Beyond the Tree of Life: Testing for Reticulate Evolution with Phylogenetic Networks in Biomedical Research

Abstract

This article provides a comprehensive guide for researchers and drug development professionals on the application of phylogenetic networks to test for and interpret reticulate evolution. As high-quality genomic data becomes ubiquitous, recognizing and modeling non-treelike evolutionary processes—such as hybridization, introgression, and gene flow—is critical for accurate phylogenetic inference and its downstream applications. We explore the foundational shift from trees to networks, detail cutting-edge methodological frameworks and computational tools, address common challenges in model selection and interpretation, and validate network approaches against traditional methods. By synthesizing theory and practice, this review empowers scientists to leverage phylogenetic networks for more biologically realistic evolutionary investigations in genomics, disease tracing, and biodiversity research.

From Tree to Web: Foundational Concepts of Reticulate Evolution and Phylogenetic Networks

Reticulate evolution represents a fundamental departure from the strictly branching patterns of the traditional "Tree of Life" metaphor. It encompasses evolutionary processes where genetic material is exchanged between distinct lineages, creating network-like relationships. This framework is essential for understanding the full complexity of evolutionary history, particularly in rapidly diversifying groups and across all levels of the biological hierarchy. The primary processes underlying reticulate evolution include hybridization (the interbreeding of individuals from genetically distinct populations), introgression (the transfer of genetic material between species through repeated backcrossing), and horizontal gene transfer (the movement of genetic material between organisms other than by vertical inheritance) [1] [2]. These processes can act as significant mechanisms for the origin and growth of biological diversity and complexity, providing the raw material for evolutionary innovation alongside the more established forces of mutation and selection [1].

The study of reticulate evolution has gained substantial importance with the advent of phylogenomics, as whole-genome sequencing projects have provided convincing evidence that such processes are not exceptional but rather ubiquitous across the tree of life [1]. This realization has prompted the development of new analytical models and tools that move beyond strictly bifurcating trees to phylogenetic networks, which can simultaneously represent both divergent and convergent evolutionary pathways [3]. For researchers in fields including drug development, understanding these processes is critical, as the reticulate exchange of genetic material can influence the evolution of pathogenicity, drug resistance, and functional traits in model organisms used for biomedical research [4].

Key Reticulate Processes and Their Definitions

Table 1: Core Processes in Reticulate Evolution

Process	Definition	Key Characteristics	Evolutionary Scale
Hybridization	The interbreeding of individuals from two genetically distinct populations, lineages, or species [3].	Creates novel genetic combinations; can lead to hybrid speciation or introgression.	Typically occurs between populations or closely related species.
Introgression	The transfer of genetic material from one species into the gene pool of another by repeated backcrossing of hybrids with their parent species [3] [2].	Results in the incorporation of alien genes; can facilitate adaptive evolution.	Occurs between closely related species following hybridization.
Horizontal Gene Transfer (HGT)	The movement of genetic material between organisms other than by vertical transmission from parent to offspring [1] [2].	Allows for acquisition of novel traits across distant lineages; common in prokaryotes and increasingly recognized in eukaryotes.	Can occur between distantly related species, even across different kingdoms.
Endosymbiosis	An intimate symbiotic relationship where one organism lives inside the cells of another, potentially leading to genome integration [1].	Can result in major evolutionary transitions (e.g., origin of organelles); a form of whole-genome fusion.	Ultra-deep reticulation, as in the origin of eukaryotic organelles.

Phylogenomic Workflow for Detecting Reticulate Evolution

Distinguishing the signals of reticulate evolution from other sources of gene tree discordance, such as Incomplete Lineage Sorting (ILS), is a central challenge in phylogenomics. ILS occurs when ancestral polymorphisms persist through multiple speciation events, leading to gene genealogies that diverge from the species tree. The co-occurrence of ILS and introgression can create complex evolutionary signals that require specialized models for accurate interpretation [3]. A robust phylogenomic workflow for testing reticulate evolution involves multiple steps, from data collection to model selection and hypothesis testing.

Table 2: Key Steps in a Phylogenomic Reticulation Detection Workflow

Step	Description	Common Tools/Methods	Notes on Reticulation Detection
1. Data Collection & Locus Sampling	Selection of multiple independent genomic loci or windows from the species of interest.	Genome alignments, targeted sequencing.	Loci should be independent (e.g., spaced apart in the genome) to satisfy model assumptions [3].
2. Gene Tree Estimation	Inferring phylogenetic trees for each individual locus.	RAxML, IQ-TREE	Bootstrap resampling is used to assess confidence in individual gene trees [3].
3. Species Tree/Network Inference	Reconstructing the overarching evolutionary history from the set of gene trees.	Multispecies Coalescent (MSC) models, Phylogenetic Networks (e.g., PhyloNet)	MSC models account for ILS but not reticulation; Network models (MSNC) account for both [3].
4. Incongruence Assessment	Quantifying and visualizing discordance among gene trees and between gene trees and the species tree/network.	Topological comparisons, quartet methods.	Widespread, strongly supported incongruence can signal reticulation.
5. Model Testing & Hypothesis Validation	Statistically testing whether a reticulate model provides a significantly better fit to the data than a purely bifurcating tree.	Likelihood-based tests, parameter estimation (e.g., inheritance probabilities, γ).	Methods can help determine the timing of introgression relative to speciation [2].

Visualizing the Phylogenomic Workflow

The following diagram illustrates the logical flow of a phylogenomic analysis designed to detect reticulate evolution, highlighting the critical decision points for distinguishing between different evolutionary processes.

Experimental Protocols for Key Studies

Protocol: Re-analysis ofAnopheles gambiaeComplex Using Phylogenetic Networks

This protocol is based on a landmark re-analysis of mosquito genomes that demonstrated the power of phylogenetic networks to uncover complex evolutionary histories involving both ILS and introgression [3].

1. Data Acquisition and Preparation:

• Source: Obtain the whole-genome multiple sequence alignment (MAF format) for the six species of the Anopheles gambiae complex, using Anopheles christyi as an outgroup.
• Locus Sampling: To ensure independence of loci, sample genomic windows such that each new locus is at least 64 kb apart from the previous one. This mitigates the effects of linkage disequilibrium.
• Region Separation: Partition chromosomes into standard and inverted regions (e.g., 2La and 3La inversions) using reference genome coordinates, as these regions may have distinct evolutionary histories.

2. Gene Tree Estimation:

• For each sampled locus, infer phylogenetic trees using a maximum-likelihood method like RAxML under the GTRGAMMA model.
• Assess the uncertainty in each gene tree by generating 100 bootstrap replicates per locus. These bootstrap trees are used directly in the subsequent network inference to account for gene tree estimation error.

3. Phylogenetic Network Inference:

• Use the PhyloNet software package, which implements the Multi-Species Network Coalescent (MSNC) model.
• Input the bootstrap gene trees for all loci into PhyloNet's inference command (e.g., InferNetwork_ML).
• The MSNC model will simultaneously estimate a phylogenetic network topology, its branch lengths (in coalescent units), and inheritance probabilities (γ) at reticulation nodes, which represent the proportion of genetic material inherited from each parent in a hybridization event.

4. Analysis and Interpretation:

• The output network is interpreted as the reticulate evolutionary history of the species complex.
• The inheritance probabilities (γ) quantify the genomic impact of each hybridization event.
• To elucidate the patterns of introgression across the genome, a method quantifying the usage of reticulation branches in the network by different genomic regions (e.g., different chromosomes) can be applied.

Protocol: Network-Based Visualization of Transposable Element (TE) Evolution

This protocol describes a novel network-based method to study the evolution of transposable elements (TEs), which can be subject to horizontal transfer and provide insights into reticulate evolution [5].

1. TE Identification and Data Collection:

• Select a focal TE superfamily (e.g., Tc1/mariner) and a set of study species.
• Identify all copies of the TE superfamily in the genome of each species using a tool like RepeatMasker.

2. Network Construction:

• Node Definition: Each individual TE sequence is represented as a single node in the network.
• Edge Definition: Perform all-vs-all BLAST comparisons (nucleotide-level) between the TE sequences. The strength of the connection (edge) between two nodes is determined by the BLAST bit score, which reflects sequence similarity.
• This creates a monopartite sequence similarity network.

3. Network Visualization and Cluster Analysis:

• Visualize the network using a layout algorithm (e.g., OpenOrd in Gephi) that clusters nodes based on connection strength.
• Identify distinct clusters within the network using community detection algorithms, such as the Louvain Method for modularity, implemented in Gephi.
• Annotate clusters based on metadata, such as the species of origin or the presence of functional protein domains (e.g., the DDE3 domain for transposition).

4. Hypothesis Testing:

• Statistically test properties of the network. For example, compare the connectivity (e.g., node degree) of TEs that contain a functional domain versus those that do not using a non-parametric test like the Mann-Whitney U test.
• The network structure can reveal evolutionary patterns, such as convergent acquisition of domains or the evolutionary history of TE subfamilies across species, which are difficult to discern with traditional phylogenetic trees.

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Key Research Reagents and Computational Tools for Reticulate Evolution Studies

Item Name	Type	Function in Research	Example Use Case
Whole-Genome Alignment Data	Dataset	Provides the fundamental nucleotide/protein sequences for comparative analysis across taxa.	Used as input for locus sampling and gene tree estimation [3].
RAxML / IQ-TREE	Software	Implements maximum-likelihood phylogenetic inference for estimating gene trees from sequence alignments.	Estimating individual gene trees for each genomic locus, with bootstrap support [3] [4].
PhyloNet	Software	A comprehensive package for inferring and analyzing phylogenetic networks under the Multi-Species Network Coalescent (MSNC) model.	Inferring species networks from multi-locus data while accounting for both ILS and hybridization [3].
Gephi	Software	An open-source network visualization and exploration platform.	Visualizing and conducting cluster analysis on TE sequence similarity networks [5].
OrthoFinder	Software	Infers orthologous groups of genes across multiple species, a critical step in phylogenomics.	Identifying groups of genes that share a common ancestry before phylogenomic analysis [4].
RepeatMasker	Software	Screens DNA sequences for interspersed repeats and low-complexity DNA sequences.	Identifying and classifying transposable elements in genome sequences for network analysis [5].

Comparative Analysis of Reticulation Detection Methods

Choosing the appropriate method is crucial for accurately inferring reticulate evolutionary histories. Different methods have varying strengths, assumptions, and data requirements.

