stochastically over a long-term horizon (e.g., multiple years) to generate numerous realizations of the pandemic.
• Outcome Comparison (Ground Truth Validation): The model's outputs are directly compared to observed ground truth data:
  • Capability 1 (Epidemic Patterns): Simulated incidence and prevalence curves are compared to reported case data to check for aligned peaks and waves [8] [40].
  • Capability 2 (Pathogen Evolution): The simulated trajectory of accumulated mutations and relative transmissibility is compared to empirical estimates derived from genomic surveillance [8] [40].
  • Capability 3 (Variant Emergence): The model's ability to generate new variants with similar phylogenetic patterns and diversity dynamics to those observed (e.g., Omicron BA.1, XBB) is assessed [8] [40].

Essential Research Reagents and Computational Tools

The experiments and models discussed rely on a suite of specialized computational tools and reagents.

Table 2: Key Research Reagent Solutions for Phylodynamic Modeling

Reagent / Tool	Type	Primary Function in Research
Whole Genome Sequence (WGS) Data [38]	Data	Provides the highest-resolution molecular evidence for distinguishing pathogen lineages and inferring transmission links. The fundamental input for analysis.
Time-Scaled Phylogenies [39] [41]	Data / Model Output	Represents the evolutionary relationships among pathogen sequences with branch lengths in units of time. Essential for estimating transmission rates and evolutionary history.
Agent-Based Modeling Frameworks (e.g., Repast [42])	Software Platform	Toolkits for building simulations of individual agents (people) and their interactions within a defined environment, used to model disease spread.
Phylodynamic Software (e.g., BEAST [6])	Software Platform	Infers pathogen population history, evolutionary rates, and demographic parameters from genetic sequence data and phylogenies.
Source Attribution Model [39]	Algorithm	A specific model that calculates infector probabilities ((W_{ij})) between cases using the phylogeny, incidence, prevalence, and clinical data.
Structured Coalescent Model [41] [6]	Mathematical Model	A population genetic model that describes how lineages coalesce in a structured population (e.g., with different demes or subpopulations), used to estimate migration rates.

The comparative analysis reveals a clear trajectory in the field of phylodynamics for outbreak source attribution. While traditional clustering methods offer simplicity and speed, they are susceptible to statistical artifacts and provide a relatively coarse picture of transmission dynamics [39]. Model-based source attribution methods offer a more robust and statistically sound framework for inferring transmission links and risk factors, though at a higher computational cost [39].

The most comprehensive approach, embodied by multi-scale phylodynamic ABMs like PhASE TraCE, represents a paradigm shift [8] [40]. These models synthesize the strengths of agent-based modeling—simulating heterogeneous populations and intervention scenarios—with the power of phylodynamics to track evolving pathogens. This integration directly addresses the "who infected whom" question while simultaneously modeling the complex feedback loops between human behavior, public health policy, and viral evolution that characterize modern outbreaks.

However, model specification remains a critical consideration. As highlighted in recent research, even sophisticated phylodynamic models can suffer from inductive bias if the model is misspecified or provides an overly simplistic representation of the underlying evolutionary and epidemiological processes [6]. Therefore, ongoing validation and refinement of these integrated models against real-world data are paramount. For researchers, the choice of method should be guided by the specific research question, the quality and volume of available data, and computational resources, with an understanding that the field is moving towards ever more integrated and dynamic multi-scale frameworks.

The COVID-19 pandemic, caused by the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2), triggered an unprecedented global public health crisis characterized by rapid international spread and the continual emergence of novel variants [43]. The Arabian Peninsula, particularly the Gulf Cooperation Council (GCC) countries, represents a critical region for studying viral transmission dynamics due to its status as a major global travel hub and its unique demographic characteristics [43]. Following the initial detection of the first SARS-CoV-2 cases in the United Arab Emirates in January 2020, the virus rapidly spread across the region, with subsequent cases reported in Kuwait, Bahrain, Qatar, Oman, and Saudi Arabia between mid-February and early March 2020 [43].

Phylodynamic methods, which combine evolutionary, demographic, and epidemiological concepts, have emerged as powerful tools for reconstructing viral transmission patterns, estimating growth rates, and identifying the origins and spread of concerning variants [2]. This case study employs a comparative framework to evaluate the application of structured phylodynamic models—specifically the structured coalescent and multi-type birth-death models—for investigating the introduction and dispersal of major SARS-CoV-2 variants across the Arabian Peninsula. The insights gained from this analysis provide valuable guidance for selecting appropriate methodological approaches for outbreak source attribution research.

Methodology: Phylodynamic Inference Framework

Data Collection and Genome Sequencing

The foundational step in phylodynamic analysis involves comprehensive data collection and genome sequencing. Between November 2020 and June 2021, multiple GCC member states implemented SARS-CoV-2 genomic surveillance programs, generating thousands of complete viral genomes [43]. The present analysis focuses on five World Health Organization-designated variants: Alpha (B.1.1.7), Beta (B.1.351), Delta (B.1.617.2), Kappa (B.1.617.1), and Eta (B.1.525) [43].

Experimental Protocol: Genome Sequencing and Processing

• Sample Collection: Respiratory specimens were collected from confirmed COVID-19 cases using nasopharyngeal/oropharyngeal swabs immersed in viral transport media.
• RNA Extraction: Viral RNA was extracted using commercial kits (e.g., QIAamp Viral RNA Kits) following manufacturer protocols.
• Library Preparation and Sequencing: Libraries were prepared using reverse transcription and amplification, followed by whole-genome sequencing on Illumina or Oxford Nanopore platforms.
• Genome Assembly: Raw sequencing reads were processed through quality control, adapter trimming, and reference-based mapping to generate consensus sequences.
• Data Deposition: Curated genome sequences were deposited in public repositories such as GISAID and GenBank to facilitate global surveillance efforts [43].

Phylodynamic Models for Variant Spread

Two primary structured phylodynamic models were applied to estimate viral spread between populations:

Structured Coalescent Model (Constant Population Size)

• Principle: This model extends the Kingman coalescent to structured populations, estimating the time to the most recent common ancestor of sequences from different subpopulations.
• Assumptions: Constant effective population size within each demographic; migration rates follow a continuous-time Markov process.
• Implementation: Bayesian inference using Markov Chain Monte Carlo (MCMC) sampling implemented in BEAST2 and MrBayes software.
• Parameters Estimated: Effective population sizes, symmetric or asymmetric migration rates, and time-scaled phylogenies [44].

Multi-type Birth-Death Model (Constant Rate)

• Principle: This model directly describes the transmission, recovery, and sampling processes, categorizing individuals into types based on geographic location or other traits.
• Assumptions: Constant birth (transmission) and death (recovery) rates across types; type-change events equivalent to migration.
• Implementation: MCMC sampling in specialized software packages such as BEAST2 with the MultiTypeTree package.
• Parameters Estimated: Reproduction numbers (Rt), transmission rates, sampling proportions, and migration rates between subpopulations [44].

The following workflow diagram illustrates the comprehensive phylodynamic analysis process:

Comparative Analysis of Phylodynamic Methods

Quantitative Performance Comparison

A recent simulation study quantitatively compared the performance of structured coalescent and multi-type birth-death models for estimating pathogen spread across different epidemiological scenarios [44]. The table below summarizes the key findings:

Table 1: Performance Comparison of Phylodynamic Models for Pathogen Spread Estimation

Performance Metric	Epidemic Outbreak Scenario	Endemic Disease Scenario
Migration Rate Accuracy	Birth-death model superior	Comparable performance between models
Migration Rate Precision	Coalescent model less precise	Coalescent model more precise
Source Location Estimation	Comparable performance	Comparable performance
Computational Demand	Higher for birth-death model	Higher for birth-death model
Sensitivity to Sampling	Coalescent model more sensitive	Birth-death model more robust

The superior performance of the multi-type birth-death model during epidemic outbreaks stems from its inherent capacity to directly capture exponential growth dynamics, which aligns with the rapid expansion phase characteristic of emerging variants [44]. In contrast, the structured coalescent model with constant population size assumptions fails to adequately account for these dynamic population changes, leading to less accurate migration rate estimates.

Application to SARS-CoV-2 Variants in the Arabian Peninsula

Applying these phylodynamic methods to SARS-CoV-2 genomic data from the Arabian Peninsula revealed distinct patterns of variant introduction and spread:

Table 2: Spatiotemporal Origins and Transmission Dynamics of SARS-CoV-2 Variants in the Arabian Peninsula

Variant	Primary Introduction Source	Introduction Period	Population Growth Pattern	Dominant Dispersal Routes
Alpha (B.1.1.7)	Europe	Mid-2020 to Early 2021	Sequential growth and decline	Europe Arabian Peninsula
Beta (B.1.351)	Africa	Mid-2020 to Early 2021	Sequential growth and decline	Africa Arabian Peninsula
Delta (B.1.617.2)	East Asia	Early 2021 to Mid-2021	Sequential growth and decline	East Asia Arabian Peninsula
Kappa (B.1.617.1)	Multiple Sources	2021	Sporadic, inconclusive	Limited international dispersal
Eta (B.1.525)	Multiple Sources	2021	Sporadic, inconclusive	Limited international dispersal

Bayesian phylodynamic analyses indicated that Alpha and Beta variants underwent sequential periods of exponential growth followed by decline, a pattern linked to the implementation and subsequent relaxation of non-pharmaceutical interventions (NPIs) between mid-2020 and early 2021 [43]. The Delta variant exhibited more complex dynamics, with its progression likely shaped by the combination of NPIs and the rapidly expanding vaccination coverage across the region.

The discrete trait phylogeographic analysis implemented in the Bayesian evolutionary framework revealed significant and intense dispersal routes between the Arabian Peninsula and other global regions, with air travel patterns strongly correlating with variant spread [43]. The restricted dispersal and stable effective population sizes of Kappa and Eta variants suggest they did not establish significant community transmission networks in the region.

The following diagram illustrates the structural differences between the two primary phylodynamic modeling approaches:

Successful implementation of phylodynamic analyses requires specialized computational tools and analytical resources. The following table catalogs key research reagent solutions essential for conducting robust variant spread investigations:

Table 3: Essential Research Reagents and Computational Tools for Phylodynamic Analysis

Research Tool	Category	Primary Function	Application in Variant Spread Analysis
QIAamp Viral RNA Kits	Laboratory Reagent	Viral RNA extraction from clinical specimens	Isolate high-quality RNA for genome sequencing
Illumina Sequencing Platforms	Laboratory Equipment	High-throughput genome sequencing	Generate complete viral genomes for analysis
BEAST2 Software Package	Computational Tool	Bayesian evolutionary analysis	Implement coalescent and birth-death models
MAFFT Algorithm	Computational Tool	Multiple sequence alignment	Prepare homologous sequences for phylogenetics
IQ-TREE Software	Computational Tool	Maximum likelihood phylogenetics	Reconstruct evolutionary relationships
RDP4 Software	Computational Tool	Recombination detection	Identify recombinant sequences in datasets
GISAID Database	Data Resource	Genomic data repository	Access global SARS-CoV-2 sequence data
NextStrain Platform	Visualization Tool	Real-time pathogen tracking	Visualize spatiotemporal spread patterns

These research reagents and tools formed the foundation for the genomic epidemiology infrastructure established across the Arabian Peninsula during the pandemic, enabling regional molecular surveillance programs that informed public health decision-making regarding intervention strategies targeting the most relevant variants [43].

Discussion

Methodological Recommendations for Outbreak Source Attribution

Based on our comparative analysis of phylodynamic methods applied to SARS-CoV-2 variant spread in the Arabian Peninsula, we propose the following methodological recommendations for outbreak source attribution research:

• For Rapidly Expanding Epidemics: Multi-type birth-death models should be prioritized when investigating outbreaks in their exponential growth phase, as they explicitly capture the dynamic population size changes characteristic of emerging pathogen spread [44].

• For Established Endemic Transmission: Structured coalescent models with constant population size assumptions provide comparable accuracy with greater precision for endemic diseases with stable transmission dynamics, making them suitable for persistent viral circulation patterns [44].

• For Comprehensive Outbreak Investigation: Employ both modeling frameworks complementarily when possible, as they offer different strengths in estimating migration rates and source locations, providing a more robust inference of transmission dynamics.

• For SARS-CoV-2 Specific Applications: Given the pattern of sequential variant replacement observed in the Arabian Peninsula, birth-death models are particularly appropriate for investigating the initial introduction and establishment phases of novel variants, while coalescent models may better capture the dynamics of variant decline and transition periods.

Implications for Public Health Response

The phylodynamic findings from the Arabian Peninsula have significant implications for public health preparedness and response strategies. The demonstrated role of the region as a hub for variant importation from multiple global sources underscores the critical importance of sustained genomic surveillance at points of entry and within community transmission networks [43]. The observed effectiveness of NPIs in shaping variant progression during the pre-vaccination era provides empirical support for maintaining readiness to implement such measures in response to future variant emergences.

Furthermore, the intense dispersal routes identified between the Arabian Peninsula and other global regions highlight the necessity of coordinated international surveillance efforts and data sharing agreements to enable early detection and containment of emerging variants. The establishment of regional molecular surveillance programs across the GCC countries represents a vital investment for guiding targeted intervention strategies and vaccine allocation decisions in response to evolving viral threats [43].

This case study demonstrates the powerful insights gained from applying comparative phylodynamic approaches to reconstruct SARS-CoV-2 variant spread in the Arabian Peninsula. The structured evaluation of coalescent and birth-death models provides a framework for selecting appropriate methodological approaches based on specific epidemiological contexts and research objectives. The superior performance of multi-type birth-death models for epidemic outbreaks recommends their prioritization for investigating emerging variant spread, while structured coalescent models offer advantages for endemic scenarios.

The continued evolution of SARS-CoV-2 and the persistent threat of novel viral emergences underscore the critical importance of maintaining and enhancing phylodynamic capabilities within global public health infrastructure. Future methodological developments should focus on integrating multiple data sources, improving computational efficiency, and enhancing model flexibility to better capture the complex interplay of evolutionary, demographic, and epidemiological processes shaping pathogen spread.

Navigating Phylodynamic Challenges: Bias, Scalability, and Model Selection

Phylodynamics has emerged as a crucial discipline at the intersection of evolutionary biology and epidemiology, enabling researchers to infer critical parameters of infectious disease spread from pathogen genome sequences. These analyses reconstruct transmission dynamics and geographical sources of outbreaks by leveraging the evolutionary history embedded in the topology of phylogenetic trees, which serves as a natural record of infectious agent dispersal between geographical locations [5]. The fundamental premise of all phylodynamic inference is that epidemiological spread leaves a detectable trace in the form of substitutions in pathogen genomes that can be utilized to reconstruct transmission histories [45]. Pathogen populations meeting this assumption are classified as 'measurably evolving populations,' wherein molecular evolution occurs at rates sufficient to generate genetic diversity observable over epidemiological timescales.

The two fundamental data components driving all phylodynamic analyses are pathogen genome sequences and their associated sampling dates. Genome sequences provide the molecular evidence of evolutionary divergence, while sampling dates provide the temporal framework that allows researchers to model this divergence as a rate over time [45]. Despite their intertwined importance in phylodynamic frameworks, the relative contribution and sensitivity of inferences to these two data components remain inadequately characterized. This review systematically compares the individual and combined impacts of genomic sequence data versus temporal sampling data on the accuracy and precision of phylodynamic reconstructions, with particular emphasis on applications to outbreak source attribution research for public health response.

Theoretical Foundations of Sequence and Date Information

The Informational Value of Genomic Sequence Data

Genomic sequences provide the primary signal for reconstructing evolutionary relationships among pathogen samples. The phylogenetic tree topology, inferred from patterns of shared mutations across sequences, forms the backbone upon which phylodynamic models operate. Sequence data enable the identification of specific lineages, the detection of convergent evolution, and the characterization of selective pressures acting on pathogen populations. In source attribution studies, the geographical transitions embedded in tree nodes provide the historical record of spatial spread [5]. The quantity of sequence data directly impacts resolution, with larger datasets providing increased power to distinguish between alternative phylogenetic hypotheses and reducing uncertainty in parameter estimates.

The Informational Value of Sampling Date Data

Sampling dates provide the temporal calibration essential for translating genetic divergence into evolutionary rates and for estimating the timescale of epidemic spread. Precise sampling dates allow phylodynamic models to estimate the rate of molecular evolution, the time to the most recent common ancestor (tMRCA) of sampled pathogens, and the effective reproductive number (Re) through time [45]. The temporal distribution of samples significantly influences parameter estimation, with evenly spaced sampling through an outbreak providing more reliable inference than clustered sampling. Sampling dates also carry epidemiological information about case incidence patterns, though this is often secondary to their role in calibrating molecular clocks.

Quantitative Comparison of Data Impact

Experimental Evidence from Date-Rounding Studies

Recent investigations into the effects of reduced sampling date precision provide direct evidence of the critical importance of temporal data in phylodynamic inference. One comprehensive study analyzed bias in epidemiological parameter estimation across multiple pathogens when sampling dates were rounded to different precisions (day, month, or year) [45]. The researchers hypothesized that bias emerges when date uncertainty exceeds the average time for a substitution to arise in a given pathogen, causing substitution events to become conflated in temporal analyses.

Table 1: Impact of Date-Rounding on Phylodynamic Parameter Estimation

Pathogen	Substitution Rate (subs/site/year)	Time per Substitution (days)	Significant Bias at Month Resolution	Significant Bias at Year Resolution
H1N1 Influenza	4.0 × 10⁻³	7.0	Yes	Yes
SARS-CoV-2	1.0 × 10⁻³	33.4	Yes	Yes
Staphylococcus aureus	1.0 × 10⁻⁶	345.8	No	Yes
Mycobacterium tuberculosis	1.0 × 10⁻⁷	2325.6	No	No

The experimental protocol involved conducting phylodynamic analyses for both empirical and simulated datasets with sampling dates rounded to different precisions, then measuring the resulting bias in key epidemiological parameters (Re, substitution rate, and tMRCA). For viral pathogens with faster evolutionary rates like H1N1 influenza and SARS-CoV-2, rounding dates to month resolution already introduced substantial bias, while for slower-evolving bacterial pathogens like M. tuberculosis, year-rounding produced minimal bias [45]. This demonstrates that the relative impact of sampling date precision is modulated by the underlying molecular evolutionary rate of the pathogen, with temporal data becoming increasingly critical for fast-evolving viruses.

Impact on Specific Epidemiological Parameters

The direction and magnitude of bias varied across different estimated parameters, reflecting their differential dependence on temporal calibration:

Table 2: Parameter-Specific Effects of Reduced Date Precision

Estimated Parameter	Impact of Date-Rounding	Severity of Effect	Primary Mechanism
Reproductive Number (Re)	Variable direction, dataset-dependent	Moderate to Severe	Disruption of incidence curve timing
Substitution Rate	Systematic underestimation	Severe	Conflation of evolutionary events in time
tMRCA (outbreak age)	Systematic overestimation	Moderate to Severe	Imprecise rooting of temporal phylogenies
Geographical Transition Rates	Unquantified but predicted high	Unknown	Incorrect dating of spatial spread events

The substitution rate demonstrated the most consistent directional bias, with reduced date precision causing systematic underestimation. This occurs because evolutionary divergence appears compressed when sampling times are imprecise. Conversely, the time to the most recent common ancestor often became overestimated, as the root of the tree was pushed further back in time to accommodate the apparent evolutionary divergence under a slower inferred substitution rate [45]. The effective reproductive number showed more variable effects, with the direction of bias depending on the specific dataset and tree prior used in analysis.

Methodological Protocols for Signal Quantification

Experimental Design for Data Impact Assessment

To systematically quantify the relative contributions of sequence and sampling date information, researchers have developed standardized computational experiments. The following workflow illustrates the protocol for evaluating the sensitivity of phylodynamic inference to variations in data quality and completeness:

Diagram 1: Experimental workflow for quantifying the relative impact of sequence versus sampling date data on phylodynamic inference.

The experimental protocol involves several methodical stages. First, researchers begin with a high-quality dataset containing complete genome sequences with precise sampling dates. For sequence perturbation, common approaches include progressively downsampling the number of available sequences, introducing missing data at random positions, or adding sequencing error according to empirical error profiles. For date perturbation, studies typically either round dates to lower precision (e.g., to the nearest week, month, or year) or introduce random noise to sampling times. The key measurements involve comparing posterior distributions of critical parameters (Re, tMRCA, substitution rate, and spatial transition rates) between the perturbed datasets and the baseline high-quality dataset, calculating metrics like relative bias, posterior distribution overlap, and coefficient of variation.

Bayesian Phylodynamic Analysis Framework

Phylodynamic analyses typically employ Bayesian statistical frameworks that integrate multiple modeling components to simultaneously infer phylogenetic relationships, evolutionary parameters, and epidemiological dynamics. The core analysis pipeline for outbreak source attribution integrates several data types and models:

Diagram 2: Integrated Bayesian phylodynamic framework for outbreak source attribution.

The modeling framework incorporates several key components. The substitution model describes how nucleotide or amino acid sequences evolve over time, typically using site-homogeneous or heterogeneous models. The molecular clock model translates genetic divergence into time, using either strict or relaxed clock approaches to account for rate variation across lineages. The phylogeographic model reconstructs spatial spread, with popular approaches including ancestral state reconstruction and structured population models like the structured coalescent and birth-death models [5]. Finally, the population dynamic model infers changes in effective population size through time, often parameterized through skygrid or skyride models. These components are integrated in a unified Bayesian framework that simultaneously estimates all parameters, properly accounting for uncertainty and interdependence between model components.

Case Study: SARS-CoV-2 Variant Spread in the Arabian Peninsula

Application to Real-World Surveillance Data

A recent study applying Bayesian phylodynamic methods to SARS-CoV-2 variants in the Arabian Peninsula provides a compelling case study of sequence and sampling date integration for outbreak investigation [43]. Researchers analyzed genomic surveillance data from Gulf Cooperation Council (GCC) countries to compare the evolutionary dynamics, spatiotemporal origins, and spread patterns of five variants (Alpha, Beta, Delta, Kappa, and Eta). The study utilized 7,434 high-quality SARS-CoV-2 genomes from the region, with sampling dates spanning from mid-2020 to mid-2021, providing a robust dataset for assessing data impacts.

The analysis revealed distinct patterns of variant introduction and spread: Alpha and Beta variants were frequently introduced into the Arabian Peninsula between mid-2020 and early 2021 from Europe and Africa, respectively, while the Delta variant was primarily introduced between early 2021 and mid-2021 from East Asia [43]. The research demonstrated how precise sampling dates enabled researchers to correlate variant emergence with non-pharmaceutical interventions and vaccination campaigns, showing that containment measures between mid-2020 and early 2021 likely reduced epidemic progression of Beta and Alpha variants, while the combination of interventions and rapid vaccination rollout shaped Delta variant dynamics.

Data Requirements for Effective Surveillance

The successful application of phylodynamics to SARS-CoV-2 surveillance highlights specific data requirements for robust inference. The study emphasized the importance of comprehensive genomic sampling across time and space, as sporadic introductions of variants like Kappa and Eta resulted in inconclusive population growth patterns due to insufficient data [43]. The authors advocated for establishing regional molecular surveillance programs to ensure effective decision-making regarding intervention allocation, highlighting the necessity of both sequence quality and temporal sampling density for actionable public health insights.

Table 3: Essential Research Reagents and Computational Tools for Phylodynamic Analysis

Tool/Resource	Category	Primary Function	Application in Signal Quantification
BEAST2	Software Platform	Bayesian evolutionary analysis	Primary framework for phylodynamic inference
Structured Coalescent	Model Type	Infer population structure	Source attribution with migration rates
Birth-Death Models	Model Type	Estimate reproductive number	Quantify epidemic growth and decline
NextStrain	Visualization	Real-time outbreak analytics	Data quality assessment and visualization
GISAID	Data Repository	Pathogen genome sharing	Source of sequence and date metadata
TreeTime	Software Tool	Molecular clock dating	Assess date precision impact on estimates
RASP	Software Tool	Ancestral state reconstruction	Geographical source attribution

The structured coalescent model approaches geographical inference by considering discrete populations with migration between them, inferring migration rates and ancestral locations directly from genetic data and sampling information [5]. Birth-death models provide an alternative framework that explicitly models transmission, recovery, and sampling events, offering advantages for estimating the effective reproductive number through time. The Bayesian Evolutionary Analysis by Sampling Trees (BEAST2) software platform integrates these modeling approaches, providing a flexible framework for assessing the relative impact of different data types through sensitivity analyses and model comparison.

Synthesis and Research Recommendations

The relative impact of genomic sequence data versus sampling date information in phylodynamic inference exhibits strong context-dependence, modulated by pathogen evolutionary rate, sampling scheme, and the specific epidemiological parameters of interest. For rapidly evolving RNA viruses like SARS-CoV-2 and influenza, sampling date precision proves particularly crucial, with month-level rounding introducing significant bias in substitution rate and tMRCA estimates [45]. For slower-evolving pathogens, sequence data quantity and quality may dominate inference, with temporal precision becoming less critical until rounding exceeds the average inter-substitution time.

This synthesis suggests specific guidelines for outbreak investigation resource allocation. For emerging viral pathogens with high evolutionary rates, ensuring precise sampling date documentation should receive priority comparable to sequence quality, as temporal uncertainty directly propagates to key epidemiological parameters. Research investments should focus on integrated analytical frameworks that simultaneously handle sequence uncertainty and temporal imprecision, particularly for fast-evolving pathogens where these data components interact strongly. Methodological development should prioritize models that explicitly account for date uncertainty, especially for historical outbreaks or surveillance systems where precise sampling dates are unavailable.

The evidence reviewed indicates that neither sequence nor sampling date data can be considered in isolation—their synergistic interaction drives robust phylodynamic inference. Future methodological comparisons should adopt standardized protocols for data perturbation analyses to systematically quantify the relative importance of each data type across the diverse range of pathogens confronting public health systems.

Overcoming Sampling Bias and Heterogeneous Surveillance

Phylodynamics, defined as the melding of immunodynamics, epidemiology, and evolutionary biology, has become a fundamental paradigm in infectious disease research [10]. This approach leverages pathogen genetic sequences to infer epidemiological dynamics, assuming that molecular evolutionary change and epidemiological processes occur on similar timescales [10]. For outbreak source attribution, phylodynamic methods enable researchers to reconstruct transmission networks, identify sources of infection, and quantify the contributions of different reservoirs or transmission pathways. The foundational principle is that branching times and tree topologies in phylogenetic trees reflect underlying transmission dynamics, leaving an imprint that can be decoded statistically [10].

However, the real-world application of these methods faces significant challenges from sampling bias and heterogeneous surveillance. Pathogen sequences are rarely collected systematically; instead, sampling intensity often correlates with outbreak size, resource allocation, or public attention [46]. This preferential sampling can systematically distort phylodynamic inferences, potentially leading to incorrect conclusions about transmission dynamics and source attribution [46]. Similarly, heterogeneous surveillance across geographic regions or host populations creates data gaps that complicate the reconstruction of complete transmission histories. Understanding these limitations and the methods developed to overcome them is crucial for researchers relying on phylodynamic approaches for outbreak investigation.

Comparative Analysis of Phylodynamic Method Performance

The table below summarizes key phylodynamic approaches and their performance characteristics relative to sampling challenges:

Table 1: Performance Comparison of Phylodynamic Methods Under Sampling Biases

Method Category	Key Characteristics	Performance with Sampling Bias	Data Requirements	Computational Scalability
Structured Coalescent Models	Simple representation of migration rates between populations	Recovers migration rates despite model simplicity; small bias (≤5%) with sample size ≥1000 sequences; higher migration rate estimation more accurate [6]	Partial gene sequences or complete genomes; metadata on sample location/origin	Not scalable for datasets ≥600 sequences in BEAST [6]
Nonparametric Coalescent (Skyline)	Infers effective population size changes through time; Gaussian process priors	Systematic bias when sampling times depend on population size; overestimates peaks when sampling intensifies during high prevalence [46]	Heterochronous sequences (sampled at different times)	Moderate; enhanced by INLA approximation [46]
Birth-Death Models	Models transmission, recovery, and sampling processes directly	Handles various sampling schemes; estimates sampling probability jointly with other parameters [47]	Time-stamped sequences; case count time series	Challenging for large outbreaks; approximate methods developed [47]
Preferential Sampling Correction	Explicitly models sampling times as inhomogeneous Poisson process dependent on Ne(t)	Reduces bias and improves precision when sampling correlates with prevalence [46]	Genealogy with sampling times	High; requires joint inference of population and sampling processes [46]
Integrated Phylogenetic-Epidemiological	Combines genomic data with epidemiological time series (e.g., Timtam)	Improves estimation of prevalence and reproduction number by leveraging both data types [47]	Sequences and case count time series	Efficient approximation suitable for large datasets [47]

The performance data reveal several important patterns. Simple models can yield surprisingly robust inferences despite model misspecification, with one HIV study finding that structured coalescent models could recover migration rates while adjusting for nonlinear epidemiological dynamics, with inductive bias decreasing substantially with sample sizes of ≥1000 sequences [6]. However, computational limitations persist, with phylogeographic models in BEAST failing to scale for datasets of 600 or more sequences [6].