Table 4: Comparison of Methods for Disentangling Reticulation and ILS

Method / Approach	Underlying Model	Primary Use	Strengths	Limitations / Pitfalls
Multispecies Coalescent (MSC) [3]	Statistical coalescent theory within a bifurcating species tree.	Inferring species trees from gene trees while accounting for ILS.	Robust to ILS; widely used and implemented.	Assumes no gene flow; can be misled by reticulation, forcing inaccurate parameter estimates.
Multispecies Network Coalescent (MSNC) [3]	Extends the MSC to operate on a phylogenetic network with reticulation nodes.	Jointly inferring a phylogenetic network and parameters from gene trees, accounting for both ILS and hybridization.	Explicitly models two major sources of incongruence (ILS & reticulation); does not require a known species tree a priori.	Computationally intensive; network space is complex to explore.
Parsimony-based Network Inference [3]	Minimizes the number of reticulation events needed to reconcile a set of gene trees.	Inferring a phylogenetic network that combines all input gene trees.	Intuitive objective (minimize reticulations); can handle large sets of trees.	Does not account for ILS; can overestimate reticulations if ILS is present.
D-statistics (ABBA-BABA)	A test for allele frequency patterns that deviate from a null tree-like model.	Testing for a specific pulse of introgression between four taxa.	Simple, fast, and powerful for testing specific introgression hypotheses.	Limited to a 4-taxon case; does not infer a full network or timing of events.
Network Visualization of TEs [5]	Network science and graph theory applied to sequence similarity.	Visualizing and hypothesizing about the complex evolution of repetitive elements.	Reveals patterns and connections missed by traditional phylogenetics (e.g., convergent evolution).	Primarily descriptive; requires follow-up analyses for statistical validation of hypotheses.

The representation of evolutionary relationships as a bifurcating tree is a foundational concept in biology, tracing back to Darwin. However, the increasing complexity revealed by genomic data now challenges this simplification. Reticulate evolution—the process in which organisms exchange genetic material through mechanisms like hybridization, horizontal gene transfer, and recombination—creates evolutionary pathways that are more accurately represented as interconnected webs or networks rather than simple trees [6]. This perspective is crucial for testing hypotheses in reticulate evolution, as bifurcating models inherently fail to capture the complexity of genomic landscapes shaped by these processes. The "web of life" reflects a more accurate and nuanced history of evolution for many taxa, particularly plants, bacteria, and other organisms prone to hybridization and gene flow [6].

The limitations of trees are not merely theoretical; they have practical consequences for biodiversity research, conservation, and agricultural science. When forced into a treelike structure, evolutionary histories involving gene flow can produce relationships with high uncertainty, even when whole-genome data is available [6]. This article compares the bifurcating tree model against the more flexible phylogenetic network approach, providing the data and methodologies researchers need to objectively evaluate these frameworks within a thesis on reticulate evolution.

Comparative Analysis: Bifurcating Trees vs. Phylogenetic Networks

The following tables summarize the core conceptual and quantitative differences between these two evolutionary models.

Table 1: Conceptual Framework and Applicability Comparison

Aspect	Bifurcating Tree Model	Phylogenetic Network Model
Underlying Structure	Strictly hierarchical, divergent	Reticulate, web-like, allows merging
Representation of Speciation	Assumes lineage-splitting only	Accommodates both splitting and merging via hybridization
Handling of Gene Flow	Cannot represent horizontal gene flow or hybridization	Explicitly represents gene flow and hybridization events
Computational Convenience	Mathematically convenient, computationally less intensive [6]	Computationally challenging, but now feasible [6]
Ideal for Modeling	Vertical descent without gene flow	Reticulate processes like hybridization, recombination, and horizontal gene transfer [6]

Table 2: Quantitative Data and Practical Output Comparison

Feature	Bifurcating Tree Model	Phylogenetic Network Model
Key Output	A single, rooted or unrooted tree	Ancestral Recombination Graph (ARG) or phylogenetic network [7]
Supported Data Formats	Newick, NEXUS, PhyloXML, NeXML [7]	Extended Newick, NEXUS (for annotated networks) [7]
Visualization Tools	FigTree, Dendroscope, MacClade [7]	IcyTree, tanggle, Dendroscope [8] [7]
Impact on Conservation	May misidentify hybrid species as distinct, potentially misallocating resources [6]	Clarifies hybrid origins, aiding in conservation priority-setting [6]
Role in Agriculture	Limited utility for understanding hybrid crops	Clarifies origins of crops like wheat and sweet potato [6]

Experimental Protocols for Inferring and Testing Reticulate Evolution

Genomic Offset as a Measure of Maladaptation

The concept of genomic offset (also known as genomic vulnerability or genetic offset) provides a quantitative method to assess how populations may respond to environmental changes, which is critical in complex landscapes where local adaptation is key [9].

Methodology Overview:

• Data Collection: Collect genomic data (e.g., SNP datasets from sequencing) and environmental data (e.g., bioclimatic variables) from geo-referenced individuals across the species' range.
• Genotype-Environment Association (GEA): Perform a statistical analysis (e.g., using linear models or redundancy analysis) to identify loci putatively under selection by correlating allele frequencies with environmental variables. This step controls for confounding effects like population genetic structure [9].
• Model Fitting: For each putatively adaptive locus, a model is established: ( Yi = \alpha + \beta \cdot E + \varepsilon ), where ( Yi ) is the genetic value and ( E ) is the environmental variable [9].
• Spatial/Temporal Extrapolation: Use the fitted model to predict the genomic composition required for survival in a future climate scenario or a new environment.
• Offset Calculation: The genomic offset is the genetic distance between a population's current genomic composition and the future-predicted composition. A larger offset indicates a higher risk of maladaptation [9].

Limitations: This approach relies on space-for-time substitution and assumes the genotype-environment relationship remains constant. It should be combined with other evidence, such as common garden experiments, for robust conservation recommendations [9].

Phylogenetic Network Inference

This methodology aims to reconstruct evolutionary histories that include reticulate events.

Methodology Overview:

• Data Collection: Generate a whole-genome or reduced-representation sequencing dataset for the taxa of interest.
• Alignment and Site Identification: Align sequences and identify informative genomic regions, including those that may have conflicting phylogenetic signals due to reticulation.
• Model Selection and Analysis: Use specialized software (e.g., Bacter for bacteria) that implements probability models and algorithms capable of inferring networks. These tools use advances in probability theory to calculate the likelihood of a network given the genetic data [6] [7].
• Visualization and Interpretation: Visualize the inferred network using tools like IcyTree or tanggle, which support formats like Extended Newick. Analyze the network for hybrid nodes and patterns of gene flow [8] [7].

Visualizing Complex Evolutionary Relationships

Effective visualization is critical for interpreting the complex relationships depicted by phylogenetic networks. The following workflow diagram, generated using Graphviz, outlines the process of testing for reticulate evolution.

Diagram 1: A workflow for phylogenetic network research. Key steps involve identifying conflict in tree models and using it to infer networks.

The conceptual difference between a tree and a network model is profound. The following diagram illustrates how a network more accurately represents evolutionary history when hybridization occurs.

Diagram 2: A network model correctly shows hybrid species C originating from two ancestral lineages, which a tree model must misrepresent.

Successfully researching reticulate evolution requires a suite of computational tools and resources.

Table 3: Key Research Reagent Solutions for Reticulate Evolution

Tool/Resource	Function/Brief Explanation
IcyTree	A web-based tool for visualizing phylogenetic trees and networks, particularly adept at handling Ancestral Recombination Graphs (ARGs) and supporting Extended Newick format [7].
tanggle	An R package designed for plotting phylogenetic networks and split graphs, integrating with the ggtree ecosystem for customization [8].
Extended Newick Format	A standardized file format for representing phylogenetic networks, including information on hybrid nodes and inheritance probabilities, enabling data interchange between software [7].
BEAST 2 / MrBayes	Software platforms for Bayesian evolutionary analysis, which can generate posterior distributions of trees; conflict among these trees can be a signal of reticulation [7].
Genomic Offset Software	Computational frameworks (e.g., in R) that perform Genotype-Environment Association (GEA) analysis and calculate the genomic offset to predict climate change vulnerability [9].
Whole-Genome Sequencing Data	The foundational empirical data required to detect the genomic signatures of reticulate evolution, such as introgressed regions and conflicting phylogenetic signals [6].

In the field of evolutionary biology, the classic metaphor of the "tree of life" is increasingly being supplemented by the more complex "web of life" [6]. This shift recognizes that evolution is often not a purely divergent process but can involve reticulate processes such as hybridization and gene flow, where genetic material is transferred between different species or populations [6]. Phylogenetic networks have emerged as powerful tools to model these complex evolutionary relationships, and they can be broadly categorized into explicit and implicit networks. These two classes differ fundamentally in their representation forms, computational foundations, and primary applications in biological research.

Explicit networks rely on traditional geometric methods and predefined parameters to represent biological relationships directly. In technical domains, explicit methods are known for using clearly defined, parameterized models to represent structure [10]. In phylogenetics, this translates to networks where nodes and edges have direct, interpretable biological meanings—such as representing species and evolutionary relationships—with explicitly defined probabilities or distances.

In contrast, implicit networks utilize neural representations and machine learning approaches to capture complex patterns without explicitly defining every parameter. Implicit methods can be understood as those that obtain implicit neural representations of scenes through the training of neural networks [10]. Within biodiversity research, these approaches might uncover subtle, non-obvious relationships in genetic data that traditional methods could overlook, effectively learning the structure of evolutionary relationships directly from the data itself.

Conceptual Frameworks: Explicit versus Implicit Networks

Explicit Biological Networks

Explicit networks in biological research are characterized by their transparent structure and direct interpretability. Each component in an explicit network corresponds to a defined biological entity or relationship:

• Nodes represent specific biological entities (e.g., species, genes, proteins)
• Edges represent explicitly defined relationships (e.g., evolutionary descent, protein-protein interactions)
• Parameters have direct biological interpretations (e.g., mutation rates, divergence times)

These networks are computationally convenient and mathematically tractable, making them accessible for researchers to implement and interpret [6]. The explicit representation allows for straightforward hypothesis testing and validation against established biological knowledge.

Implicit Biological Networks

Implicit networks employ learned representations where the relationships are encoded in the parameters of a model rather than being directly specified:

• Pattern recognition from complex data without explicit programming of relationships
• Neural network-based approaches that capture non-linear and emergent properties
• Feature learning that automatically identifies relevant patterns without prior specification

While technically more complex, these methods can reveal relationships that might be missed by explicit models, particularly when dealing with complex genomic data where traditional mathematical models may be insufficient [10].