Methods that explicitly account for sampling biases generally outperform those that assume random or fixed sampling schemes. A preferential sampling model that treated sampling times as an inhomogeneous Poisson process dependent on effective population size demonstrated both bias reduction and improved estimation precision compared to standard approaches [46]. Similarly, integrated approaches like Timtam that combine genomic and epidemiological time series data provide better estimates of key parameters like prevalence and effective reproduction numbers [47].

Experimental Protocols for Evaluating Sampling Bias

Protocol for Simulating Preferential Sampling Effects

To quantify the effects of preferential sampling on phylodynamic inference, researchers have developed rigorous simulation protocols:

• Simulate Effective Population Size Trajectory: Generate a known demographic scenario, typically with seasonally varying population size to reflect realistic infectious disease dynamics [46].

• Generate Sampling Time Distributions: Create multiple sampling schemes:

  • Functionally independent: Sampling times follow a distribution independent of population size (e.g., uniform distribution)
  • Preferentially sampled: Sampling times depend on effective population size, with more intensive sampling during high-prevalence periods [46]

• Simulate Genealogies: Given the sampling times and population size trajectory, simulate genealogies using coalescent process simulation tools [46].

• Perform Phylodynamic Inference: Apply state-of-the-art phylodynamic methods to the simulated genealogies while incorrectly assuming sampling times are fixed or independent of population size [46].

• Quantify Bias: Compare estimated effective population size trajectories to the known simulated truth, measuring systematic deviations and precision [46].

This protocol demonstrated that ignoring preferential sampling can produce systematically biased effective population size estimates, with the size of bias depending on local properties of the population trajectory [46].

Protocol for Testing Model Misspecification Robustness

To evaluate how simplified models perform when applied to complex epidemiological dynamics:

• Simulate Complex Epidemics: Use an individual-based model with realistic population structure and transmission dynamics, calibrated to actual epidemic data (e.g., men who have sex with men in San Diego, USA) [6].

• Generate Evolutionary Histories: Simulate genealogies and sequence evolution along the simulated epidemic trajectory, creating alignments equivalent to specific genetic regions (e.g., HIV partial pol gene and complete genome) [6].

• Apply Simplified Inference Models: Analyze the simulated data using simpler phylodynamic models that provide simplistic representations of the true epidemiological process [6].

• Compare Estimates to Known Truth: Quantify inductive bias by comparing estimated parameters (e.g., migration rates) to their known values from the simulation [6].

• Evaluate Sample Size Effects: Repeat analyses with different sample sizes (e.g., 100, 500, 1000 sequences) to determine how bias changes with data quantity [6].

This approach revealed that even misspecified models could recover certain parameters, with estimation accuracy of migration rates depending on both method and sample size [6].

Figure 1: Experimental workflow for evaluating sampling bias effects on phylodynamic inference.

Methodological Approaches to Overcome Sampling Biases

Preferential Sampling Modeling

The most direct approach to address sampling bias involves explicitly modeling the sampling process itself. Rather than treating sampling times as fixed or independent of population dynamics, preferential sampling models:

• Formulate sampling times as an inhomogeneous Poisson process with intensity functionally dependent on effective population size [46]
• Jointly estimate population and sampling processes using Bayesian inference with Gaussian process priors [46]
• Integrate this sampling model into nonparametric phylodynamic frameworks (e.g., Gaussian Markov random fields, Gaussian processes) [46]

Application of this method to seasonal influenza data demonstrated large improvements in precision over sampling-unaware methods, with varying strengths of preferential sampling detected across geographic regions [46]. The approach successfully eliminated systematic bias while providing more precise estimates of effective population size trajectories.

Integrated Genomic and Epidemiological Data Analysis

Another powerful strategy combines genomic data with traditional epidemiological surveillance:

• Develop joint models that analyze both time-stamped pathogen genomes and time series of case counts [47]
• Use approximations to make joint inference computationally feasible for large outbreaks [47]
• Leverage the complementary strengths of both data types: sequences provide resolution on transmission links, while case counts inform on prevalence dynamics [47]

The Timtam package implements this approach within the BEAST2 framework, enabling estimation of both effective reproduction numbers and prevalence trajectories while accounting for the fact that typically only a small fraction of cases are sequenced [47]. In empirical applications to SARS-CoV-2 and poliomyelitis outbreaks, this integrated approach produced estimates consistent with previous analyses while providing novel insights into prevalence dynamics [47].

Model-Based Correction with Informed Priors

Bayesian phylodynamic methods offer natural mechanisms to incorporate prior knowledge about sampling processes:

• Specify informed prior distributions for sampling parameters based on surveillance metadata
• Use Bayesian model averaging to account for uncertainty in sampling schemes
• Implement mixture models that accommodate multiple sampling strategies simultaneously

These approaches are particularly valuable when sampling intensity varies systematically across regions or time periods, allowing researchers to formally incorporate knowledge about surveillance heterogeneity into their analyses.

Table 2: Key Computational Tools and Databases for Phylodynamic Research

Tool/Resource	Type	Primary Function	Application Context
BEAST2	Software Package	Bayesian evolutionary analysis sampling trees; implements coalescent, birth-death models, and phylogeography [10]	General phylodynamic inference; accommodates multiple evolutionary and population genetic models
Timtam	BEAST2 Package	Approximate likelihood combining phylogenetic information and epidemiological time series [47]	Joint analysis of sequenced and unsequenced case data; estimation of prevalence and reproduction number
GalaxyTrakr	Bioinformatics Platform	User-friendly interface for quality control, assembly, and annotation of genomic data [48]	Foodborne pathogen surveillance; whole genome sequence analysis
Pathogen Detection Website	Database	FDA's real-time pathogen identification and tracking system [48]	Outbreak investigation; comparison of clinical, food, and environmental isolates
SplitsTree	Software Tool	Network visualization of phylogenetic relationships; detection of recombination events [49]	Assessment of phylogenetic uncertainty; identification of conflicting signals
INLA	Statistical Method	Integrated nested Laplace approximation for efficient Bayesian inference [46]	Approximation of coalescent likelihoods; simulation studies of sampling bias

The comparative analysis of phylodynamic methods reveals significant differences in their robustness to sampling bias and heterogeneous surveillance. Methods that explicitly model the sampling process, such as preferential sampling corrections, or that integrate multiple data types, like Timtam, generally provide more reliable inference under realistic sampling scenarios [46] [47]. However, computational constraints remain a significant challenge, particularly for large datasets or complex models [6].

Future methodological development should focus on improving computational efficiency, developing more flexible models of sampling processes, and creating better diagnostic tools for detecting sampling biases in empirical datasets. Additionally, greater integration between traditional epidemiological approaches and genomic methods will likely yield more robust frameworks for outbreak source attribution. As sequencing technologies continue to become more accessible and affordable, addressing these methodological challenges will be crucial for maximizing the public health impact of phylodynamic approaches.

Addressing Computational Bottlenecks and Ensuring Scalability

A critical challenge in modern phylodynamics, the field that combines epidemiological dynamics with phylogenetic evolutionary analysis, is developing methods that are both accurate and computationally scalable for outbreak source attribution. This guide compares the performance of three distinct approaches—the mechanistic Bayesian framework of ScITree, the deep learning-based PhyloDeep, and the established BEAST2 software ecosystem—highlighting how they address the inherent computational bottlenecks in the field.

Breaking the Bottleneck: A Comparison of Scalable Phylodynamic Methods

The table below summarizes the core characteristics, performance, and ideal use cases for the three methods compared in this guide.

Method	Core Approach	Computational Scaling	Key Innovation	Inferential Target	Ideal Use Case
ScITree [50] [35]	Bayesian Mechanistic Model	Linear with outbreak size	Infinite sites assumption for mutations	Transmission tree, epidemiological parameters	Accurate and scalable inference of who-infected-whom in large outbreaks.
PhyloDeep [51]	Deep Learning (Simulation-based)	Fast inference after training	Compact Bijective Ladderized Vector (CBLV) for tree representation	Epidemiological parameters (e.g., R0), model selection	Ultra-fast parameter estimation and model selection from very large phylogenies.
BEAST2 [52] [53]	Bayesian Evolutionary Analysis	Can be computationally intensive [51]	Cohesive ecosystem (MCMC, TreeAnnotator) [53]	Time-calibrated phylogenies, evolutionary rates	Detailed, model-flexible evolutionary analysis where time is less constrained.

A key bottleneck in previous methods was the explicit, nucleotide-level modeling of pathogen mutation, which created a massive parameter space. ScITree overcomes this by adopting an infinite sites assumption, modeling mutations as accumulating over time via a Poisson process rather than at each individual base pair [35]. This shifts computational scaling from exponential to linear with outbreak size, enabling full Bayesian inference for larger outbreaks [50] [35].

PhyloDeep bypasses traditional likelihood calculations entirely. It is a likelihood-free, simulation-based method that uses deep neural networks trained on millions of simulated trees. It employs either a large set of summary statistics or a novel Compact Bijective Ladderized Vector (CBLV)—a complete, compact representation of a phylogenetic tree that enables efficient learning [51].

Experimental Protocols and Performance Benchmarks

ScITree Validation Workflow

The performance of ScITree was rigorously tested using a standard protocol for validating phylodynamic methods [35]:

• Simulation: Multiple outbreaks are simulated using a known spatial transmission model (e.g., an SEIR framework) and evolutionary process.
• Inference: The ScITree model is applied to the simulated data to infer the transmission tree and key parameters like the reproductive number.
• Comparison: Inferred transmission trees are compared to the known "ground truth" from the simulation. Accuracy is measured by the proportion of correct transmission links identified. Computational time is tracked relative to outbreak size.
• Real-Data Application: The validated model is applied to a real-world dataset, such as the 2001 Foot-and-Mouth Disease outbreak in the UK, to demonstrate practical utility and compare results with previous methods [35].

Experimental data demonstrates that ScITree achieves inference accuracy comparable to the previous Lau method while offering a dramatic improvement in scalability. The Lau method's computing time increases exponentially with outbreak size, while ScITree's increases only linearly, making it feasible for larger outbreaks [50] [35].

PhyloDeep Training and Inference

The experimental approach for PhyloDeep involves a pre-training phase [51]:

• Data Generation: Millions of phylogenetic trees are simulated under a broad range of parameter values for specific birth-death models (e.g., BD, BDEI, BDSS).
• Representation: Each simulated tree is converted into a numerical vector using either the 83 summary statistics or the CBLV representation.
• Training: Neural networks (Feed-Forward for summary statistics, Convolutional for CBLV) are trained on these vectors to learn the mapping from tree features to epidemiological parameters (regression) or model type (classification).
• Inference: A user's empirical phylogenetic tree is fed into the trained network, which rapidly outputs parameter estimates and a model prediction.

Benchmarks on simulated data show that PhyloDeep provides accuracy comparable to or better than BEAST2, but is significantly faster, especially on very large trees with thousands of tips [51].

The Scientist's Toolkit: Essential Research Reagents and Software

The table below lists key software tools and their functions in phylodynamic analysis.

Tool Name	Type	Primary Function
BEAST2 [52]	Software Package	Bayesian evolutionary analysis using MCMC to infer time-calibrated phylogenetic trees.
TreeAnnotator [53]	Analysis Tool	Summarizes the posterior sample of trees from BEAST2 into a single Maximum Clade Credibility (MCC) tree.
ggtree [52] [54]	R Package	A highly customizable and programmable R package for visualizing and annotating phylogenetic trees.
ScITree [35]	R Package	A Bayesian phylodynamic model for scalable inference of the transmission tree from outbreak data.
PhyloDeep [51]	Software Tool	A deep learning-based tool for fast likelihood-free estimation of parameters and model selection from phylogenies.

Relationships and Workflows in Phylodynamic Methods

The following diagram illustrates the logical relationships and fundamental differences in approach between the phylodynamic methods discussed.

Navigating Method Selection

The choice of a phylodynamic method involves a strategic trade-off between computational speed, inferential target, and model flexibility.

• For direct transmission tree inference in large outbreaks, where understanding "who-infected-whom" is the primary goal, ScITree offers a powerful balance of mechanistic accuracy and computational feasibility [50] [35].
• For rapid assessment of very large phylogenies to estimate key parameters like the reproductive number or to identify the best-fitting model, PhyloDeep provides unparalleled speed without sacrificing accuracy [51].
• For detailed evolutionary analysis requiring fine-grained inference of substitution rates, complex evolutionary models, and precise dating of ancestral nodes, the BEAST2 ecosystem remains a comprehensive and gold-standard solution, though often at a higher computational cost [52] [51] [53].

The ongoing innovation in both model-aware algorithms and simulation-based deep learning is decisively addressing the computational bottlenecks that have historically constrained phylodynamics, empowering researchers to fully leverage the richness of modern genetic data for effective outbreak response.

Detecting and Correcting for Model Misspecification and Inductive Bias

In the field of infectious disease epidemiology, phylodynamic methods have become fundamental building blocks for understanding outbreak characteristics by integrating pathogen genetic sequences with epidemiological data [49]. These methods enable researchers to infer key aspects of disease outbreaks, including geographic origins, dispersal routes, and the number of transmission events between populations [55]. However, the robustness and statistical efficiency of these models can be compromised when the underlying assumptions do not adequately represent the complex reality of epidemiological and evolutionary processes [6] [37].