Table 1: Fundamental Characteristics of Explicit and Implicit Networks

Characteristic	Explicit Networks	Implicit Networks
Representation	Direct parameterization	Learned representation
Interpretability	High	Variable (often "black box")
Computational Demand	Generally lower	Generally higher
Data Requirements	Can work with smaller datasets	Typically requires large datasets
Biological Basis	Directly encoded relationships	Emergent from data patterns
Implementation	Mathematically convenient [6]	Computationally intensive [10]

Methodological Implementation in Phylogenetic Research

Experimental Protocols for Explicit Network Construction

Constructing explicit phylogenetic networks involves clearly defined steps based on traditional geometric and statistical methods:

• Data Collection and Alignment: Obtain molecular sequences (DNA, RNA, or amino acid) from the taxa of interest and perform multiple sequence alignment using tools such as MAFFT or Clustal Omega.

• Distance Matrix Calculation: Compute genetic distances between all pairs of sequences using appropriate substitution models (e.g., Jukes-Cantor, Kimura 2-parameter, or more complex models selected through model testing).

• Network Reconstruction: Apply explicit algorithms such as Neighbor-Net or Minimum Spanning Networks that use the distance matrix to construct the phylogenetic network with explicit splits and conflict representation.

• Parameter Estimation: Calculate explicit parameters for edges and nodes, including bootstrap support values, posterior probabilities, or direct interpretations of genetic distance.

• Visualization and Interpretation: Render the network using visualization tools such as SplitsTree or Dendroscope, with nodes colored according to biological attributes and edges weighted by supported distance measures.

This methodology shares conceptual ground with explicit methods in other fields that obtain geometric structure presented with explicit parameters [10], adapting these principles to evolutionary biological data.

Experimental Protocols for Implicit Network Construction

Implicit phylogenetic network construction employs machine learning approaches to infer relationships:

• Data Preparation and Feature Engineering: Compile genomic datasets and optionally extract features, though many implicit methods can operate directly on raw sequence data.

• Model Architecture Selection: Choose an appropriate neural network architecture (e.g., convolutional neural networks for sequence data, graph neural networks for relational data, or autoencoders for dimensionality reduction).

• Training Phase: Iteratively present data to the model to learn representations, using optimization algorithms (e.g., stochastic gradient descent, Adam) to minimize a loss function that captures the difference between predicted and actual relationships.

• Representation Learning: Allow the network to develop internal representations that capture the complex patterns of evolutionary relationships without explicit programming.

• Network Extraction: Interpret the trained model to extract the implicit phylogenetic relationships, which may involve visualization techniques such as t-SNE or UMAP to project high-dimensional learned representations into interpretable networks.

This approach mirrors implicit methods in neural rendering that obtain implicit neural representation of scenes through the training of neural networks [10], applying similar concept to evolutionary relationships.

Comparative Workflow Diagram

The following diagram illustrates the key methodological differences between explicit and implicit network construction:

Performance Comparison and Applications

Comparative Analysis of Network Performance

The choice between explicit and implicit network approaches involves trade-offs across multiple performance dimensions relevant to phylogenetic research:

Table 2: Performance Comparison of Explicit vs. Implicit Networks in Biological Research

Performance Metric	Explicit Networks	Implicit Networks
Computational Efficiency	Higher efficiency [10]	Lower efficiency; can take hours to optimize [10]
Interpretability	High; direct biological interpretation	Variable; can be "black box"
Handling Reticulate Evolution	Good for known hybridization events	Potentially better for complex or unknown reticulation
Data Efficiency	Effective with smaller datasets	Requires larger datasets for training
Implementation Complexity	Lower; established algorithms	Higher; specialized expertise needed
Accuracy with Known Models	Excellent when model matches reality	May outperform when complex patterns exist
Scalability to Large Genomic Datasets	Can face challenges with very large datasets	Designed to handle large, complex datasets

Applications in Reticulate Evolution Research

Both explicit and implicit networks provide valuable approaches for studying reticulate evolution, each with distinctive strengths:

Explicit networks empower research on:

• Known hybridization events in plants like sunflowers, wheat, and sweet potatoes [6]
• Conservation prioritization where distinct evolutionary units must be identified for protection decisions
• Microendemic species assessment to determine evolutionary independence versus recent hybridization
• Phylogenetic diversity measurements that inform biodiversity conservation strategies

Implicit networks show promise for:

• Uncovering ancient gene-flow events that have created confusion in the tree of life
• Identifying complex reticulation patterns in hyper-diverse groups like pitcher plants
• Revealing non-obvious evolutionary relationships from whole-genome data
• Clarifying uncertain relationships that persist despite extensive genomic data

Research Reagent Solutions and Computational Tools

Implementing explicit and implicit network approaches requires specific computational tools and resources:

Table 3: Essential Research Reagents and Computational Tools for Phylogenetic Network Analysis

Tool/Reagent	Type	Function/Purpose	Implementation Context
Cytoscape	Software Platform	Network visualization and analysis [11]	Both explicit and implicit networks
Splitstree	Specialized Software	Explicit phylogenetic network construction	Primarily explicit networks
Neural Network Frameworks	Computational Libraries	Implement implicit learning approaches	Implicit networks
Phylogenetic Markov Chain Monte Carlo	Statistical Tool	Bayesian inference of evolutionary parameters	Both, particularly explicit
Whole Genome Sequence Data	Biological Data	Primary input for network construction	Both explicit and implicit networks
Multiple Sequence Alignment Tools	Bioinformatics Software	Preprocessing genetic data for analysis	Both explicit and implicit networks
R/Python Visualization Libraries	Programming Tools	Create custom network visualizations [11]	Both explicit and implicit networks

Visualization Guidelines for Biological Networks

Effective visualization is crucial for interpreting and communicating phylogenetic network results. The following diagram illustrates a recommended workflow for creating biological network figures:

Color Selection for Biological Data Visualization

Choosing appropriate color schemes requires understanding your data type [12]:

• Qualitative/Categorical Data: Use distinct colors for unrelated categories (e.g., different biological species)
• Sequential Data: Use color intensity to represent ordered values (e.g., level of gene expression)
• Diverging Data: Use contrasting colors to emphasize deviation from a central point (e.g., fold change in expression)

Always check color accessibility for readers with color vision deficiencies and ensure sufficient contrast between text and background colors [12].

Layout Selection for Network Figures

Different layout strategies support different communicative goals [11]:

• Node-Link Diagrams: Familiar to readers and effective for showing relationships beyond immediate neighbors
• Adjacency Matrices: Superior for dense networks and encoding edge attributes with color
• Fixed Layouts: Position nodes based on external data (e.g., geographic maps)
• Implicit Layouts: Useful for tree structures (e.g., icicle plots, sunburst plots)

Select layouts that minimize unintended spatial interpretations that could lead to misinterpretation of biological relationships [11].

Explicit and implicit phylogenetic networks represent complementary approaches for studying reticulate evolution. Explicit methods provide mathematically convenient, interpretable frameworks that are particularly valuable when biological processes are reasonably well-understood and computational efficiency is important [10] [6]. Implicit methods offer powerful pattern recognition capabilities that can uncover complex, non-obvious relationships in large genomic datasets, though at higher computational cost [10].

The emerging "web of life" perspective in evolutionary biology benefits from both approaches—explicit networks provide biological transparency and validation, while implicit networks offer the potential to discover novel evolutionary patterns without strong prior assumptions. As both methodologies continue to develop, their integration promises a more comprehensive understanding of reticulate evolutionary processes across the tree—or web—of life.

For researchers implementing these approaches, the selection between explicit and implicit networks should be guided by research questions, data characteristics, and computational resources, recognizing that these methods represent different points on a continuum of approaches for understanding biological complexity rather than mutually exclusive alternatives.

The reconstruction of evolutionary histories has traditionally relied on phylogenetic trees, which model divergence through vertical descent. However, comparative genomics has increasingly revealed evolutionary patterns that cannot be explained by strictly treelike relationships. Processes such as hybridization, horizontal gene transfer (HGT), and introgression create reticulate relationships where lineages combine genetic material from multiple ancestors [13] [14]. Phylogenetic networks generalize phylogenetic trees to model these complex histories by incorporating reticulation vertices (or nodes) that represent such merging events [15]. The accurate interpretation of these reticulation vertices and their associated inheritance probabilities (γ) is fundamental to testing hypotheses about reticulate evolution. This guide examines the core components of phylogenetic networks, comparing modeling approaches and their performance in inferring reticulate evolutionary histories.

Core Components of a Reticulate Network

Reticulation Vertices

In graph-theoretic terms, a (rooted binary) phylogenetic network is a rooted, directed, acyclic graph (DAG). Within this structure, most vertices are tree vertices (in-degree one, out-degree two). The key distinguishing features are the reticulation vertices (or hybrid nodes), which have an in-degree of two and an out-degree of one [15] [14]. The two edges directed into a reticulation vertex are called reticulation edges.

• Biological Interpretation: A reticulation vertex represents an evolutionary event where two distinct lineages contribute genetic material to form a single descendant lineage. The specific biological process can include:
  • Hybrid speciation: The formation of a new species through interbreeding between two existing species [14].
  • Horizontal Gene Transfer (HGT): The movement of genetic material from one organism to another that is not its offspring [14].
  • Recombination: The exchange of genetic material between similar sequences [14].
• Level of a Network: The level of a phylogenetic network is a measure of its complexity, defined as the maximum number of reticulation vertices in a biconnected component (a blob) of the network's skeleton [15]. Level-1 networks are a key class where no vertex belongs to more than one cycle in the network's skeleton, meaning blobs are disjoint cycles [15]. This class offers a good balance between biological realism and mathematical tractability.

Inheritance Probabilities (γ)

To each reticulation edge e, the model assigns an inheritance probability, denoted by γ(e) [14]. These parameters quantify the proportional genetic contribution from each parent lineage involved in a reticulate event. Formally, for the two reticulation edges, e1 and e2, that lead into the same reticulation vertex, the inheritance probabilities must satisfy the constraint: γ(e1) + γ(e2) = 1 [14].

• Biological Interpretation: The value γ(e) represents the expected fraction of genetic material in the hybrid lineage that was inherited via the reticulation edge e from that specific parent. An inheritance probability of 0.5 suggests an equal contribution from both parents, as might be expected in a symmetric hybridization. Values skewed towards 0 or 1 may indicate an introgression event where a small fraction of the genome was transferred.