The problem of model misspecification occurs when analytical models provide an overly simplistic representation of the evolutionary system, leading to inductive bias in parameter estimation [6] [37]. This challenge is particularly acute in phylodynamic inference for infectious diseases where time scales are short but epidemiological dynamics, driven by human behavior and pathogen biology, are potentially quite complex [37]. The consequences of misspecification can distort central conclusions of epidemiological studies, affecting estimates of ancestral outbreak origins, dispersal routes, and the number of transmission events between populations [55].

Experimental Evidence of Misspecification Effects

Quantitative Evidence from Simulation Studies

Table 1: Experimental Findings on Model Misspecification in Phylodynamic Inference

Study Focus	Misspecification Scenario	Key Finding	Impact on Inference
HIV Migration Rates [6] [37]	Complex epidemic trajectory vs. simple structured coalescent	Inductive bias occurred with model misspecification	Bias was small with sample size ≥1000 sequences; higher migration rates estimated more accurately
Spatial Epidemiology [55]	Unrealistic prior assumptions in discrete geographic models	≈93% of surveyed studies used biologically unrealistic priors	Distorted conclusions about relative dispersal rates, route importance, and ancestral origins
Avian Coronavirus Spread [56]	Structured coalescent with and without "ghost" deme	Different compartmentalization patterns emerged	Revealed undocumented transmission routes via unsampled populations
TB Transmission Clusters [20]	SNP cut-offs vs. phylodynamic transmission inference	Phylodynamic approaches revealed overlooked transmission	4-SNP cut-off captured 98% of inferred transmission events

Methodological Protocols for Misspecification Testing

Researchers have developed rigorous experimental protocols to evaluate the impact of model misspecification. A comprehensive approach involves:

• Complex Model Simulation: Developing sophisticated epidemiological models calibrated to real-world data. For example, one study used a structured compartmental model with 120 ordinary differential equations for HIV transmission, incorporating five stages of infection, four age groups, three diagnosis stages, and two risk groups [37].

• Data Generation: Using the complex model to simulate genealogies and genetic sequence alignments equivalent to both partial genes and complete pathogen genomes [6] [37].

• Simplified Model Inference: Applying simplified phylodynamic models (reflecting current standard practice) to estimate parameters from the simulated data [37].

• Bias Quantification: Comparing parameter estimates from simplified models against known values from the complex simulation to quantify inductive bias [6] [37].

This approach allows researchers to test how well standard methods perform when confronted with data generated from more complex, realistic systems, thereby evaluating the real-world applicability of these methods.

Comparative Analysis of Phylodynamic Methods

Performance Under Misspecification

Table 2: Method Comparison for Handling Model Misspecification

Method Category	Key Assumptions	Strengths	Vulnerabilities to Misspecification
Structured Coalescent [37]	Constant population sizes and migration rates	Computational efficiency; well-established theoretical foundation	Struggles with nonlinear dynamics; sensitive to sampling schemes
Structured Birth-Death Models [37]	Exponential growth or specified population dynamics	Accommodates changing population sizes; intuitive parameters	Misspecification of growth model distorts divergence time estimates
Discrete-Trait Phylogeographic [37] [55]	Migration as continuous-time Markov chain	Flexible for geographic inference; widely implemented	Highly sensitive to prior specifications; inflated false-positive dispersal routes
Model-Based Phylodynamics [37]	Connection between genealogy and population dynamics	Accommodates nonlinear dynamics and time-varying rates	Requires correct specification of population dynamics model

Correction Strategies and Their Efficacy

Several strategies have emerged to detect and correct for model misspecification:

• Prior Specification Adjustments: Research has demonstrated that default priors in popular software packages like BEAST make strong and biologically unrealistic assumptions [55]. Developing more biologically reasonable priors significantly improves inference accuracy, particularly for discrete geographic models.

• Structured Coalescent with Ghost Populations: Incorporating unsampled populations ("ghost demes") in structured coalescent models can reveal undocumented transmission routes and provide more accurate estimates of migration rates between sampled populations [56].

• Joint Inference Frameworks: Approaches like the EpiFusion framework integrate phylogenetic and epidemiological data within a unified particle filtering framework, reducing misspecification errors by jointly modeling both data types [22].

• Sample Size Considerations: Evidence suggests that increasing sample size to ≥1000 sequences can mitigate some biases introduced by model misspecification, though it does not eliminate them entirely [6] [37].

Visualization of Methodological Workflows

Phylodynamic Model Validation Workflow

Structured Coalescent with Ghost Populations

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Analytical Tools for Misspecification Research

Tool/Software	Primary Function	Application in Misspecification Research	Implementation Considerations
BEAST [55] [10]	Bayesian evolutionary analysis	Testing discrete phylogeographic models; evaluating prior sensitivity	Default priors may introduce bias; requires careful specification
phydynR [37]	Model-based phylodynamics	Implementing structured coalescent models with population dynamics	Accommodates nonlinear dynamics; scalable for large datasets
EpiFusion [22]	Joint phylogenetic-epidemiological inference	Dual observation model for case incidence and phylogenetic data	Particle filtering approach; reduces integration bias
phybreak [20]	Transmission tree inference	Alternative to contact tracing for transmission clusters	Suitable for low-incidence settings; does not impute single unobserved cases
SplitsTree [49]	Phylogenetic network analysis	Detecting recombination events that complicate phylogenetic inference	Handles network structures; distinguishes recombination from uncertainty

The evidence consistently demonstrates that model misspecification presents a substantial challenge in phylodynamic inference, potentially distorting estimates of key epidemiological parameters. However, methodological advances offer promising approaches for detection and correction.

Based on current experimental findings, researchers should:

• Validate model assumptions using simulation studies calibrated to real-world systems before applying methods to empirical data [6] [37].
• Carefully specify prior distributions rather than relying on software defaults, particularly for discrete geographic models [55].
• Consider structured coalescent approaches that incorporate ghost populations when working with partially sampled epidemics [56].
• Utilize joint inference frameworks that integrate multiple data types to reduce reliance on any single model structure [22].
• Report sensitivity analyses that quantify how inferences change under different model specifications and prior assumptions [55].

As the field continues to develop, researchers must maintain critical awareness of the limitations of their analytical frameworks and actively employ strategies to detect and correct for model misspecification in outbreak source attribution research.

Best Practices for Parameterization and Interpreting Model Outputs

Phylodynamics, a term coined by Grenfell et al., represents the melding of evolutionary biology, immunodynamics, and epidemiology to infer population dynamics of pathogens from genetic sequence data [10]. This field is grounded in the principle that epidemiological processes leave recognizable signatures in pathogen genomes, which can be decoded through phylogenetic analysis combined with mathematical models [57] [10]. For researchers focused on outbreak source attribution, phylodynamic methods provide powerful tools for reconstructing transmission chains, estimating key epidemiological parameters, and understanding spatial spread patterns. The robustness of these inferences, however, depends critically on appropriate model specification, parameterization, and interpretation—particularly when dealing with the complex scenarios typical of real-world outbreaks.

The fundamental assumption underlying phylodynamics is that epidemiological and evolutionary processes occur on similar timescales for rapidly evolving pathogens [10]. This enables researchers to use time-scaled phylogenies to estimate parameters such as the basic reproduction number (R0), effective population size (Ne), and migration rates between populations [57] [10]. Two primary classes of statistical models dominate the field: coalescent-based approaches, which model the process of lineage merging backward in time, and birth-death models, which forward-model transmission and sampling events [57] [10]. Understanding the strengths, limitations, and appropriate application contexts for these frameworks is essential for reliable source attribution research.

Comparative Framework for Phylodynamic Methods

Methodological Approaches and Theoretical Foundations

Phylodynamic methods can be broadly categorized into three methodological paradigms, each with distinct theoretical foundations, data requirements, and computational characteristics. The choice between these approaches involves important trade-offs between statistical efficiency, biological realism, and computational tractability that researchers must navigate based on their specific outbreak investigation goals.

Table 1: Comparison of Primary Phylodynamic Methodological Frameworks

Method	Theoretical Basis	Key Parameters	Strengths	Limitations
Fully Bayesian (e.g., BEAST2)	Bayesian evolutionary analysis with MCMC sampling	R0, evolutionary rates, population sizes, migration rates	Naturally integrates uncertainty in all parameters; flexible model specification [58]	Computationally intensive for large datasets (>500 sequences) [6] [58]
Birth-Death Sampling Models	Forward-time transmission process with sampling	R0, becoming uninfectious rate, sampling proportion	Direct epidemiological interpretation; models continuous sampling [59]	Sensitive to model misspecification; date data can dominate inference [59]
Hybrid Approaches	Maximum likelihood tree estimation + Bayesian parameter inference	Evolutionary rates, growth rates, population sizes	Computational efficiency for large datasets; maintains temporal structure [58]	Ignores uncertainty in tree estimation; requires clocklike behavior [58]

The fully Bayesian approach, implemented in software packages like BEAST2, enables simultaneous inference of phylogenetic trees, evolutionary timescales, and epidemiological parameters within a unified probabilistic framework [58]. This method excels at naturally integrating uncertainty across all model components, but becomes computationally prohibitive for datasets exceeding several hundred sequences [6] [58]. In contrast, hybrid approaches that combine maximum likelihood tree estimation with subsequent Bayesian parameter inference can achieve comparable accuracy while dramatically reducing computational burdens, making them particularly valuable for outbreak investigations involving large numbers of pathogen genomes [58].

Performance Comparison Across Methodologies

Recent comparative studies have quantitatively evaluated the performance of different phylodynamic approaches under controlled conditions, providing evidence-based guidance for method selection. These comparisons have examined accuracy, precision, and computational efficiency across a range of outbreak scenarios and dataset characteristics.

Table 2: Experimental Performance Comparison of Phylodynamic Methods

Study	Comparison	Dataset	Key Findings	Implications for Source Attribution
BMC Evol Biol (2018) [58]	Fully Bayesian vs. Hybrid	Bacterial WGS datasets (63-329 samples)	Estimates between methods were very similar when temporal structure was strong	Hybrid methods valid for large datasets with clocklike behavior; reduces compute time from weeks to days
Mol Biol Evol (2023) [59]	Date vs. Sequence data influence	600 simulated outbreaks (500 cases each)	62% of analyses were date-driven; sequence data more informative with high evolutionary rates	Sampling times critical for R0 estimation; genome sequence value increases with substitution rate
PLoS Comput Biol (2017) [60]	Regression-ABC vs. Likelihood-based	Simulated trees under SIR model	Comparable accuracy for large phylogenies; superior for host population size estimation	Machine learning approach avoids likelihood computation; useful for complex models
Sci Rep (2025) [20]	Phybreak for transmission inference	2,008 M. tuberculosis genomes	SNP cut-off of 4 captured 98% of inferred transmissions	Phylodynamics provides alternative to contact tracing for cluster definition

The 2018 comparative study of bacterial genomic data analysis revealed that hybrid methods produced highly congruent parameter estimates compared to fully Bayesian approaches when applied to data with strong temporal structure [58]. This finding was particularly pronounced for evolutionary rate estimates, where the 95% credible intervals from BEAST2 and confidence intervals from least-squares dating (LSD) showed substantial overlap across multiple bacterial pathogens [58]. The practical implication is significant: for outbreak investigations involving hundreds of genomes, hybrid approaches can reduce computation time from weeks to days while maintaining analytical rigor, enabling more rapid public health response.

Research published in 2023 introduced a novel framework for quantifying the relative contributions of sampling dates versus sequence data to phylodynamic inference [59]. Through analysis of 600 simulated outbreaks, the study demonstrated that approximately 62% of analyses were predominantly driven by date information, with sequence data becoming more influential only at higher evolutionary rates (10⁻³ substitutions/site/time) [59]. This finding has crucial implications for study design: careful documentation of sampling dates is essential, while the marginal value of additional sequences may diminish once certain thresholds are exceeded, depending on the pathogen's evolutionary rate.

Experimental Protocols and Methodologies

Protocol for Fully Bayesian Phylodynamic Inference

The fully Bayesian approach implemented in BEAST2 remains the gold standard for phylodynamic inference when computational resources allow. The following protocol outlines the key steps for proper parameterization and implementation:

• Data Preparation: Compile pathogen sequences with exact collection dates. For best results, use whole genomes or sufficiently informative genetic regions (e.g., complete HIV genome versus partial pol gene) [6]. Ensure temporal structure in the data through date-randomization tests [58].

• Model Specification: Select appropriate substitution models (e.g., GTR+Γ+I), molecular clock models (strict vs. relaxed), and tree priors (coalescent vs. birth-death) based on dataset characteristics. Use Bayesian model testing to compare alternatives [58].

• Parameterization: Set informed priors for key parameters rather than relying on default settings. For birth-death models, specify informed sampling proportions and becoming-uninfectious rates based on epidemiological data [59] [58].

• MCMC Execution: Run multiple independent Markov Chain Monte Carlo chains with sufficient length to achieve convergence (effective sample size >200 for all parameters). Use chain combining after confirming stationarity [58].

• Output Interpretation: Analyze posterior distributions of parameters of interest (R0, growth rates, migration rates). Use Bayesian model averaging when uncertainty exists between competing models [58].

This protocol was applied in a study of HIV transmission dynamics, which demonstrated that simple structured coalescent models could recover migration rates even when adjusting for nonlinear epidemiological dynamics, though some inductive bias occurred with model misspecification [6].

Protocol for Hybrid Phylodynamic Analysis

For larger datasets (>500 sequences) where fully Bayesian analysis becomes computationally prohibitive, the hybrid approach provides a viable alternative:

• Phylogram Estimation: Infer maximum likelihood phylogenies using software such as PhyML or RAxML under an appropriate substitution model. Use non-parametric bootstrapping to assess topological uncertainty [58].

• Molecular Clock Dating: Convert phylograms to time-scaled chronograms using least-squares dating (LSD) methods, which assume a strict molecular clock but provide computational efficiency comparable to Bayesian methods [58].

• Phylodynamic Inference: Analyze the fixed chronograms using Bayesian inference in BEAST2 or RevBayes to estimate demographic parameters, using the tree as fixed input rather than a inferred parameter [58].

• Validation: Confirm clocklike behavior in the data through comparison with Bayesian date-randomization tests. For the S. aureus ST239 dataset, this approach yielded root age estimates of 1945 (LSD) versus 1958 (BEAST2), demonstrating temporal congruence [58].