• Role in Likelihood Calculations: Under the maximum likelihood (ML) framework for phylogenetic networks, the probability of observing a specific gene tree T, given a network N and its inheritance probabilities γ, is defined as [14]: P(T | N, γ) = Σ_{η(T) ∈ I(T)} Π_{e ∈ η(T)} γ(e) Here, I(T) is the collection of all possible induction sets of tree T within network N—the sets of reticulation edges that, when kept, lead to the gene tree T. The overall likelihood of sequence data S is then computed by integrating over all possible gene trees contained within the network [14].

Comparative Analysis of Phylogenetic Network Classes

Different classes of phylogenetic networks impose constraints on their structure to balance biological realism, mathematical tractability, and computational feasibility. The following table compares several key classes referenced in the literature.

Table 1: Comparison of Major Phylogenetic Network Classes

Network Class	Key Structural Constraint	Biological Interpretation & Rationale	Mathematical & Computational Properties
Level-1 Networks [15]	No vertex is in more than one biconnected component (blob). Blobs are disjoint cycles.	Models reticulate events that do not overlap complexly. A foundational, tractable model for well-separated hybridization/HGT.	The network parameter is generically identifiable under Markov models for triangle-free, level-1 networks [15].
Tree-Child (TC) Networks [16]	Every internal node has at least one child that is a tree node.	Ensures every extinct or hypothetical ancestor has a lineage evolving only through mutation, prohibiting "stacked" reticulations.	One of the most studied and permissive classes; allows efficient generation algorithms [16].
Normal Networks [17]	A subclass of tree-child networks with additional constraints on the placement of reticulations.	Proposed as a class sitting in the "sweet spot" between biological relevance and mathematical tractability.	Emerging as a leading class due to a strong combination of identifiability results and biological plausibility [17].
Orchard Networks [16]	Defined by a specific sequence of reduction operations ("cherry-picking").	Another "well-behaved" class with a clear combinatorial structure.	Allows efficient, injective generation algorithms [16].

Maximum Likelihood Inference and Performance

The Maximum Likelihood Framework

The ML framework for inferring phylogenetic networks from sequence data aims to find the network topology, branch lengths, and inheritance probabilities γ that maximize the probability of observing the given sequence alignments [14]. For a set of genes (loci) ( S = {S1, S2, ..., Sk} ), the likelihood function is [14]: [ L(N, \gamma | S) = \prod{Si \in S} \sum{T \in T(N)} \left[ P(Si | T) \cdot P(T | N, \gamma) \right] ] Here, ( P(Si | T) ) is the standard phylogenetic likelihood of the sequence alignment for gene i given a gene tree T, and ( P(T | N, \gamma) ) is the probability of the gene tree given the network, as defined in Section 2.2.

Performance and Identifiability of Reticulation Events

The performance of ML in correctly identifying reticulation events is influenced by several biological and methodological factors.

• Evolutionary Diameter and Height: The evolutionary diameter of a reticulation event—the phylogenetic distance between the donor and recipient lineages—has a significant effect on its detectability. Reticulations with a larger diameter are generally more difficult to identify accurately [14].
• Inheritance Probability (γ): The value of γ itself impacts performance. Events with extreme inheritance probabilities (close to 0 or 1) can be more challenging to detect than those with balanced probabilities [14].
• Generic Identifiability: A key theoretical guarantee for model-based methods is that the semi-directed network parameter of a level-1 network is generically identifiable under common Markov models of sequence evolution (e.g., Jukes-Cantor, Kimura 2-parameter) [15]. This means that, for such networks, the data almost surely determines the correct network topology and reticulation vertices, preventing confusion between different network structures.

The Model Selection Problem

A fundamental challenge is that more complex networks (with more reticulations) will always fit the data at least as well as simpler ones, risking overfitting [14]. Information criteria are used to penalize model complexity and select the simplest network that adequately explains the data.

• Bayesian Information Criterion (BIC): Simulation studies have shown that BIC performs very well in controlling model complexity and preventing ML from grossly overestimating the number of reticulation events [14]. Its formula is ( BIC = -2 \ln L + K \ln n ), where ( K ) is the number of parameters, ( L ) is the maximized likelihood, and ( n ) is the sample size.
• Akaike Information Criterion (AIC): While also used, AIC has been found to be less effective than BIC at preventing overestimation of reticulation in this context [14]. Its formula is ( AIC = 2K - 2 \ln L ).

Table 2: Impact of Factors on Reticulation Detectability and Model Selection

Factor	Impact on Inference Performance	Supporting Experimental Evidence
Evolutionary Diameter	High diameter significantly reduces detectability and placement accuracy of reticulation edges [14].	Simulation studies analyzing ML performance under different diameter conditions [14].
Inheritance Probability (γ)	Extreme values (near 0/1) can reduce detectability compared to balanced values [14].	Analyses of ML accuracy in estimating γ and locating the corresponding edges [14].
Number of Loci/Genes	A larger number of independent loci improves the accuracy of the inference [14].	Simulation studies using multi-locus datasets [14].
Model Selection Criterion	BIC effectively controls overfitting; AIC can lead to overestimation of reticulations [14].	Comparisons of inferred vs. true number of reticulations under AIC and BIC [14].

Experimental Protocols for Key Methodologies

Protocol 1: Maximum Likelihood Inference with Information Criteria

This protocol outlines the core methodology for inferring a phylogenetic network from sequence data using maximum likelihood, as derived from simulation studies [14].

• Input Data Preparation: Assemble a set of ( k ) multiple sequence alignments ( S = {S1, S2, ..., S_k} ), each corresponding to a non-recombining genomic region (e.g., a gene).
• Model Assumption Selection: Choose an underlying nucleotide substitution model (e.g., Jukes-Cantor, Kimura 2-parameter) for calculating the tree likelihoods ( P(S_i | T) ).
• Likelihood Calculation: For a candidate network ( N ) and inheritance probabilities ( \gamma ), compute the composite likelihood ( L(N, \gamma | S) ) using the formula in Section 4.1.
• Parameter Optimization: Use numerical optimization methods to find the network topology, branch lengths, and inheritance probabilities that maximize ( L(N, \gamma | S) ). This is typically a heuristic search due to the vast space of possible networks.
• Model Selection: For a set of candidate networks with differing numbers of reticulations (e.g., ( r = 0, 1, 2, ... )), compute the BIC score for each. The network with the lowest BIC score is selected as the preferred model.

Protocol 2: Testing Generic Identifiability for Level-1 Networks

This protocol summarizes the theoretical and combinatorial approach used to establish the generic identifiability of level-1 networks, a crucial result for justifying model-based methods [15].

• Define the Model Class: Restrict the analysis to the class of ( n )-leaf, triangle-free, level-1 semi-directed networks with ( r ) reticulation vertices.
• Algebraic Parameterization: Associate the network-based Markov model with an algebraic variety. The goal is to show that the map from the parameter space (network topology, branch lengths, γ) to the probability distributions of site patterns is injective, except on a subset of measure zero.
• Leverage Small Network Results: Use established algebraic identifiability results for small networks (e.g., networks with a single reticulation vertex and a cycle of length ≥4) as foundational building blocks [15].
• Combinatorial Decomposition: Prove that any larger, triangle-free level-1 network can be uniquely decomposed or distinguished based on the properties of its smaller subnetworks, using combinatorial arguments about their graph structure [15].
• Conclusion: Combine the algebraic results for small building blocks with the combinatorial decomposition to conclude that the network parameter is generically identifiable for the entire defined class [15].

Visualizing Network Inference and Structure

Workflow for Phylogenetic Network Inference

The following diagram illustrates the primary workflow for inferring phylogenetic networks via maximum likelihood, incorporating model selection to avoid overfitting.

Diagram 1: Maximum likelihood phylogenetic network inference workflow.

Core Components of a Reticulate Network

This diagram deconstructs the essential elements of a phylogenetic network, highlighting the different vertex types and the critical inheritance probabilities on reticulation edges.

Diagram 2: Core components of a phylogenetic network with a single reticulation vertex.

Table 3: Essential Software and Analytical Tools for Network Inference

Tool / Resource	Type	Primary Function in Analysis	Relevance to Reticulation Vertices & γ
PhyloNet [18] [15]	Software Package	A toolbox for inferring and analyzing phylogenetic networks.	Implements algorithms for inferring networks and calculating inheritance probabilities from gene trees or sequences.
SNaQ(Solís-Lemus & Ané, 2016) [15]	Inference Algorithm	A maximum likelihood method for inferring phylogenetic networks under the network multispecies coalescent.	Estimates network topology and γ values; relies on identifiability results for level-1 networks.
NANUQ(Allman et al., 2019) [15]	Inference Algorithm	A method for inferring phylogenetic networks from quartet distances.	Used to reconstruct semi-directed level-1 networks, the parameter class proven to be identifiable.
Information Criteria (BIC) [14]	Statistical Criterion	A model selection criterion that penalizes model complexity.	Critical for selecting the correct number of reticulation vertices and avoiding model overfitting.
Tree-Child Network Generators [16]	Algorithmic Tool	Algorithms for systematically generating all possible tree-child networks with a given number of leaves.	Useful for exploring the space of plausible network hypotheses and for theoretical combinatorial studies.

The Biological Significance of Reticulation in Speciation and Trait Diversity

For decades, the phylogenetic tree has served as the primary model for representing evolutionary relationships among species. However, accumulating genomic evidence reveals that evolution is often not strictly treelike, particularly in plants, bacteria, and viruses where hybridization, horizontal gene transfer (HGT), and interspecific recombination are widespread [6]. Reticulate evolution—characterized by the merging of lineages—creates a complex "web of life" that more accurately captures the evolutionary history of many taxa [6]. This paradigm shift from trees to networks empowers researchers to better understand biodiversity patterns, trait evolution, and species boundaries with significant implications for conservation prioritization and drug discovery from biological resources.

Phylogenetic networks are rooted, directed, acyclic graphs that extend tree models by incorporating non-vertical inheritance of genetic material [19]. They provide a mathematical framework for modeling hybridization, HGT, and recombination events through reticulation nodes (nodes with in-degree 2 and out-degree 1) and associated inheritance probabilities (γ) that quantify the genetic contribution from each parent lineage [19]. This model flexibility allows researchers to detect and quantify historical gene flow, providing crucial insights into evolutionary mechanisms driving trait diversity across species radiations.

Performance Comparison: Phylogenetic Networks Versus Traditional Trees

The comparative analysis between phylogenetic networks and traditional trees reveals significant differences in their methodological approaches, performance characteristics, and biological insights. The table below summarizes key quantitative comparisons between these approaches based on empirical studies and simulation tests.