This methodology was successfully applied to analyze large genomic datasets of Shigella dysenteriae type 1 (n=329), which would have been computationally intractable using fully Bayesian approaches [58].

Workflow Visualization

This workflow outlines the key decision points in selecting an appropriate phylodynamic methodology based on dataset characteristics and research objectives. The pathway diverges based on dataset size and computational constraints, with the fully Bayesian approach recommended for smaller datasets (<500 sequences) where computational resources allow comprehensive uncertainty integration, while hybrid methods provide a viable alternative for larger outbreaks where computational efficiency is paramount [6] [58].

Parameterization Best Practices

Managing Model Misspecification and Inductive Bias

Model misspecification represents a significant challenge in phylodynamic inference, potentially introducing inductive bias that distorts parameter estimates. A 2025 study of HIV transmission dynamics demonstrated that even simple structured coalescent models could recover migration rates when adjusting for nonlinear epidemiological dynamics, but noted that inductive bias could occur if the model provided an overly simplistic representation of the evolutionary process [6]. The study found this bias was minimal with sample sizes ≥1000 sequences, suggesting that larger genomic datasets provide some robustness against model misspecification [6].

To mitigate inductive bias, researchers should:

• Test model sensitivity by comparing estimates across alternative model specifications
• Use simulation-based calibration to verify model adequacy before analyzing real data
• Incorporate known epidemiological parameters as informed priors rather than relying solely on sequence data
• Validate against independent data sources such as incidence records or contact tracing data

The HIV phylodynamic study further demonstrated that estimation of higher migration rates was more accurate than for lower migration rates, highlighting the importance of considering parameter-specific performance when interpreting results [6].

Accounting for Reporting Delays and Sampling Biases

Real-time phylodynamic analyses must contend with reporting delays between sample collection and sequence availability, which can severely impact parameter estimates near the present time. A 2025 method proposed incorporating reporting delay distributions into the sampling model to mitigate these effects [15]. This approach uses historically observed times between sampling and reporting for a population of interest to account for missing samples in recent time periods.

Key considerations for addressing sampling biases include:

• Modeling preferential sampling when sampling intensity correlates with disease prevalence
• Incorporating lineage-specific reporting delays when different variants have different sequencing priorities
• Using multi-type birth-death models to account for heterogeneous sampling across populations
• Integrating incidence data through joint models like EpiFusion, which combines case incidence and phylogenetic trees within a particle filtering framework [22]

The EpiFusion framework exemplifies the trend toward integrating multiple data sources, using a "single process model, dual observation model" structure that simulates outbreak trajectories evaluated against both phylodynamic and epidemiological data [22].

Interpretation of Model Outputs

Connecting Genetic Parameters to Epidemiological Quantities

Interpreting phylodynamic output requires careful translation of genetic parameters into epidemiological quantities with appropriate consideration of underlying assumptions. The effective population size (Ne) estimated from genetic data represents genetic diversity rather than the absolute number of infected individuals, though these quantities are often correlated under stable demographic conditions [15]. Similarly, growth rates estimated from phylogenies reflect the expansion of genetic diversity, which may lag behind epidemic growth depending on the proportion of cases sampled.

When interpreting phylodynamic estimates of R0, researchers should consider:

• Generation time distribution assumptions, which strongly influence R0 estimates
• Sampling proportion across the outbreak, as undersampling can bias estimates
• Population structure that may violate model assumptions of homogeneity
• Time-varying reproduction numbers that may be averaged in the phylogeny

For source attribution studies, phylogeographic methods can estimate migration rates between locations, but these inferences are sensitive to sampling heterogeneity across regions [57]. Discrete trait analysis (DTA) offers computational efficiency for reconstructing spatial spread, while structured birth-death models provide more epidemiologically interpretable parameters at greater computational cost [2].

Managing Computational and Statistical Constraints

Computational limitations present practical constraints on phylodynamic analysis that impact interpretation. A study of HIV phylodynamics found that phylogeographic models in BEAST were not scalable for datasets of 600 or more sequences, necessitating alternative approaches for larger outbreaks [6]. Similarly, the fully Bayesian analysis of a Shigella dysenteriae dataset (n=329) required substantial computational resources, making hybrid approaches preferable for routine application [58].

Statistical power in phylodynamic inference depends on multiple factors:

• Evolutionary rate determines the number of informative sites; higher rates provide more signal
• Sample size impacts precision, with diminishing returns beyond certain thresholds
• Temporal spacing of samples affects clock calibration; evenly distributed samples improve dating
• Population structure may require more complex models with additional parameters

Researchers should report effective sample sizes for MCMC analyses and convergence diagnostics to ensure statistical reliability. For hybrid approaches, confidence intervals from maximum likelihood estimation should be complemented with sensitivity analyses to tree uncertainty.

Essential Research Reagents and Computational Tools

Successful implementation of phylodynamic methods requires familiarity with both conceptual frameworks and practical computational tools. The following table summarizes key software solutions and their applications in outbreak source attribution research.

Table 3: Research Reagent Solutions for Phylodynamic Analysis

Tool/Software	Primary Function	Application Context	Implementation Considerations
BEAST2 [58] [60]	Bayesian evolutionary analysis	Fully Bayesian phylodynamic inference; integrates tree and parameter uncertainty	Computationally intensive; requires MCMC diagnostics; appropriate for datasets <500 sequences
EpiFusion [22]	Joint inference from incidence and genetic data	Particle filtering framework combining case data and phylogenies	Java-based command line tool; uses XML input files; available via GitHub repository
LSD (Least Squares Dating) [58]	Molecular clock dating	Rapid estimation of evolutionary timescales from phylogenies	Assumes strict clock; computational efficient for large trees; validated against Bayesian methods
PhyML [58]	Maximum likelihood tree estimation	Phylogenetic tree inference under substitution models	Fast compared to Bayesian tree search; enables bootstrap support values
phybreak [20]	Transmission inference	Reconstruction of transmission trees from genetic data	Does not impute unobserved cases; suitable for low-incidence settings with imported cases
Regression-ABC [60]	Approximate Bayesian Computation	Likelihood-free inference for complex models	Uses machine learning (LASSO) for summary statistic selection; comparable accuracy to likelihood methods for large trees

These tools represent the evolving landscape of phylodynamic software, with ongoing developments focused on improving computational efficiency, model flexibility, and integration of diverse data sources. The trend toward hybrid approaches that combine the strengths of different methodological paradigms reflects the field's response to the challenges posed by increasingly large genomic datasets collected during outbreaks.

Phylodynamic methods have transformed our ability to reconstruct outbreak dynamics from pathogen genetic sequences, providing powerful approaches for source attribution research. The comparative analysis presented here demonstrates that method selection involves fundamental trade-offs between statistical efficiency, biological realism, and computational tractability. Fully Bayesian approaches remain the gold standard for smaller datasets where computational resources allow comprehensive uncertainty quantification, while hybrid methods offer practical alternatives for larger outbreaks.

Robust parameterization requires careful attention to model specification, with particular consideration of sampling biases, reporting delays, and potential misspecification. Interpretation of results must acknowledge the fundamental connection between genetic parameters and epidemiological quantities, recognizing that estimates represent inferences rather than direct observations. As the field continues to evolve, integration of multiple data sources through frameworks like EpiFusion and development of efficient approximate methods will further enhance our capacity to unravel transmission dynamics from genetic data.

For researchers embarking on outbreak source attribution studies, the best practices outlined here provide a foundation for implementing phylodynamic methods that balance analytical rigor with practical constraints. By selecting appropriate methodologies based on dataset characteristics and research questions, carefully parameterizing models to reflect biological reality, and interpreting outputs with appropriate caution, scientists can maximize the insights gained from pathogen genomic data to inform public health response.

Benchmarking Phylodynamic Inference: Accuracy, Validation, and Comparative Performance

This guide objectively compares the performance of modern phylodynamic methods, focusing on their validation through simulated outbreaks and ground-truth comparisons. This approach is crucial for verifying the accuracy of models in reconstructing transmission trees, estimating key epidemiological parameters, and ultimately building confidence in their application for outbreak source attribution.

Performance Comparison of Phylodynamic Methods

The table below summarizes the quantitative performance and validation frameworks of several phylodynamic methods as reported in the scientific literature.

Table 1: Comparative Performance of Phylodynamic Methods in Outbreak Reconstruction

Method Name	Core Approach	Validation Framework (Simulated Outbreaks)	Key Performance Metrics	Comparative Performance
ScITree [61]	Scalable Bayesian mechanistic model; uses infinite sites assumption for mutations.	Assessed using multiple simulated outbreak datasets.	Inference accuracy of transmission tree & parameters; computational time; scalability.	Achieved accuracy comparable to the Lau method. Computing time scales linearly with outbreak size, a significant improvement.
Lau Method [61]	Full Bayesian mechanistic model; models mutation explicitly at nucleotide level.	Used as a benchmark in ScITree evaluation due to its high accuracy.	Accuracy in estimating joint epidemiological-evolutionary dynamics and transmission tree.	High accuracy but faces major computational bottlenecks; computing time scales exponentially with outbreak size.
Nanopore SNP Polishing + Birth-Death Models [62]	Random forest classifiers for polishing nanopore SNP calls; phylodynamic inference with birth-death skyline models.	Validation of SNP calls against Illumina references; phylodynamic inference on two real MRSA outbreaks.	SNP call accuracy/precision; recall; inference of phylogenetic topology and origin.	Reproduced phylogenetic topology and outbreak origin; enabled phylodynamic inference from low-coverage nanopore data.
EpiFusion [22]	Joint inference from case incidence and phylogenetic trees via particle filtering and MCMC.	Tested on both simulated and real outbreak datasets to infer effective reproduction number (Rt).	Accuracy in estimating Rt and infection trajectories.	Validated as a tool for joint inference, providing a framework to integrate different data types.
DeepDynaForecast [63]	Phylogenetic-informed graph deep learning for forecasting transmission dynamics.	Trained and tested on simulated outbreak data; applied to empirical HIV data.	Accuracy in classifying transmission dynamics (growth/static/decline).	Achieved 91.6% accuracy in classifying dynamics on simulated data; demonstrated utility on real HIV data.

Detailed Experimental Protocols for Method Validation

A critical component of phylodynamic research is the use of robust experimental protocols to validate methods before their application to real-world data.

Protocol for Validation via Simulated Outbreaks

The following workflow, utilized by methods like ScITree and DeepDynaForecast, outlines the standard process for validating a phylodynamic model using simulations [61] [63].

Step 1: Define Ground-Truth Parameters: The process begins by defining the complete, known parameters of a simulated outbreak. This includes the reproductive number (R), the transmission tree (who-infected-whom), and epidemiological rates (e.g., incubation and infectious periods). For the evolutionary process, parameters like the mutation rate and substitution model are specified [61].

Step 2: Simulate the Epidemiological Process: Using the ground-truth parameters, a stochastic epidemiological process is simulated. This often employs a continuous-time SEIR framework, where individuals transition from Susceptible to Exposed to Infectious to Removed. The force of infection from an infectious individual i to a susceptible individual j is typically modeled with a spatial kernel function, such as an exponentially decaying rate ( \beta e^{-\kappa d{ij}} ), where ( d{ij} ) is the distance between them [61].

Step 3: Simulate Genetic Evolution: Alongside the epidemiological process, the genetic evolution of the pathogen is simulated along the branches of the known transmission tree. Different methods make different assumptions at this stage. The Lau method simulates mutations explicitly at the nucleotide level, while ScITree adopts the infinite sites assumption, modeling mutations as a Poisson process accumulating within an individual [61].

Step 4: Generate Simulated Datasets: The output of the simulations is a synthetic dataset that mimics what researchers would obtain from a real outbreak. This includes the sampling times and genetic sequences of a subset of infected individuals, representing the observed data, while the complete transmission tree and other parameters are retained as the ground-truth for validation [61].

Step 5: Apply the Phylodynamic Method: The simulated observed data (genetic sequences and sampling times) are fed into the phylodynamic method being validated (e.g., ScITree, EpiFusion). The method then performs inference without any prior knowledge of the ground-truth [61] [22].

Step 6: Reconstruct Parameters and Compare to Ground-Truth: The method's output—including the inferred transmission tree, reproductive number, and other parameters—is systematically compared to the known ground-truth. Key performance metrics include the accuracy of the transmission tree reconstruction (e.g., the proportion of correct transmission links identified) and the coverage of credible intervals for parameter estimates [61].

Protocol for Assessing Intervention Strategies

Phylodynamic methods can also be validated for their utility in assessing public health interventions by using historical outbreaks or simulated scenarios where the outcome is known [64].

Table 2: Research Reagent Solutions for Phylodynamic Validation

Reagent / Tool	Primary Function in Validation	Application Example
Stochastic SEIR Simulators	Generates synthetic outbreak data with known transmission trees and parameters.	Creating ground-truthed datasets for testing method accuracy and scalability [61].
Evolutionary Simulators (e.g., with Infinite Sites)	Simulates genetic sequence evolution along transmission trees under defined models.	Producing realistic pathogen genetic sequences for input into phylodynamic models [61].
Markov Chain Monte Carlo (MCMC)	Bayesian inference algorithm for exploring parameter space and estimating posterior distributions.	Inferring posterior distributions of model parameters (e.g., in ScITree, EpiFusion) from input data [61] [22].
Particle MCMC (pMCMC)	A hybrid algorithm that uses a particle filter (for state variables) within an MCMC framework.	Used in EpiFusion to fit parameters like recovery rate and sampling rate while integrating case incidence data [22].
Tree Pruning & Posterior Predictive Simulation	Computationally modifies phylogenetic trees to test "what-if" intervention scenarios.	Quantifying the hypothetical impact of travel restrictions by removing long-distance viral movements from a tree [64].
Random Forest Classifiers for SNP Polishing	Machine learning model to filter false-positive SNP calls from nanopore sequencing data.	Enabling accurate phylogenetic reconstruction from cost-effective, low-coverage bacterial sequencing [62].

The process involves using a well-characterized outbreak, such as the 2013-2016 West African Ebola virus epidemic, for which a robust phylogenetic tree has been established [64]. Researchers then define a specific hypothetical intervention, such as preventing long-distance viral dispersal or restricting spread to major urban hubs. This intervention is computationally applied to the posterior distribution of phylogenetic trees, for example, by "pruning" branches that represent transmission events that would have been blocked by the intervention. Finally, the impact is quantified by comparing the epidemic size and duration in the pruned trees to the original, full reconstruction, estimating the potential reduction in cases had the intervention been in place [64].