Table 1: Performance comparison between phylogenetic networks and traditional trees

Performance Metric	Phylogenetic Networks	Traditional Phylogenetic Trees
Model Complexity	Higher: Accounts for both vertical descent and horizontal gene flow [19]	Lower: Assumes strictly divergent evolution [6]
Computational Demand	High: Requires probability theory and significant computational resources [6]	Moderate: Established efficient algorithms [6]
Gene Tree Incongruence	Explicitly models as reticulation events [19]	Treats as statistical noise or incomplete lineage sorting
Trait Diversity Prediction	Superior: Explains trait patterns from hybridization events [6]	Limited: May misinterpret convergent evolution [20]
Reticulation Detection Power	Dependent on diameter, inheritance probability, and number of genes [19]	Unable to detect hybridization/HGT by design
Model Selection	Uses BIC/AIC to prevent overparameterization [19]	Uses likelihood ratio tests or information criteria

Key Performance Differentiators

• Biological Accuracy: Phylogenetic networks provide more accurate representations for groups with known hybridization, such as sunflowers, wheat, sweet potatoes, and pitcher plants, where treelike models have historically struggled to resolve certain relationships despite extensive genomic data [6].

• Parameter Sensitivity: The detectability of reticulation events depends heavily on their evolutionary diameter (phylogenetic distance between donor and recipient) and the number of genes transferred, with larger diameters and more genes increasing detection power [19].

• Model Selection Performance: The Bayesian Information Criterion (BIC) effectively controls model complexity in network inference, preventing overestimation of reticulation events while maintaining detection power for biologically significant gene flow [19].

Experimental Frameworks for Reticulate Evolution Research

Maximum Likelihood Framework for Phylogenetic Network Inference

The maximum likelihood (ML) framework for inferring phylogenetic networks from molecular sequence data extends the standard phylogenetic tree likelihood model to account for both mutational processes within genomic regions and reticulation across regions [19].

Likelihood Function: The likelihood of a phylogenetic network ( N ) with inheritance probabilities ( \gamma ), given a set of sequence alignments ( S = {S1, S2, ..., S_k} ) from ( k ) non-recombining genomic regions, is given by:

[ L(N,\gamma|S) = \prod{Si \in S} \sum{T \in T(N)} [P(Si|T) \cdot P(T|N,\gamma)] ]

where:

• ( P(Si|T) ) represents the standard tree likelihood for alignment ( Si ) given tree ( T )
• ( P(T|N,\gamma) ) denotes the probability of gene tree ( T ) within network ( N ) with inheritance probabilities ( \gamma )
• ( T(N) ) represents the set of all trees contained within network ( N ) [19]

Inheritance Probability Estimation: For a reticulation node with incoming edges ( e1 ) and ( e2 ), the inheritance probabilities ( \gamma(e1) ) and ( \gamma(e2) ) are estimated from the data and satisfy ( \gamma(e1) + \gamma(e2) = 1 ), representing the proportional genetic contribution from each parent lineage [19].

Table 2: Information criteria for model selection in phylogenetic network inference

Information Criterion	Formula	Performance in Network Inference
Akaike Information Criterion (AIC)	( AIC = 2K - 2\ln L ) [19]	Moderate complexity control
Bayesian Information Criterion (BIC)	( BIC = K\ln n - 2\ln L ) [19]	Superior performance in preventing overestimation of reticulations

Experimental Workflow for Reticulate Evolution Detection

The following workflow diagram illustrates the key steps in detecting and verifying reticulate evolution using phylogenetic networks:

Eco-Evolutionary Model of Trait Variance

The relationship between species diversity and functional diversity is modulated by reticulate evolutionary processes. An eco-evolutionary model integrating quantitative genetics and species interactions demonstrates how trait variances evolve in response to competitive pressures:

Model Components:

• State Variables: Population density, trait means, and genetic variances for each species
• Selection Mechanism: Phenotypes experience differential growth based on position in trait space and competition for shared resources
• Distribution Assumption: Trait distributions follow normal distributions with evolvable means and variances [20]

Key Finding: In species-rich communities, increased competition drives the evolution of narrower intraspecific trait breadths as species minimize niche overlap. This can paradoxically reduce functional diversity despite higher species richness [20]. The following diagram illustrates this counterintuitive relationship:

Table 3: Essential research reagents and computational tools for reticulate evolution research

Research Tool	Specification/Function	Application Context
Non-recombining Genomic Regions	Multiple independent loci or whole genomes	Provide phylogenetic signal for detecting incongruence [19]
Inheritance Probability (γ)	Estimates proportion of genetic material from each parent	Quantifies strength of reticulation events [19]
Hidden State Speciation & Extinction (HiSSE) Models	Accounts for unmeasured traits affecting diversification	Controls for hidden variables in trait-dependent diversification [21]
Bayesian Information Criterion (BIC)	Penalizes model complexity; K·ln n - 2ln L	Prevents overestimation of reticulation events [19]
Maximum Likelihood Framework	Computes probability of data given network model	Estimates network topology and parameters [19]
Binary State Speciation & Extinction (BiSSE)	Models trait-dependent diversification	Baseline comparison for HiSSE models [21]

Discussion: Implications for Biodiversity Research and Conservation

The adoption of phylogenetic networks has profound implications for interpreting biodiversity patterns and establishing conservation priorities. Network approaches reveal that what appear to be evolutionarily distinct units based on tree models may actually represent recent hybridization products or populations with ongoing gene flow [6]. This clarification is particularly valuable for assessing conservation status of microendemic species with limited ranges and small population sizes, where accurate classification directly impacts resource allocation decisions [6].

In agricultural and pharmaceutical research, understanding reticulate evolution provides insights into the genetic origins of valuable traits in crop plants like wheat and sweet potato, which have experienced ancient hybridization events [6]. Identifying the genomic contributions from different parental lineages through network analysis facilitates more targeted breeding strategies and gene discovery efforts for medically relevant compounds from plants.

The counterintuitive finding that species richness does not necessarily increase functional diversity [20] underscores the importance of incorporating eco-evolutionary dynamics into biodiversity assessments. This relationship has practical implications for predicting ecosystem responses to species loss and understanding how trait diversity accumulates over evolutionary timescales.

Methodologies in Practice: A Guide to Inferring and Applying Phylogenetic Networks

The Network Multispecies Coalescent (NMSC) is an advanced statistical framework that extends the Multispecies Coalescent (MSC) model to explicitly account for reticulate evolutionary processes, such as hybridization and introgression, alongside incomplete lineage sorting (ILS). The MSC model itself provides a powerful framework for inferring species phylogenies by integrating the phylogenetic process of species divergences with the population genetic process of coalescence, effectively modeling gene tree-species tree discordance [22]. The NMSC builds upon this foundation by incorporating non-tree-like evolutionary events, recognizing that the history of life cannot always be properly represented as a tree, particularly in groups with extensive hybridization [23] [24].

This framework has emerged as a critical tool in phylogenomics, as it simultaneously models multiple sources of gene tree incongruence, including both ILS and hybridization, which often co-occur when closely related species are capable of exchanging genetic material [25]. The NMSC provides a probabilistic approach for analyzing genomic sequence data from multiple species, enabling researchers to infer species networks rather than being constrained to strictly bifurcating trees. This is particularly relevant in plant evolution and other groups where reticulate evolution is widespread [23] [24].

Table: Key Concepts in the NMSC Framework

Concept	Description	Biological Significance
Reticulate Evolution	Evolutionary pattern involving network-like relationships due to hybridization or gene flow	Explains non-tree-like evolutionary histories in many taxonomic groups
Incomplete Lineage Sorting (ILS)	Failure of gene lineages to coalesce in a population, carrying ancestral polymorphisms forward	Causes gene tree-species tree discordance even without hybridization
Gene Tree Incongruence	Different gene trees having different topologies across genomic loci	Can result from ILS, hybridization, or other biological processes
Anomalous Gene Trees (AGTs)	Gene trees that are more probable than the gene tree matching the species tree	Can occur with postspeciation gene flow and high ILS

Theoretical Foundations and Model Comparison

From MSC to NMSC: Theoretical Extension

The Multispecies Coalescent (MSC) forms the foundational model for the NMSC, describing the genealogical relationships of DNA sequences sampled from multiple species [22] [26]. Under the MSC, gene trees are modeled as evolving within a species tree, with coalescence events occurring backward in time within populations. The model involves parameters for species divergence times (τ) and population sizes (θ), measured in expected mutations per site [22]. The MSC predicts that gene trees can differ from the species tree due to the stochastic nature of the coalescent process, particularly when internal branches of the species tree are short relative to population sizes [22] [26].

The NMSC extends this model by incorporating reticulation nodes that represent historical hybridization or introgression events. These nodes allow lineages to originate from multiple parental species, creating a network structure rather than a strictly bifurcating tree [25]. This extension enables the model to account for gene tree discordance that results from both ILS and hybridization, providing a more comprehensive framework for analyzing genomic data from groups with complex evolutionary histories [25].

Comparative Framework: NMSC vs. Alternative Models

Table: Comparison of Phylogenetic Statistical Frameworks

Model/Framework	Handles ILS?	Handles Hybridization?	Key Assumptions	Primary Applications
Network MSC (NMSC)	Yes	Yes	Known species network topology; isolation after divergence with possible historical hybridization events	Species network inference; hybridization detection; parameter estimation with gene flow
Multispecies Coalescent (MSC)	Yes	No	Known species tree topology; complete isolation after divergence	Species tree estimation; divergence time and population size estimation; species delimitation
Isolation-with-Migration	Limited	Yes (continuous)	Continuous gene flow over limited time periods	Phylogeography; population divergence with ongoing gene flow
Concatenation	No	No	Gene tree heterogeneity is noise; all sites share same evolutionary history	Species tree estimation when ILS is minimal
Structured Coalescent	Yes	Indirectly	Population structure with migration	Phylogeography; population structure inference

The NMSC differs fundamentally from the standard MSC in its ability to model gene flow events explicitly through reticulation nodes. While the MSC assumes complete isolation after species divergence, the NMSC allows for historical hybridization at specific points in evolutionary history [25]. This distinction is crucial when analyzing groups where hybridization has played a significant role, as the MSC may incorrectly attribute hybridization-induced discordance to ILS alone.

Compared to the Isolation-with-Migration model, which assumes continuous gene flow over specified periods, the NMSC typically models hybridization as discrete events, making it more suitable for scenarios involving sporadic hybridization rather than continuous migration [22] [25]. The NMSC also differs from population genetic models that use summary statistics like allele frequencies and SNPs to infer demographic processes, as it works directly with sequence data and explicitly models genealogical histories [22].