Synthesis of Comparative Findings

The collective evidence from these validation studies reveals critical trade-offs and shared challenges in phylodynamic inference.

Trade-offs Between Accuracy and Scalability

A central finding across studies is the inherent tension between model complexity and computational feasibility. The Lau method sets a high benchmark for accuracy in transmission tree reconstruction by using a complete, nucleotide-level mechanistic model [61]. However, this comes at the cost of exponential scaling of computing time with outbreak size, making it impractical for very large datasets [61]. In contrast, ScITree demonstrates that by incorporating simplifying assumptions like the infinite sites model, it is possible to achieve comparable accuracy while the computing time scales linearly with the outbreak size, offering a deployable solution for larger outbreaks [61].

Performance Under Imperfect Data

Validation frameworks also test methods against real-world data imperfections, such as incomplete sampling. ScITree has been shown to maintain reasonable accuracy in estimating the transmission tree even when not all infected individuals are sampled [61]. Similarly, the application of random forest models to polish nanopore SNP calls demonstrates that phylodynamic inference can be successfully performed with cost-effective, lower-accuracy sequencing data, making it accessible in resource-limited settings [62].

Emerging Directions: Integration and Forecasting

Recent developments focus on integrating diverse data types and moving from reconstruction to forecasting. EpiFusion exemplifies the trend of joint inference, combining traditional case incidence data with phylogenetic trees within a single model to sharpen estimates of the effective reproduction number [22]. Furthermore, methods like DeepDynaForecast leverage deep learning trained on simulated outbreaks to predict future transmission dynamics directly from phylogenetic data, achieving high accuracy in classifying growth trends [63]. This represents a shift from descriptive phylodynamics to a more predictive framework.

In genomic epidemiology, phylogenetic trees reconstructed from pathogen genomes are crucial for understanding the emergence of new variants and tracing transmission dynamics during outbreaks [34]. However, for large-scale analyses, such as those involving millions of SARS-CoV-2 genomes, assessing the confidence and reliability of these trees presents a monumental computational and interpretive challenge [34]. Traditional methods like Felsenstein’s bootstrap, among the most widely used in modern science, require enormous computational capacity and are unsuitable for pandemic-scale datasets [34]. Furthermore, these methods focus on evaluating confidence in clades (groupings of taxa), a topological perspective that is often less relevant for genomic epidemiology than understanding specific evolutionary histories and lineage placements [34]. This guide compares a new efficient method, Subtree Pruning and Regrafting-based Tree Assessment (SPRTA), against established phylogenetic confidence measures, providing a objective analysis of their performance, scalability, and applicability for outbreak source attribution research.

Established Methods for Phylogenetic Confidence Assessment

Before the advent of pandemic-scale genomics, several methods were developed to assign confidence scores to branches of phylogenetic trees. The performance and limitations of these methods are summarized below.

Traditional and Local Support Methods

• Felsenstein’s Bootstrap [34]: The traditional benchmark method, it creates numerous replicate datasets by resampling the original genetic data with replacement. Phylogenetic trees are inferred from each replicate, and the support for a clade is calculated as the proportion of replicate trees containing that clade. Its major drawback is excessive computational demand, which makes it infeasible for massive datasets.
• Local Branch Support Measures [34]: This category includes methods like approximate Likelihood Ratio Test (aLRT) and Bayesian-like transformation of aLRT (aBayes). They are more computationally efficient than full bootstrap as they compare the likelihood of the inferred tree against the likelihoods of similar alternative trees locally around each branch. They share the topological focus of the bootstrap.

The Emerging Solution: SPRTA

Subtree Pruning and Regrafting-based Tree Assessment (SPRTA) is a new approach designed for pandemic-scale phylogenetics [34]. It shifts the paradigm from a topological focus (evaluating clades) to a mutational or placement focus (evaluating evolutionary histories). For a given branch in the tree, SPRTA efficiently approximates the probability that a lineage evolved directly from its proposed ancestor, as opposed to alternative evolutionary origins [34]. It achieves this by evaluating the likelihood of alternative tree topologies generated by relocating a subtree as a descendant of other parts of the tree, a process known as a Subtree Pruning and Regrafting (SPR) move [34].

Table 1: Core Principles and Computational Characteristics of Phylogenetic Confidence Methods.

Method	Principle	Computational Demand	Primary Output
Felsenstein’s Bootstrap [34]	Resampling data to assess clade repeatability	Extremely high; infeasible for millions of sequences	Confidence in clade membership
Local Support (aLRT, aBayes) [34]	Comparing likelihoods of local tree rearrangements	Moderate to high; more efficient than bootstrap	Confidence in local branch topology
SPRTA [34]	Evaluating likelihood of alternative lineage placements via SPR moves	Very low; designed for pandemic scale	Confidence in evolutionary origin of a lineage

Comparative Performance Analysis

A direct comparison of computational demand and application scope reveals the distinct advantages of SPRTA for large-scale phylodynamic studies.

Computational Efficiency and Scalability

Empirical assessments demonstrate that SPRTA reduces runtime and memory demands by at least two orders of magnitude compared to existing methods, including Felsenstein’s bootstrap, transfer bootstrap expectation (TBE), and ultrafast bootstrap approximation (UFBoot) [34]. This performance gap widens as the dataset size increases. While other methods often fail to complete analyses on very large datasets (indicated by premature termination in benchmarks), SPRTA remains feasible [34]. Its efficiency stems from leveraging the SPR search that is already part of the phylogenetic tree search in scalable maximum-likelihood methods like MAPLE and RaxML, requiring minimal additional computation [34].

Table 2: Experimental Performance and Applicability for Genomic Epidemiology.

Method	Scalability (Number of Taxa)	Robustness to Rogue Taxa	Interpretability in Outbreak Studies
Felsenstein’s Bootstrap	Low (Suits smaller datasets)	Low; rogue taxa can substantially lower support throughout the tree [34]	Low; clade-based support is less relevant than lineage placement [34]
Local Support Methods	Moderate	High [34]	Moderate; still topologically focused
SPRTA	Very High (Millions of genomes) [34]	High; placement of uncertain sequences has negligible effect on scores [34]	High; directly assesses confidence in lineage origins and mutational histories [34]

Application in a Real-World Pandemic Context

SPRTA has been successfully applied to investigate a global public SARS-CoV-2 phylogenetic tree comprising more than two million genomes [34]. This analysis highlighted plausible alternative evolutionary origins for many SARS-CoV-2 variants and assessed the reliability of the Pango outbreak lineage classification system [34]. Furthermore, it demonstrated the effect of phylogenetic uncertainty on inferred mutation rates, enabling a detailed probabilistic assessment of transmission and mutational histories at a true pandemic scale [34].

Experimental Protocols for Method Evaluation

To objectively benchmark phylogenetic confidence methods, researchers rely on simulations and defined workflows.

Benchmarking Protocol via Simulation

A standard protocol for evaluating methods like SPRTA involves simulating genome data (e.g., SARS-CoV-2-like sequences) where the true evolutionary tree and mutational history are known [34]. In this controlled environment, branch support scores from each method are interpreted as estimates of the posterior probability of the mutation events implied by the inferred tree. The accuracy of each method is then determined by how well its calculated support scores correlate with the known correctness of branches and mutations from the simulation ground truth [34].

General Phylodynamic Workflow for Outbreak Research

The broader context of phylodynamic inference for outbreak source attribution involves a multi-stage process, from data collection to tree interpretation. The following workflow diagram outlines the key steps, highlighting where confidence assessment integrates into the pipeline.

Diagram 1: Phylodynamic Analysis Workflow.

Successful phylogenetic and confidence analysis requires a suite of software tools and data resources.

Table 3: Key Research Reagent Solutions for Phylogenetic Confidence Analysis.

Tool/Resource	Type	Primary Function	Relevance to Confidence
MAPLE / RaxML [34]	Software Package	Scalable maximum-likelihood phylogenetic inference	Provides the foundational tree and likelihood calculations used by SPRTA [34]
ggtree [54]	R Package	Visualization and annotation of phylogenetic trees	Enables visualization of confidence scores (e.g., SPRTA values) on tree branches [54]
treeio [54]	R Package	Phylogenetic data input/output	Parses diverse tree file formats and associated data into R for analysis with ggtree and other packages [54]
Aligned Genomic Sequences	Data	Primary input data (e.g., from GISAID, GenBank)	The multiple sequence alignment from which trees and confidence scores are inferred [65]
Simulated Datasets	Data	Benchmarking ground truth	Provides known evolutionary histories to validate and compare confidence methods like SPRTA [34]

The scale of data generated during modern pandemics has rendered traditional phylogenetic confidence measures like Felsenstein's bootstrap computationally impractical. The emergence of SPRTA represents a significant paradigm shift, offering a scalable, efficient, and highly interpretable method for assessing confidence in evolutionary histories. For researchers focused on outbreak source attribution, SPRTA provides a direct probabilistic assessment of lineage origins and mutational pathways, which is more actionable than traditional clade-based support. While local support methods offer a more efficient alternative to the bootstrap, SPRTA stands out as the only method currently capable of providing detailed confidence assessments for phylogenetic trees comprising millions of sequences, thereby enhancing our ability to respond to future pandemics.

Comparative Accuracy of Transmission Tree Reconstruction Across Methods

Reconstructing transmission trees—inferring "who infected whom" in disease outbreaks—is a cornerstone of modern infectious disease epidemiology and public health response. These reconstructions provide critical insights into pathogen spread dynamics, help quantify key parameters like the effective reproductive number (R), and allow for the evaluation of mitigation strategies such as vaccination or non-pharmaceutical interventions [66] [4]. The increasing affordability of pathogen genomic sequencing has spurred the development of numerous computational methods that combine this molecular data with traditional epidemiological information to infer transmission chains.

However, these methods differ substantially in their underlying assumptions, data requirements, statistical frameworks, and computational approaches, leading to variations in their accuracy and applicability. For researchers, scientists, and drug development professionals, selecting an appropriate method is complicated by the lack of direct, standardized comparisons. This guide provides an objective comparison of transmission tree reconstruction methods, summarizing their performance based on published experimental data and benchmarks. It is framed within the broader thesis of advancing phylodynamic methods for outbreak source attribution research, focusing on practical accuracy and implementation considerations.

Methodological Families of Reconstruction Algorithms

Methods that combine genomic and epidemiological data can be categorized into distinct families based on how they handle phylogenetic information and integrate it with the transmission process. A systematic review of the literature defines three primary families [67].

Table 1: Core Methodological Families for Transmission Tree Reconstruction

Method Family	Core Approach	Representative Tools	Key Distinction
Non-Phylogenetic (NPF)	Uses pairwise genetic distances between pathogen sequences, without inferring a phylogenetic tree.	Aldrin 2011 [67]	Does not rely on a pre-estimated or co-estimated phylogeny.
Sequential Phylogenetic (SeqPF)	A two-step process: a phylogenetic tree is first reconstructed, and then a transmission tree is inferred from it.	[Various tools [67]]	Assumes the phylogenetic tree is independent of the transmission process.
Simultaneous Phylogenetic (SimPF)	The phylogenetic tree and transmission tree are inferred simultaneously in an integrated framework.	JUNIPER, BORIS, TransPhylo [66] [67]	Joint inference accounts for the dependency between evolution and transmission.

The following workflow diagram illustrates the conceptual and procedural relationships between these families and the data they utilize.

Comparative Performance Analysis

The accuracy of transmission tree reconstruction is influenced by multiple factors, including the method's ability to model complex biological processes and its computational feasibility for large outbreaks.

Benchmarking on Real and Simulated Outbreaks

A critical assessment of performance comes from benchmarking tools on datasets where the true transmission links are known, either from simulated outbreaks or real-world outbreaks with highly reliable contact tracing. Recent benchmarks highlight key trade-offs.

JUNIPER, a tool from the Simultaneous Phylogenetic family, was specifically designed to overcome computational and methodological limitations in existing tools. It incorporates a novel statistical model for within-host variant frequencies and uses parallelization to handle large datasets [66]. On a dataset of over 160,000 deep-sequenced SARS-CoV-2 genomes, its model for intrahost single nucleotide variant (iSNV) frequencies showed minimal discrepancy between the empirical and theoretical probability density, validating its approach [66]. The tool has been demonstrated on large-scale datasets, including over 1,500 bovine H5N1 cases and over 13,000 human COVID-19 cases, quantifying elevated transmission rates and the efficacy of vaccination [66].

Methods that do not account for within-host diversity or that assume a complete transmission bottleneck (where only a single genotype is transmitted) can be misled when this assumption is violated, as is common for pathogens like HIV and Mycobacterium tuberculosis [4] [67]. Furthermore, most methods (17 out of 22 according to the systematic review) model the transmission process itself, but only a minority (8 out of 22) account for imperfect case detection, which can introduce bias if unaccounted for in an outbreak with many unreported cases [67].

Quantitative Comparison of Method Characteristics

The table below synthesizes data from the systematic review and recent preprints to compare the characteristics of different method families across several critical dimensions.

Table 2: Performance and Characteristic Comparison of Method Families

Characteristic	Non-Phylogenetic (NPF)	Sequential Phylogenetic (SeqPF)	Simultaneous Phylogenetic (SimPF)
Within-Host Evolution Model	Varies; often not explicit.	Commonly a coalescent process [67].	Coalescent or pure-birth process (e.g., JUNIPER [66]).
Transmission Process Model	Majority model this process [67].	Majority model this process [67].	Majority model this process [67].
Accounts for Unsamp. Cases	Few methods (e.g., 2/8 in review [67]).	Few methods [67].	More common (e.g., JUNIPER, TransPhylo [66] [67]).
Use of iSNVs	Limited.	Limited.	High (e.g., JUNIPER's core model [66]).
Computational Scalability	Generally high.	Can be limited by phylogenetic step.	Varies; JUNIPER uses parallelization for scale [66].
Ease of Implementation	Straightforward, two-step.	Requires choosing/phylogenetic tool.	Can be complex, but more integrated.

Experimental Protocols for Benchmarking

To ensure robust and reproducible comparisons between reconstruction methods, benchmark studies typically follow a structured protocol. The workflow below outlines the key stages in a comprehensive benchmarking experiment, from data preparation to performance evaluation.

Data Preparation and Simulation

• Simulated Outbreaks: Use a stochastic transmission model (e.g, an SIR or branching process) to generate a realistic outbreak with known transmission events. The model should incorporate key parameters such as the reproduction number (R), generation interval distribution, and potentially overdispersion (superspreading) [66] [67].
• Within-Host Evolution: For each infected host in the simulated outbreak, simulate the evolution of the pathogen population within the host. A common approach is to use a coalescent model or a pure-birth process to generate genetic diversity [66] [67]. The simulation should model mutation rates and potential selection pressures.
• Sampling: Simulate the sampling of pathogens from infected hosts at specified times. This step should account for incomplete sampling, where not all cases are observed, and for the potential bias in which cases are selected for sequencing [67].
• Real Outbreaks with Known Links: Complement simulated data with real outbreak datasets where transmission links have been strongly confirmed through intensive epidemiological investigation and contact tracing. These serve as a valuable ground-truth validation [66].

Inference and Evaluation

• Method Execution: Run the transmission tree reconstruction methods (from different families) on the prepared datasets using their standard inference procedures (e.g., MCMC for Bayesian methods). For sequential methods, this includes first running a dedicated phylogenetic inference tool [67].
• Performance Metrics: Compare the inferred trees to the known ground truth using standardized metrics:
  • Sensitivity and Specificity: For identifying direct transmission pairs.
  • Ancestor-Descendant Accuracy: The proportion of correctly identified ancestor-descendant relationships in the tree.
  • Tree Distance Metrics: Metrics like the Robinson-Foulds distance can be adapted to compare the topological similarity between the inferred and true transmission trees.
  • Parameter Estimation Error: The difference between inferred key parameters (e.g., R, mutation rate) and their true values.

Success in transmission tree reconstruction relies on a combination of computational tools, data resources, and laboratory reagents. The following table details essential components of the research pipeline.

Table 3: Essential Research Reagents and Resources for Transmission Tree Studies

Item Name	Type	Function in Research
Next-Generation Sequencing (NGS)	Laboratory Technology	Generates whole-genome sequence data or deep sequencing data for intrahost variant identification from pathogen samples. It is the foundation for genomic analysis [66] [4].
JUNIPER	Software Tool	A highly-scalable, simultaneous phylogenetic tool for reconstructing transmission trees that incorporates intrahost variation and incomplete sampling. Ideal for large outbreaks [66].
TransPhylo	Software Tool	A Bayesian method in the Simultaneous Phylogenetic family that infers transmission trees while accounting for unsampled cases. Useful for smaller outbreaks or when used as a component in other methods [66] [67].
axe-core / axe DevTools	Software Library / Tool	An open-source accessibility engine for testing web-based data visualization dashboards, ensuring that color-coded results meet contrast guidelines for interpretability by all researchers [68].
Reference Genome	Data Resource	A high-quality, annotated genome sequence of the pathogen used as a reference for aligning short reads from NGS data during the sequence assembly process [4].
Multi-locus Sequence Typing (MLST) Database	Data Resource	A curated database that defines strain types based on sequences of a set of housekeeping genes. Provides a standardized nomenclature for initial pathogen classification and clustering [4].

Evaluating Robustness in Estimating Key Parameters like Effective Reproduction Number (Rₜ)

Robustness in phylodynamic inference refers to the reliability and stability of parameter estimates, such as the effective reproduction number (Rₜ), when confronted with real-world data challenges including model misspecification, incomplete sampling, and genetic sequence limitations. As genomic data become increasingly integral to outbreak investigation, understanding the performance characteristics of different phylodynamic methods is essential for researchers, scientists, and drug development professionals who depend on these tools for source attribution and transmission dynamics reconstruction. This guide provides a systematic comparison of leading phylodynamic methods, evaluating their robustness through published experimental data and simulation studies to inform method selection for outbreak research.

Methodological Foundations of Phylodynamic Approaches

Phylodynamic methods integrate evolutionary models with epidemiological dynamics to reconstruct transmission parameters from genetic sequence data. The core approaches differ in their conceptual foundations and mathematical structure, which directly impacts their robustness for parameter estimation.

Table 1: Core Phylodynamic Methodologies

Method Class	Fundamental Principle	Key Parameters Estimated	Theoretical Strengths	Theoretical Limitations
Structured Birth-Death Models	Models population dynamics through birth (infection), death (recovery/removal), and sampling rates [3]	Rₜ, migration rates, population sizes	Naturally accommodates changing population sizes; directly models sampling process	Computationally intensive; sensitive to model specification
Structured Coalescent Models	Based on the probability that lineages coalesce in reverse time, dependent on effective population size [37]	Rₜ, effective population size, migration rates	Efficient for large datasets; well-established theoretical foundation	Assumes constant population sizes between sampling events; sensitive to sampling density
Discrete Trait Analysis	Treats location transitions as a substitution process using continuous-time Markov chains [37] [5]	Migration rates, source probabilities	Computationally efficient; flexible for complex discrete state models	May oversimplify epidemiological dynamics; potentially lower statistical power for migration inference

The multi-type birth-death model has seen significant advancements recently, with the BEAST2 package bdmm implementing algorithmic improvements that dramatically increased numerical robustness and efficiency. These changes enabled analysis of datasets containing several hundred genetic samples, overcoming previous limitations of approximately 250 samples that severely constrained parameter estimation precision in structured models [3] [69]. This enhancement is particularly crucial for Rₜ estimation in complex outbreaks with multiple populations.

Figure 1: Phylodynamic Analysis Workflow for Parameter Estimation

Experimental Protocols for Robustness Assessment

Robustness evaluation in phylodynamics employs carefully designed simulation studies that test methodological performance under controlled conditions with known parameter values. These protocols typically follow a three-stage approach that mirrors real-world analytical challenges.

Simulation-Based Validation Framework

A comprehensive robustness assessment begins with model calibration to empirical data to ensure simulated outbreaks reflect realistic epidemiological dynamics. In one HIV study, researchers developed a complex 120-compartment model structured by infection stages, age groups, diagnosis status, and risk behaviors, then calibrated it to surveillance data from men who have sex with men in San Diego [37]. This model incorporated five HIV infection stages, four age groups, three diagnosis stages, and two risk groups, creating a sophisticated testing ground for simpler phylodynamic models.

The second stage involves simulating genealogies and genetic sequences under the calibrated model. Researchers typically generate multiple replicate datasets of varying sizes (e.g., 175, 500, and 1,000 sequences) to evaluate how sampling density affects parameter estimation [3] [37]. For the HIV study, sequences equivalent to both the HIV partial pol gene and complete genome were simulated to test the impact of genetic information content [37].

The final stage employs phylodynamic inference using simplified models that represent standard analytical practice. The key robustness question is whether these simpler models can accurately recover known parameters from the complex simulation truth. Performance is quantified through statistical measures of bias, precision, and coverage across multiple replicates [37].

Experimental Dataset for Comparative Validation

The Influenza A virus HA sequence dataset provides a practical experimental framework for robustness comparison. One study analyzed two partly overlapping datasets—500 samples versus 175 samples—to quantify the information gained with larger sample sizes [3] [69]. This design directly tests robustness to sampling density, a critical practical constraint in outbreak investigations. The comparison assessed global migration patterns and seasonal dynamics inferred from each dataset, with the larger dataset demonstrating improved precision of parameter estimates, particularly for structured models with high numbers of inferred parameters [3].

Comparative Performance Analysis

Direct comparison of phylodynamic methods reveals significant differences in robustness characteristics, particularly for estimating key parameters like Rₜ under realistic analytical conditions.

Table 2: Robustness Performance Across Methodologies

Method	Sample Size Requirements	Performance with Model Misspecification	Computational Efficiency	Best Application Context
Structured Birth-Death (bdmm)	250-500 sequences for reliable inference [3]	Moderate robustness; improved with algorithmic enhancements to reduce numerical instability [3]	Moderate; improved with recent algorithmic changes [3]	Structured populations with known sampling biases; when sampling process must be explicitly modeled
Structured Coalescent	≥1,000 sequences for reliable migration rate estimation [37]	Variable performance; simpler models show bias with complex dynamics but still provide useful estimates [37]	High for approximations; lower for exact implementations	Large datasets with complex population structure; when computational efficiency is prioritized
Discrete Trait Analysis	Not explicitly quantified in studies	Sensitive to model complexity; may show bias with strong population structure [37]	High	Preliminary analysis; when computational resources are limited

The impact of model misspecification was systematically evaluated in the HIV simulation study, which tested whether simple models could accurately estimate migration rates when applied to data generated from a complex ground truth. The results demonstrated that even misspecified models could provide useful estimates, with sample size being a critical factor—models with at least 1,000 sequences showed significantly better performance despite structural simplicity [37].

Algorithmic improvements have substantially impacted robustness characteristics. The bdmm package overcame numerical instability issues that previously limited analysis to approximately 250 samples through implementation of techniques that prevented numerical underflow in probability density calculations [3]. This enhancement was particularly valuable for structured models with high numbers of inferred parameters, where sufficient samples from each subpopulation are essential for reliable estimation.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Research Reagents for Phylodynamic Robustness Evaluation

Reagent/Resource	Function in Robustness Assessment	Implementation Considerations
BEAST2 Platform	Bayesian evolutionary analysis software providing implementation of multiple phylodynamic methods [3] [13]	Modular architecture allows method comparison; requires careful prior specification and MCMC diagnostics
bdmm Package	Implements multi-type birth-death model with sampling for structured population inference [3]	Recently improved numerical robustness enables larger datasets; flexible sampling scheme specification
PhyDynR Package	Implements structured coalescent models with nonlinear population dynamics [37]	Useful for testing robustness under complex population dynamics; R-based implementation
Reference Genomic Sequences	Empirical datasets for method validation and calibration [3] [64]	Influenza A and Ebola virus datasets provide realistic test cases with different evolutionary characteristics
Scenario Simulation Pipeline	Custom computational framework for generating synthetic outbreaks with known parameters [37]	Enables controlled robustness testing; requires careful calibration to realistic epidemiological dynamics

Figure 2: Robustness Evaluation Decision Pathway

Discussion and Practical Recommendations

The robustness of phylodynamic methods depends critically on the interplay between model complexity, sample size, and implementation details. Structured birth-death models offer the advantage of explicitly modeling the sampling process, which is particularly valuable for outbreak investigation where sampling biases are common. Recent algorithmic improvements have substantially enhanced their numerical robustness, enabling application to larger datasets [3]. Structured coalescent approaches provide computational efficiency for large datasets but may require sample sizes exceeding 1,000 sequences for reliable migration rate estimation [37].

For researchers estimating time-varying parameters like Rₜ, model misspecification presents a persistent challenge. Studies indicate that simpler models can provide useful estimates even when the true data-generating process is more complex, particularly with sufficient sample sizes [37]. This robustness to misspecification is encouraging for practical outbreak analytics where the true epidemiological dynamics are never fully known.

Practical robustness evaluation should prioritize sample size adequacy, with different methods having distinct requirements. The significant improvement in parameter precision when analyzing 500 versus 175 Influenza A sequences demonstrates that larger samples partially compensate for model limitations [3]. Additionally, multiple prior sensitivity analysis is essential, as posterior inferences can be sensitive to prior selection, particularly for evolutionary parameters [13].

Future methodological development should focus on enhancing robustness to common data limitations, including uneven sampling across regions and time periods, while maintaining computational tractability for real-time outbreak analytics. Integration of phylodynamic methods with traditional epidemiological approaches will likely provide the most robust framework for estimating critical parameters like Rₜ during ongoing outbreaks.

Source attribution is a critical methodology in epidemiology for reconstructing the transmission of infectious diseases from a specific source, such as a population, individual, or location. It plays a vital role in public health surveillance and outbreak management [38]. Molecular source attribution, which utilizes pathogen genetic data, has become increasingly powerful with the advent of whole-genome sequencing (WGS), enabling high-resolution tracing of transmission pathways [70] [38].

The field encompasses a diverse array of computational methods, each with distinct strengths, requirements, and applications. This creates a challenging landscape for researchers and public health professionals who must select the most appropriate technique for a specific outbreak scenario or research question. This guide provides a structured comparison of predominant source attribution methods, focusing on their operational profiles, performance characteristics, and implementation requirements to inform method selection.

Source attribution methods can be broadly categorized by their underlying computational approach and the primary data they utilize. The following table summarizes the core characteristics of four prominent methods.

Table 1: Core Characteristics of Major Source Attribution Methods

Method	Core Principle	Primary Data Input	Typical Output	Key Applications
Phylogenetic Clustering [39] [38]	Groups cases based on genetic similarity thresholds (genetic distance, phylogenetic credibility).	Molecular sequences (Single locus, WGS).	Cluster membership (e.g., clustered vs. non-clustered), cluster size.	Identifying transmission clusters and risk factors associated with clustering.
Source Attribution (SA) [39]	Estimates infector probabilities between cases using time-scaled phylogenies and epidemiological data.	Molecular sequences, incidence, prevalence, clinical data (e.g., CD4).	Infector probability matrix, individual out-degree (estimated number of transmissions).	Quantifying individual transmission rates and identifying transmission risk factors.
RandomForest (Supervised Machine Learning) [70]	A classification algorithm trained on sequences from known sources to predict the source of human cases.	Whole Genome Sequencing (WGS) data (core and/or accessory genome).	Probabilistic assignment of human cases to source classes.	Attributing human infections to animal or food reservoirs.
Bayesian Frequency Matching (e.g., Hald method) [70]	Compares the frequency of bacterial subtypes in human cases to their frequency in animal/food sources.	Microbial subtyping data or WGS-based subtypes.	Estimated number/proportion of human cases attributed to each source.	Partitioning the human disease burden of foodborne illnesses to specific reservoirs.

The following diagram illustrates the logical decision pathway for selecting among these primary methods based on the research question and data availability.

Performance Comparison of Key Methods

Comparison of Phylogenetic Clustering vs. Source Attribution

A simulation study of HIV transmission among men who have sex with men compared phylogenetic clustering with a phylodynamic source attribution method. The study assessed their ability to correctly identify patient attributes as transmission risk factors [39].