Key Methodological Approaches and Experimental Protocols

Inference Methods and Their Implementation

Several methodological approaches have been developed for inference under the NMSC framework, each with distinct strengths and computational requirements. Full-likelihood methods (both maximum likelihood and Bayesian) offer the best statistical properties but are computationally intensive [22]. These methods compute the probability of the sequence data given a species network and model parameters, integrating over all possible gene trees and coalescent histories [25].

Pseudolikelihood approaches applied to subnetwork summary statistics provide a computationally efficient alternative, enabling analysis of larger datasets [27]. These methods decompose the network into smaller components, analyze them separately, and combine the results. Another strategy involves inference of small subnetworks combined with combinatorial network building, which can handle more complex evolutionary scenarios while managing computational complexity [27].

Recent advances have also addressed the identifiability properties of the NMSC model. Rhodes (2023) demonstrated that the "Tree of Blobs" of a species network - where all biconnected components are collapsed to nodes - is identifiable regardless of network structure, and developed a consistent algorithm for its inference [27]. This represents a significant theoretical advancement, as identifiability issues have previously complicated network inference under the NMSC.

Experimental Design and Data Requirements

Robust inference under the NMSC requires careful experimental design and appropriate data collection strategies:

• Locus Selection: Data should consist of sequence alignments from hundreds or thousands of independent loci, ideally short segments sampled from far-apart genomic regions to ensure independent coalescent histories [22]. Non-recombining loci or very short sequences where recombination is unlikely are preferred, as all sites within a locus must share the same gene tree [26].

• Taxon Sampling: Multiple individuals per species are recommended when possible, as this provides information about population sizes and improves parameter estimation [26]. Dense sampling across the group of interest helps distinguish between different sources of discordance.

• Genomic Resources: The unprecedented increase in genomic data availability has been crucial for NMSC applications, as genome-scale data provide both the necessary number of independent loci and the resolution to detect reticulate events [28].

Workflow for NMSC-based Phylogenomic Analysis

Distinguishing Hybridization from Incomplete Lineage Sorting

A critical methodological challenge in NMSC analysis is distinguishing hybridization from other sources of gene tree incongruence, particularly ILS. The NMSC framework provides several approaches to address this challenge:

• Asymmetry in Discordant Topologies: Under the pure MSC model for a three-species case, the two discordant gene trees are equally probable. Significant deviation from this equal probability provides evidence against the MSC null hypothesis and can indicate hybridization [25].

• Site Pattern Frequencies: Methods based on invariants in site pattern probabilities can detect hybridization by analyzing the distribution of site patterns across loci [25].

• Integration of Multiple Data Types: Combining information from gene tree topologies, branch lengths, and coalescent times provides more power to distinguish between hybridization and ILS than using any single source of information [22] [25].

It is important to note that hybridization and ILS are not mutually exclusive and often occur together. High levels of ILS can actually be beneficial for inferring hybridization, as when lineages fail to coalesce, they can trace multiple paths through a network topology, providing information about how often lineages come from different ancestors [25].

Research Reagent Solutions for NMSC Studies

Table: Essential Research Tools for NMSC Analysis

Research Tool	Function/Purpose	Example Implementations
Sequence Alignment Software	Multiple sequence alignment of loci	MAFFT, MUSCLE, Clustal Omega
Gene Tree Estimation Packages	Inferring gene trees from sequence data	RAxML, IQ-TREE, BEAST
NMSC Inference Software	Species network inference under coalescent	SNaQ, BPP, PhyloNet
Model Testing Frameworks	Comparing MSC vs. NMSC models	AIC, BIC, likelihood ratio tests
Data Visualization Tools	Visualizing networks and gene trees	Dendroscope, IcyTree, FigTree

The research reagents required for NMSC studies extend beyond traditional laboratory supplies to include specialized software and analytical tools. Computational resources for handling genome-scale datasets are essential, as NMSC analyses typically involve processing thousands of loci across multiple taxa [22] [25]. High-performance computing clusters are often necessary for full-likelihood Bayesian implementations, while pseudolikelihood approaches can be run on standard workstations for smaller datasets [27].

For researchers designing NMSC studies, several practical considerations are crucial. The selection of appropriate molecular markers should prioritize non-recombining loci or very short sequences where recombination is minimal, as the basic NMSC model assumes no recombination within loci [26]. The use of biparentally inherited markers is necessary to detect hybridization signals, in contrast to tree-based analyses that often use uniparentally inherited markers to avoid complications [23]. Additionally, reference genomes for the studied species can greatly enhance locus selection and orthology assessment, particularly when designing probes for targeted sequencing approaches.

Current Challenges and Research Frontiers

Despite significant advances, several challenges remain in the application and development of the NMSC framework. Computational complexity continues to limit the application of full-likelihood methods to large datasets or complex networks [27]. Bayesian inference has been effective only on relatively small problems, prompting the development of approximate methods [27].

Model identifiability represents another significant challenge, as different network structures and parameter combinations can sometimes produce similar patterns in gene tree distributions [27] [25]. This is particularly problematic when trying to distinguish hybridization from other processes such as ancestral population structure, which can produce similar asymmetries in gene tree frequencies [25].

Recent research has begun to extend the NMSC framework to new data types and evolutionary questions. The development of multispecies coalescent models for quantitative traits allows for the integration of phenotypic data while accounting for genealogical discordance [28]. Methods for analyzing trait evolution on networks rather than trees are also emerging, promising to expand comparative methods beyond strictly bifurcating phylogenies [25].

Future directions in NMSC research include the development of more efficient inference algorithms, improved methods for assessing statistical support for reticulation events, and approaches for integrating additional biological processes such as recombination and continuous gene flow into the model framework [22] [25]. As these methodological advances progress, the NMSC is poised to become an increasingly powerful framework for unraveling the complex web of evolutionary relationships in groups with reticulate evolutionary histories.

Scalable Computational Tools for Network Inference from Genome-Scale Data

The paradigm for understanding evolutionary history is shifting from a simple "tree of life" to a complex "web of life," driven by the recognition that reticulate evolution—processes like hybridization and gene flow—plays a fundamental role in shaping biodiversity [6]. This shift necessitates computational tools capable of inferring phylogenetic networks rather than simple trees. Simultaneously, in biomedical research, accurately mapping gene regulatory networks (GRNs) is crucial for understanding disease mechanisms and identifying therapeutic targets [29]. Both fields face a common challenge: inferring complex network structures from genome-scale data in a way that is both biologically accurate and computationally scalable. This guide provides an objective comparison of current computational methods for network inference, evaluating their performance, scalability, and applicability to reticulate evolution research.

Comparative Performance Analysis of Network Inference Methods

Benchmarking Frameworks and Key Metrics

Evaluating network inference methods requires robust benchmarking frameworks that use real-world data and biologically meaningful metrics. CausalBench is a prominent benchmark suite that uses large-scale single-cell perturbation data with over 200,000 interventional data points [29]. It employs two primary evaluation types:

• Biology-driven evaluation: Approximates ground truth through biological knowledge.
• Statistical evaluation: Uses causal effect estimators, including:
  • Mean Wasserstein Distance: Measures the strength of predicted causal effects.
  • False Omission Rate (FOR): Measures the rate at which true causal interactions are missed by the model [29].

Another common approach uses Receiver Operating Characteristic (ROC) curves and Precision-Recall (PR) curves, which evaluate performance across all possible prediction thresholds by comparing true positive rates against false positive rates [30].

Performance Comparison of State-of-the-Art Methods

The table below summarizes the performance of various network inference methods based on comprehensive benchmarking studies:

Table 1: Performance Comparison of Network Inference Methods on Real-World Single-Cell Data

Method	Type	Key Characteristics	Performance Summary
Mean Difference [29]	Interventional	Top-performing in CausalBench challenge	High on statistical evaluation (Mean Wasserstein-FOR trade-off)
Guanlab [29]	Interventional	Top-performing in CausalBench challenge	High on biological evaluation
GRNBoost [29]	Observational	Tree-based GRN inference	High recall but low precision
NOTEARS [29]	Observational	Continuous optimization with acyclicity constraint	Extracts limited information from data
PC [29]	Observational	Constraint-based causal discovery	Extracts limited information from data
GES/GIES [29]	Observational/Interventional	Score-based greedy equivalence search	Extracts limited information from data
Betterboost & SparseRC [29]	Interventional		Perform well on statistical but not biological evaluation
Pearson Correlation [30]	Observational	Simple statistical dependence	Moderate accuracy, better than random but far from perfect

Performance analysis reveals a consistent trade-off between precision and recall across methods [29]. Some methods achieve high recall (identifying many true interactions) but suffer from low precision (many false positives), while others exhibit the opposite pattern. Notably, benchmarking on real biological data has shown that methods using interventional data do not consistently outperform those using only observational data, contrary to results from synthetic benchmarks [29].

Scalability and Computational Efficiency

Scalability remains a significant limitation for many traditional methods. The poor scalability of existing approaches often limits their performance on large, genome-scale datasets [29]. However, newer methods like RAMEN (Random walk- and genetic algorithm-based network inference) demonstrate advantages in computational efficiency and scalability compared to conventional approaches [31]. RAMEN integrates absorbing random walks and genetic algorithms to efficiently learn Bayesian network structures while focusing on disease-relevant variables.

Experimental Protocols for Method Evaluation

Standardized Benchmarking Methodology

To ensure fair comparisons, benchmarking studies follow standardized protocols:

• Data Splitting: A non-standard data split is used where no perturbation condition appears in both training and test sets.
• Handling of Directly Targeted Genes: The directly perturbed gene is set to zero (for knockout) or its observed post-intervention value to avoid illusory success.
• Multiple Random Seeds: Methods are typically trained multiple times (e.g., five times) with different random seeds to ensure robustness [29] [32].

Synthetic Data Generation for Validation

When ground truth networks are unknown, synthetic data generation tools like Biomodelling.jl create realistic scRNA-seq data with known underlying network topology [33]. This approach allows for controlled benchmarking by:

• Simulating stochastic gene expression in growing and dividing cells
• Incorporating global transcription-cell volume relationships
• Modeling binomial partitioning of molecules during cell division
• Including capture efficiency of scRNA-seq protocols

Research Reagent Solutions for Network Inference

Table 2: Essential Research Reagents and Computational Tools for Network Inference

Item/Tool	Function	Application Context
CausalBench Suite [29]	Benchmarking network inference methods with real-world data	General GRN inference; single-cell perturbation studies
PEREGGRN Platform [32]	Evaluating expression forecasting performance	Prediction of genetic perturbation effects on transcriptomes
Biomodelling.jl [33]	Generating synthetic scRNA-seq data with known ground truth	Method validation and benchmarking
GGRN Software [32]	Forecasting expression based on candidate regulators	GRN-based expression prediction
Phylogenetic Network Tools [34] [17]	Inferring evolutionary relationships with reticulation	Biodiversity research, reticulate evolution studies
Single-Cell Perturbation Data [29]	Providing intervention and control measurements	Causal network inference in cellular systems

Network Inference Workflows and Method Relationships

The following diagram illustrates the typical workflow for benchmarking network inference methods and the relationships between different methodological approaches:

Diagram 1: Network inference workflow and method relationships. This workflow shows how different data types feed into various methodological approaches, which are then evaluated through standardized benchmarking frameworks. The results inform applications in both gene regulatory network inference and reticulate evolution research.

Implications for Reticulate Evolution Research

The Emerging Significance of Phylogenetic Networks

The field of phylogenetics is undergoing a fundamental transformation as researchers recognize that reticulate evolution—including hybridization and gene flow—is a key mechanism contributing to genetic and trait diversity [34]. Phylogenetic networks provide a biologically intuitive approach to depicting evolutionary processes that cannot be represented by simple trees, such as:

• Hybrid speciation
• Introgressive hybridization
• Historical gene flow [34]

These networks are particularly crucial for groups of conservation concern that lack reference genome resources and explicit hypotheses from prior investigation [34].

Methodological Advances in Phylogenetic Network Inference

Recent computational advances have made phylogenetic network inference more feasible. While family trees have been mathematically convenient and computationally tractable, new probability theory approaches and computational advances now enable researchers to estimate the likelihood of network structures [6]. Among network classes, normal networks are emerging as a leading contender, sitting in the "sweet spot between biological relevance and mathematical tractability" [17].

However, practical challenges remain. Due to the massively larger search space when going from trees to networks, researchers often must compromise by using reduced sets of samples for network analysis [35]. There is a continuing need for software that can feasibly identify reticulations for problems involving hundreds of taxa or more.

Applications in Biodiversity and Conservation

Phylogenetic networks are influencing conservation biology and biodiversity research by:

• Clarifying species relationships: Networks can resolve uncertainties in plant phylogenies that persist even with whole-genome data [6].
• Informing conservation priorities: Networks help distinguish between long-term evolutionarily independent units and recent hybridization events, aiding conservation resource allocation [6].
• Understanding evolutionary history: For groups like pitcher plants or Tillandsia, understanding ancient gene-flow events clarifies species distributions and diversification processes [6] [35].

The field of network inference is rapidly evolving, with significant advances in both gene regulatory network inference and phylogenetic network reconstruction. Performance benchmarking reveals that while simple methods sometimes outperform complex ones, newer approaches are steadily improving scalability and accuracy. For researchers studying reticulate evolution, phylogenetic networks now offer viable tools to uncover the complex web of life, moving beyond the limitations of traditional tree-based approaches. As these computational tools continue to develop and benchmark, they promise to provide increasingly powerful insights into both evolutionary processes and disease mechanisms.

Leveraging DNA Language Models and k-mers for Efficient Phylogenetic Placement

The paradigm for understanding evolutionary relationships is shifting from a strictly bifurcating "Tree of Life" to a more complex and accurate "Web of Life" [6]. This new perspective acknowledges the prevalence of reticulate evolutionary processes, such as hybridization and horizontal gene transfer, which create networks of genetic relationships that cannot be fully captured by traditional tree models [17] [6]. For researchers studying biodiversity, conservation, and agricultural genetics, this shift necessitates advanced computational tools capable of inferring these complex relationships from the ever-growing volumes of genomic data.

The emerging discipline of phylogenetic network reconstruction faces significant computational bottlenecks. Traditional methods struggle with the exponential growth in computational and storage resources required as sequence datasets expand [36]. DNA language models (DLMs), a class of foundational models trained on vast corpora of genomic sequences, offer a transformative solution. By capturing complex, long-range dependencies in DNA sequences through self-attention mechanisms, DLMs provide a powerful, alignment-free approach to genomic analysis [36] [37]. When combined with k-mer tokenization strategies—which break down sequences into fixed-length subunits—these models enable efficient and accurate phylogenetic placement, even in the presence of reticulate evolutionary events [38] [39]. This guide compares the performance of current methodologies at this intersection, providing researchers with the experimental data and protocols needed to implement these cutting-edge approaches.

Comparative Analysis of k-mer Tokenization Strategies for Genomic Language Models

The performance of a DNA language model is profoundly influenced by how its input sequences are broken down into tokens, a process known as tokenization. The k-mer tokenization strategy is a critical design choice, balancing computational efficiency with the model's ability to capture biological context [38].

Performance of k-mer Tokenization Methods

Table 1: Comparison of k-mer tokenization strategies for Genomic Language Models. Performance is rated qualitatively for plant genomic tasks, with "High" indicating superior performance.

Tokenization Strategy	k-mer Size (k)	Vocabulary Size	Relative Computational Cost	Context Preservation	Best-Suited Tasks
Fully Overlapping [38]	3-8	4k + 5	High	High	Splice site prediction, regulatory element discovery
Non-Overlapping [38]	3-8	4k + 5	Low	Medium	Large-scale genome screening, initial sequence annotation
AgroNT Method [38]	6 (non-overlapping)	46 + 5 = 4096 + 5	Medium	Medium	General-purpose plant genomic tasks

As evidenced in plant genomic tasks such as splice site prediction, a thoughtful design of the k-mer tokenizer plays a critical role in model performance, often outweighing model scale [38]. Fully overlapping k-mers, which slide a window of size k one nucleotide at a time, generally enhance prediction accuracy by preserving fine-grained local sequence context. In contrast, non-overlapping strategies improve computational efficiency by reducing token redundancy, achieving competitive accuracy for some tasks [38]. The vocabulary size for any k-mer strategy is defined by the formula Vk = 4k + 5, where the five additional tokens are for special functions like padding and masking [38].

Quantitative Performance in Genomic Prediction Tasks

Table 2: Performance comparison of DNA Language Models on specific phylogenetic and genomic tasks. "N/A" indicates data was not available in the provided search results.

Model Name	Core Architecture & Training	Reported Accuracy / Performance	Key Application in Phylogenetics/Genomics
PhyloTune [36]	Fine-tuned DNABERT	Enables rapid subtree updates with modest trade-off in accuracy (RF distance increase ~0.01-0.03)	Taxonomic unit identification & attention-guided subtree construction
Species-Aware DNA LM [37]	Species-token informed Transformer	Captures regulatory elements over >500 million years; improves motif discovery & expression prediction	Alignment-free capture of regulatory element evolution
k-mer BERT (Overlapping) [38]	BERT with optimized k-mer tokenization	Performs on par with larger AgroNT model on plant genomic tasks	Efficient alternative for splice site and polyadenylation site prediction

The PhyloTune method demonstrates the practical application of DLMs for phylogenetic placement. By leveraging a pre-trained DLM (DNABERT) to identify the smallest taxonomic unit for a new sequence, it circumvents the need to reconstruct a full tree from scratch. This approach significantly reduces computational time while maintaining high topological accuracy, as measured by the normalized Robinson-Foulds (RF) distance [36]. Furthermore, species-aware DNA language models show a remarkable ability to capture functional regulatory elements across vast evolutionary distances—over 500 million years—far beyond the limits of traditional sequence alignment [37].

Experimental Protocols for Phylogenetic Placement and Network Inference

Protocol 1: The PhyloTune Method for Efficient Phylogenetic Updates

The PhyloTune protocol provides a workflow for integrating new sequences into an existing phylogenetic tree or network using a pre-trained DNA language model [36].

• Input and Model Preparation:

  • Input: An existing reference phylogenetic tree (or network) with curated taxonomic classifications, and one or more new genomic sequences for placement.
  • Model: A pre-trained DNA language model (e.g., DNABERT) is fine-tuned on the taxonomic hierarchy of the reference tree. This step trains a hierarchical linear probe (HLP) for each taxonomic rank (e.g., family, genus, species) within the tree.

• Smallest Taxonomic Unit Identification:

  • The new sequence is processed by the fine-tuned DLM.
  • The model performs novelty detection to determine the lowest taxonomic rank at which the sequence belongs to a known group.
  • Simultaneously, it performs taxonomic classification to assign the sequence to a specific taxon at that rank. This combined step identifies the "smallest taxonomic unit," which defines the specific subtree for update.

• High-Attention Region Extraction:

  • The genomic sequences within the identified subtree are fed through the DLM.
  • The attention weights from the model's final transformer layer are analyzed. These weights indicate nucleotides considered most important for the classification task.
  • Each sequence is divided into K regions, which are scored based on their average attention weight.
  • A voting method (e.g., minority-majority) selects the top M regions (M < K) with the highest scores as the "high-attention regions."

• Targeted Subtree Construction:

  • Only the high-attention regions extracted from the sequences in the target subtree and the new sequence are used for subsequent analysis.
  • These regions are aligned using standard tools like MAFFT.
  • A new subtree is inferred from the alignment using a tool like RAxML, and the original tree is updated with this new topology.

Protocol 2: Optimizing k-mer Tokenization for Domain-Specific GLMs

This protocol outlines the process of pre-training an efficient genomic language model with an optimized k-mer tokenizer for specific downstream tasks, such as identifying regions under reticulate evolution [38].

• Data Curation and Preparation:

  • Select a domain-specific genomic corpus (e.g., plant genomes for agronomic research).
  • Extract sequences of a fixed length (e.g., 510 bp) with a defined stride (e.g., 255 bp for 50% overlap) from reference genomes.

• Tokenizer Configuration and Training:

  • Define k-mer parameters: Choose a range of k-mer sizes (e.g., k=3 to 8) and select an overlap scheme (fully overlapping vs. non-overlapping).
  • Pre-train BERT models: Using the Hugging Face Transformers library, pre-train separate BERT models for each tokenizer configuration.
  • Apply masked language modeling: During training, 15% of random k-mers are masked, and the model is trained to predict the original tokens. This self-supervised task teaches the model contextual dependencies in the DNA sequence.

• Model Evaluation and Selection:

  • Fine-tuning: Take the pre-trained models and fine-tune them on specific downstream tasks relevant to phylogenetic networks, such as splice site prediction or alternative polyadenylation site prediction.
  • Performance analysis: Evaluate models based on prediction accuracy and computational efficiency. A model with a well-optimized k-mer strategy can achieve performance comparable to much larger, general-purpose models like AgroNT [38].