Table 2: Performance Comparison: Clustering vs. Source Attribution for Identifying Transmission Risk Factors

Performance Metric	Phylogenetic Clustering	Source Attribution (SA) Method
Error Rates	Higher error rates	Lower error rates
Sensitivity	Lower sensitivity	Higher sensitivity
Robustness of Estimates	Does not provide robust estimates of transmission risk ratios	Can alleviate drawbacks of phylogenetic clustering, but may not provide robust risk ratio estimates without formal population genetic modeling
Key Limitation	Misleading associations with covariates correlated with time since infection (e.g., CD4 count, viral load, age)	Requires additional epidemiological data and independent estimates of incidence/prevalence

Comparison of Methods for Genomic Source Attribution

A study on Salmonella Typhimurium compared three WGS-based source attribution methods using a dataset of 902 isolates from the British Isles and Denmark [70].

Table 3: Performance Comparison of WGS-Based Methods for Salmonella Source Attribution

Performance Metric	RandomForest (ML)	AB_SA (Accessory genes)	Bayesian (Frequency Matching)
Attribution Accuracy	Higher accuracy when including accessory genome features	Lower accuracy than RandomForest	Overall attribution estimates varied little with or without accessory genome
Impact of Accessory Genome	Improved attribution accuracy	N/A (Method is inherently based on accessory genes)	Minimal impact on overall estimates
Computational Execution Time	Much slower execution	Much faster execution	Much faster execution
Primary Advantage	High accuracy with sufficient genomic features	Fast execution, model-based probabilistic assignment	Fast execution, provides population-level attribution estimates

Experimental Protocols for Source Attribution Studies

Protocol for Comparing Clustering and SA Methods

This protocol is adapted from a simulation study comparing phylogenetic clustering and source attribution methods for HIV [39].

• Epidemic Simulation: Simulate an epidemic trajectory using a complex model calibrated to a real-world population (e.g., men who have sex with men in San Diego, USA).
• Genealogy and Sequence Simulation: Use the epidemic trajectory to simulate genealogies and sequence alignments. To reflect real-world constraints, sequences can be limited to specific genomic regions (e.g., the partial pol gene) or can simulate the complete pathogen genome.
• Method Application:
  • Clustering Analysis: Apply genetic distance or phylogeny-based clustering algorithms to the simulated sequences. Regress cluster membership or size on patient covariates to identify transmission risk factors.
  • Source Attribution Analysis: Apply a phylodynamic SA method (e.g., [39]) to the same data, using the time-scaled phylogeny and incorporating available epidemiological data (e.g., incidence, prevalence) to calculate infector probabilities.
• Performance Evaluation: Compare the error rates and sensitivity of both methods for identifying the known transmission risk factors built into the simulation model. Specifically evaluate the propensity of clustering methods to show spurious associations with covariates correlated with time since infection.

Protocol for Comparing WGS-Based Attribution Methods

This protocol is based on a study comparing RandomForest, AB_SA, and Bayesian methods for Salmonella source attribution [70].

• Strain Selection and WGS: Select a panel of bacterial isolates from both reservoir hosts (animals) and human cases. Perform Whole Genome Sequencing on all isolates.
• Data Processing:
  • Perform quality control on the sequenced data.
  • Conduct in-silico serotyping and multilocus sequence typing (cgMLST and wgMLST) to characterize the core and accessory genome.
  • For methods like RandomForest and Bayesian, create separate feature sets comprising only core genome loci and both core and accessory genome loci to evaluate the impact of the accessory genome.
• Model Training and Testing:
  • RandomForest: Train a supervised classification random forest model on the animal (source) isolates with known primary sources. Use a held-out test set of animal isolates to evaluate attribution accuracy before applying to human cases.
  • AB_SA: Apply the multinomial logistic regression classifier, which uses the presence/absence of accessory genes for attribution.
  • Bayesian: Apply the modified Hald model (Bayesian frequency matching) using the generated subtype data.
• Performance Evaluation: For RandomForest and AB_SA, compare the attribution accuracy on the test set with known sources. For all methods, compare the resulting attribution estimates for human cases and the computational execution time.

The Scientist's Toolkit: Essential Reagents & Computational Tools

Table 4: Key Research Reagents and Computational Solutions for Source Attribution

Item Name	Function / Application	Specific Examples / Notes
Whole Genome Sequencing (WGS) Data	Provides the highest resolution data for discriminating between pathogen strains and inferring transmission links.	Essential for methods like RandomForest and AB_SA; can be used with clustering and SA methods [70] [38].
Pathogen Genomic Sequences	The fundamental input for all molecular source attribution methods.	Can range from single-locus to whole-genome data [38].
BEAST/BEAST X Software	A leading software platform for Bayesian evolutionary analysis. Used for phylogenetic reconstruction, divergence time dating, and phylodynamic inference, forming the basis for many SA and phylogeographic methods [6] [29].	Enables complex trait evolution, molecular clock models, and scalable inference [29].
BEAGLE Library	A high-performance computational library for phylogenetic inference.	Used to accelerate likelihood calculations in BEAST and other software [29].
Reference Genome	Used for aligning sequence reads and calling variants in WGS data.	Critical for consistent analysis across samples [38].
Epidemiological Metadata	Clinical, demographic, and temporal data associated with each sequenced sample.	Informs models (e.g., SA methods), helps validate predictions, and is crucial for interpreting results [39].
Hamiltonian Monte Carlo (HMC)	A Markov chain Monte Carlo (MCMC) algorithm for efficient sampling from high-dimensional probability distributions.	Implemented in BEAST X to improve scalability and sampling efficiency for large datasets and complex models [29].

This framework synthesizes performance data and operational requirements for four prominent source attribution methods. The optimal choice is contingent on the specific research objective, the nature of available data, and computational constraints. Phylogenetic clustering offers a rapid, accessible entry point for identifying transmission clusters but carries a higher risk of biased inference. Phylodynamic source attribution methods provide more powerful, quantitative estimates of transmission flows but demand richer epidemiological data and greater computational investment. For attributing illnesses to reservoir sources, supervised learning methods like RandomForest achieve high accuracy when source data is available for training, while Bayesian frequency methods offer a robust, faster alternative for population-level attribution. By aligning the research question with the methodological strengths and limitations outlined here, researchers can make more informed decisions to enhance the accuracy and reliability of their source attribution studies.

Conclusion

The comparative analysis of phylodynamic methods reveals that no single approach is universally superior; rather, the choice depends on outbreak scale, data quality, and specific public health questions. Foundational Bayesian phylogeography provides detailed historical reconstructions, while novel scalable methods like SPRTA and ScITree are essential for pandemic-speed responses. Success hinges on understanding the drivers of inference—particularly the interplay between genomic and temporal data—and rigorously validating models against simulated and real-world benchmarks. Future directions must focus on developing integrated, multi-scale models that capture feedback between evolution, epidemiology, and interventions, standardizing data sharing, and building accessible tools to transform phylodynamic insights into actionable public health strategies for future outbreaks.

References

• 1. Epidemiological inference from pathogen genomes: A review ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9241095/]
• 2. Phylogenetic and phylodynamic approaches to ... [https://www.nature.com/articles/s41576-022-00483-8]
• 3. Robust Phylodynamic Analysis of Genetic Sequencing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9413058/]
• 4. Molecular source attribution - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC9671344/]
• 5. Approaches and challenges to inferring the geographical ... [https://www.sciencedirect.com/science/article/pii/S2666524723002963]
• 6. Robust phylodynamic inference and model specification for ... [https://www.sciencedirect.com/science/article/pii/S1755436525000349]
• 7. PhyloTune: An efficient method to accelerate phylogenetic ... [https://www.nature.com/articles/s41467-025-61684-3]
• 8. Multi-scale phylodynamic modelling of rapid punctuated ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12270310/]
• 9. Phylodynamic Analysis | 176 genomes | 6 Mar 2020 [https://virological.org/t/phylodynamic-analysis-176-genomes-6-mar-2020/356]
• 10. Phylodynamic Analysis - an overview [https://www.sciencedirect.com/topics/medicine-and-dentistry/phylodynamic-analysis]
• 11. Deep learning from phylogenies to uncover... | Nature Communications [https://www.nature.com/articles/s41467-022-31511-0?error=cookies_not_supported&code=07b79713-36fc-4311-baa5-2bbe61133545]
• 12. The EpiFusion Analysis Framework for joint phylodynamic and... [https://pubmed.ncbi.nlm.nih.gov/40469802/]
• 13. Frontiers | Animal Disease Surveillance in the 21st Century... [https://www.frontiersin.org/journals/veterinary-science/articles/10.3389/fvets.2020.00176/full]
• 14. The influence of phylodynamic model specifications on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5789282/]
• 15. Accounting for reporting delays in real-time phylodynamic ... [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012970]
• 16. Source attribution [https://en.wikipedia.org/wiki/Source_attribution]
• 17. Source Attribution - an overview [https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/source-attribution]
• 18. Source Attribution of Foodborne Diseases: Potentialities ... [https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2018.01983/full]
• 19. A Review on Microbiological Source Attribution Methods of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10924193/]
• 20. Phylodynamic assessment of SNP distances from whole ... [https://www.nature.com/articles/s41598-025-94646-2]
• 21. Applications of Bayesian Phylodynamic Methods in a ... [https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2016.00067/full]
• 22. The EpiFusion Analysis Framework for joint phylodynamic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12134729/]
• 23. Integrating molecular epidemiology and social network ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5135626/]
• 24. Identification of transmission clusters with high influence on ... [https://www.medrxiv.org/content/10.1101/2025.11.18.25340546v1]
• 25. Integration of Epidemiological and Genomic Data to ... [https://www.mdpi.com/2076-0817/12/1/100]
• 26. Hypothesis testing using Phylogenetics (HyPhy) and ... [https://www.medrxiv.org/content/10.1101/2025.09.28.25336847v1]
• 27. Going back to the roots: Evaluating Bayesian phylogeographic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7686256/]
• 28. Improving the performance of Bayesian phylogenetic ... [https://bmcecolevol.biomedcentral.com/articles/10.1186/s12862-020-01609-4]
• 29. BEAST X for Bayesian phylogenetic, phylogeographic and ... [https://www.nature.com/articles/s41592-025-02751-x]
• 30. MTML-msBayes: Approximate Bayesian ... comparative [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-12-1]
• 31. Inference of Epidemiological Dynamics Based on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4222655/]
• 32. Estimating pathogen spread using structured coalescent and... [https://pubmed.ncbi.nlm.nih.gov/39461051/]
• 33. ModelFinder: Fast Model Selection for Accurate ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5453245/]
• 34. Assessing phylogenetic confidence at pandemic scales [https://www.nature.com/articles/s41586-025-09567-x]
• 35. ScITree: Scalable Bayesian inference of transmission tree ... [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012657]
• 36. An Innovative Approach to Quantifying Evolutionary ... [https://bioengineer.org/sparta-an-innovative-approach-to-quantifying-evolutionary-uncertainty/]
• 37. Robust phylodynamic inference and model specification for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12344737/]
• 38. Molecular source attribution | PLOS Computational Biology [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1010649]
• 39. Comparison of cluster-based and source-attribution methods ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5910297/]
• 40. Multi-scale phylodynamic modelling of rapid punctuated ... [https://arxiv.org/html/2412.03896v2]
• 41. Complex Population Dynamics and the Coalescent Under ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3249372/]
• 42. A framework for the comparison of agent-based models [https://link.springer.com/article/10.1007/s10458-022-09559-5]
• 43. Comparative phylodynamics reveals the evolutionary history ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9129158/]
• 44. Estimating pathogen spread using structured coalescent ... [https://www.sciencedirect.com/science/article/pii/S1755436524000562]
• 45. How does date-rounding affect phylodynamic inference for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11991728/]
• 46. Quantifying and Mitigating the Effect of Preferential ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4777449/]
• 47. Estimating epidemic dynamics with genomic and time series ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12134936/]
• 48. Unlocking the Secrets of Bacterial Genomics Workshop 2025 [https://hts.iit.edu/events/2025-workshop.html]
• 49. Phylodynamic applications in 21st century global infectious ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5683535/]
• 50. ScITree: Scalable Bayesian inference of transmission tree ... [https://pubmed.ncbi.nlm.nih.gov/40493703/]
• 51. Deep learning from phylogenies to uncover... | Nature Communications [https://www.nature.com/articles/s41467-022-31511-0?error=cookies_not_supported]
• 52. Visualizing BEAST2 results in R [https://taming-the-beast.org/tutorials/Visualizing-BEAST2-results/]
• 53. TreeAnnotator | BEAST Documentation [https://beast.community/treeannotator]
• 54. Chapter 4 Phylogenetic Tree Visualization [https://yulab-smu.top/treedata-book/chapter4.html]
• 55. Model misspecification misleads inference of the spatial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10089176/]
• 56. Phylodynamic analysis and evaluation of the balance ... [https://www.nature.com/articles/s41598-020-64477-4]
• 57. Phylogenetic and phylodynamic approaches to understanding ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9028907/]
• 58. Inferring demographic parameters in bacterial genomic data ... [https://bmcecolevol.biomedcentral.com/articles/10.1186/s12862-018-1210-5]
• 59. Decoding the Fundamental Drivers of Phylodynamic Inference [https://pmc.ncbi.nlm.nih.gov/articles/PMC10284498/]
• 60. Inferring epidemiological parameters from phylogenies using ... [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005416]
• 61. ScITree: Scalable Bayesian inference of transmission tree ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12176303/]
• 62. Phylodynamic Inference of Bacterial Outbreak Parameters ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8963328/]
• 63. DeepDynaForecast: Phylogenetic-informed graph deep ... [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011351]
• 64. Phylodynamic assessment of intervention strategies for the ... [https://www.nature.com/articles/s41467-018-03763-2]
• 65. Common Methods for Phylogenetic Tree Construction and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11117635/]
• 66. JUNIPER: Reconstructing Transmission Events from Next ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11908323/]
• 67. Methods Combining Genomic and Epidemiological Data in ... [https://www.mdpi.com/2076-0817/11/2/252]
• 68. Elements must meet minimum color contrast ratio thresholds [https://dequeuniversity.com/rules/axe/4.8/color-contrast]
• 69. Robust Phylodynamic Analysis of Genetic Sequencing ... [https://doaj.org/article/af5b9d6ae566420685a138019c38acf4]
• 70. Comparison of three source attribution methods applied to ... [https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2024.1393824/full]