The Scientist's Toolkit: Essential Reagents & Computational Solutions

Table 3: Key research reagents and computational tools for implementing DLM-based phylogenetic placement.

Item Name	Type	Function / Application	Example Use Case
DNABERT [36] [37]	Pre-trained Model	Foundational DNA language model providing core sequence understanding.	Base model for fine-tuning in PhyloTune for taxonomic classification.
Hierarchical Linear Probe (HLP) [36]	Computational Method	Enables simultaneous novelty detection and classification across taxonomic ranks.	Identifying the precise taxonomic unit of a new sequence in a known phylogeny.
k-mer Tokenizer [38] [39]	Data Pre-processing	Breaks continuous DNA sequence into discrete, analyzable tokens for transformer models.	Optimizing model input for specific tasks (e.g., using overlapping 6-mers for promoter prediction).
Transformer Attention Weights [36] [37]	Model Interpretation	Highlights nucleotides/regions the model deems important for its predictions.	Extracting high-attention regions for targeted phylogenetic analysis in PhyloTune.
Reference Phylogenomic Dataset [36] [37]	Curated Data	Benchmarking and fine-tuning dataset for evaluating model performance.	Plant (Embryophyta) or microbial (Bordetella) datasets for testing phylogenetic placement.
RAxML-NG / MAFFT [36]	Downstream Tool	Performs multiple sequence alignment and maximum likelihood tree inference.	Constructing the final phylogenetic subtree from high-attention regions.

The integration of DNA language models and optimized k-mer tokenization strategies marks a significant advancement for phylogenetic research, particularly for probing the complexities of reticulate evolution. The experimental data and protocols detailed in this guide demonstrate that methodologies like PhyloTune and species-aware models offer a powerful, efficient, and scalable alternative to traditional pipelines. By enabling targeted analysis and capturing functional genomic information across deep evolutionary time, these tools empower researchers to move beyond simplistic trees towards a more realistic "web of life" understanding. This progress not only clarifies evolutionary history but also provides a stronger genetic foundation for addressing pressing challenges in biodiversity conservation and crop development.

Cytonuclear discordance, the incongruence between phylogenetic trees built from nuclear versus plastid (chloroplast) genomic data, presents a significant challenge in evolutionary biology. This phenomenon, primarily driven by incomplete lineage sorting (ILS) and hybridization, complicates the reconstruction of accurate species relationships. This guide objectively compares the performance of different genomic approaches and analytical protocols for testing reticulate evolution, using recent phylogenomic studies in plants as experimental case studies. We summarize quantitative findings, detail essential methodologies, and provide a toolkit for researchers aiming to distinguish between competing evolutionary processes.

In plant phylogenomics, the standard "tree of life" model is often insufficient to describe evolutionary histories characterized by complex interactions like hybridization and introgression. This leads to cytonuclear discordance, where the evolutionary history told by the organellar (plastid) genome conflicts with the history told by the nuclear genome [40] [41]. Resolving this discordance is critical for accurate phylogenetic inference and for testing hypotheses of reticulate evolution, which describes a web-like pattern of evolution involving the exchange of genetic material between lineages.

Two primary biological processes explain most observed discordance:

• Incomplete Lineage Sorting (ILS): The failure of ancestral genetic polymorphisms to coalesce (merge into a common ancestor) in the immediate ancestral population of two or more species. This is a stochastic process related to the effective population size and the short time intervals between successive speciation events [42] [43].
• Hybridization and Introgression: The transfer of genetic material from one species into the gene pool of another through repeated backcrossing. This can lead to chloroplast capture, where the plastid genome of one species is fixed in another, distinct nuclear background [40] [44].

Distinguishing between these processes is a central goal in modern phylogenomics and requires robust genomic datasets and sophisticated analytical frameworks, moving beyond simple trees to phylogenetic networks [17] [6].

Comparative Analysis of Phylogenomic Workflows and Their Performance

Different sequencing and analytical approaches offer varying resolutions for detecting and interpreting cytonuclear discordance. The table below compares the key methodologies and findings from three recent plant studies.

Table 1: Performance Comparison of Phylogenomic Approaches in Resolving Cytonuclear Discordance

Study System / Clade	Sequencing Method	Genomic Data Analyzed	Key Discordance Findings	Inferred Primary Driver(s)	Analytical Methods Used
Rubioideae (Rubiaceae) [40] [45]	Target capture (Angiosperms353) with off-target plastome assembly	353 nuclear genes + complete plastomes	Several instances of highly supported discordance between nuclear and plastid phylogenies	Majority by ILS; plastome introgression in some cases	Coalescent simulation (ILS testing), concatenation
Ficus section Galoglychia [44]	Sanger sequencing of selected loci	Chloroplast DNA markers + nuclear ITS/ETS	Significant discordance between chloroplast and nuclear phylogenetic trees	Introgressive hybridization	Phylogenetic tree comparison, statistical tests for discordance
Major Angiosperm Lineages (Mesangiospermae) [42]	Whole genome sequencing (177 genomes)	Nuclear genes + plastomes	Extensive gene-tree heterogeneity and cytonuclear discordance at deep nodes	Pervasive ancient hybridization and ILS	Phylogenetic networks, coalescent simulations

Performance Insights:

• Workflow Efficiency: The Rubioideae study [40] [45] demonstrates the efficiency of leveraging off-target reads from target capture sequencing (e.g., Angiosperms353 kit) to assemble complete plastomes, providing a cost-effective method for generating matched nuclear and plastid datasets from the same samples.
• Data Sufficiency: The Mesangiospermae study [42] highlights that whole-genome sequencing, while resource-intensive, is necessary to resolve deep, rapid radiations. It provides the extensive data required for sophisticated network analyses and simulations to disentangle complex signals of ILS and hybridization.
• Driver Discrimination: Across studies, coalescent simulation remains a key analytical method for assessing the relative roles of ILS and introgression [40] [42]. However, there is a growing consensus that ancient hybridization plays a more substantial role in deep-level discordances than previously recognized [42] [41].

Detailed Experimental Protocols for Key Studies

Protocol: Resolving Discordance in Rubioideae via Target Capture

This protocol is adapted from Thureborn et al. (2024) [40] [45].

1. Sample and Dataset Preparation:

• Taxon Sampling: Utilize an identical set of taxa for both nuclear and plastid analyses to ensure direct comparability. The Rubioideae study included 124 species (101 Rubioideae, 23 outgroups).
• DNA Sequencing: Perform target capture sequencing using the Angiosperms353 kit, which enriches 353 conserved nuclear genes across angiosperms.
• Plastome Assembly: Assemble complete or near-complete plastomes from the off-target reads generated during the target capture process. Map these reads to a reference plastome and perform de novo assembly for non-mapping reads.

2. Phylogenetic Inference:

• Nuclear Phylogeny: Assemble the 353 target nuclear genes. Infer the species tree using both concatenation (e.g., using maximum likelihood in RAxML or IQ-TREE) and coalescent-based methods (e.g., ASTRAL-III).
• Plastid Phylogeny: Assemble the plastome data into a single alignment. Reconstruct the plastome tree using maximum likelihood.
• Support Assessment: Evaluate node support using bootstrapping (e.g., 1000 bootstrap replicates).

3. Discordance Analysis:

• Tree Comparison: Visually and statistically compare the nuclear and plastid phylogenies to identify nodes with strong, conflicting support.
• Coalescent Simulation: Use tools like SOWHAT or custom simulations to test whether the observed level of discordance at a specific node can be explained by ILS alone. This involves simulating gene trees under the null hypothesis (no gene flow) and comparing the simulated distribution of tree distances to the observed distance.

4. Hypothesis Testing:

• If the observed discordance is statistically greater than expected under ILS alone, alternative hypotheses such as plastome introgression are supported.

The workflow for this protocol, from sampling to hypothesis testing, is summarized in the diagram below.

Protocol: Testing for Ancient Hybridization in Major Angiosperm Lineages

This protocol is adapted from Huang et al. (2025) [42].

1. Genomic Dataset Curation:

• Data Collection: Compile whole-genome sequencing data from a broad taxonomic sample (e.g., 177 angiosperm genomes representing eudicots, monocots, magnoliids, Ceratophyllales, and Chloranthales).
• Orthology Inference: Identify sets of single-copy orthologous genes across all taxa using tools such as OrthoFinder.

2. Multi-method Phylogenetic Reconstruction:

• Gene Tree Inference: Reconstruct individual gene trees for each orthologous group.
• Species Tree Inference: Infer a species tree from the set of gene trees using a coalescent-based method (e.g., ASTRAL).
• Network Inference: Reconstruct a phylogenetic network using a method that models hybridization (e.g., PhyloNet, NANUQ) to identify potential reticulate events.

3. Analysis of Gene Tree Heterogeneity:

• Quantify the degree of conflict among gene trees. High levels of conflict at specific nodes are indicative of either ILS or hybridization.
• Use statistical tests (e.g., ABBA-BABA or related D-statistics) on genome-wide SNP data to test for signatures of ancient gene flow between lineages.

4. Coalescent Simulation:

• Simulate expected distributions of gene tree discordance under a model of pure ILS (without hybridization).
• Compare the empirical gene tree distribution to the simulated one. A significantly worse fit suggests that processes like hybridization are required to explain the data.

The Scientist's Toolkit: Essential Research Reagents and Solutions

Successful resolution of cytonuclear discordance relies on a suite of bioinformatic tools and genomic resources. The table below details key solutions used in the featured studies.

Table 2: Research Reagent Solutions for Phylogenomic Discordance Studies

Tool/Resource	Category	Primary Function	Application Example
Angiosperms353 Kit [40] [45]	Wet-lab Reagent	Target capture bait set for enriching 353 conserved nuclear genes across angiosperms.	Generating comparable, multi-locus nuclear datasets across diverse plant taxa.
ASTRAL-III [35]	Software	Coalescent-based species tree estimation from a set of gene trees.	Inferring the primary species tree from hundreds of nuclear gene trees, accounting for ILS.
PhyloNet / NANUQ [17]	Software	Inference and analysis of phylogenetic networks.	Modeling and testing for hybridization events in a phylogeny.
D-statistics (ABBA-BABA) [35]	Analytical Method	Test for gene flow between taxa by quantifying allele sharing patterns.	Detecting signatures of introgression in genomic data.
OrthoFinder	Software	Inference of orthologous groups and gene families from sequenced genomes.	Identifying single-copy orthologs for phylogenomic analysis from whole-genome data [42].
GATK [43]	Software	Genome Analysis Toolkit for variant discovery from high-throughput sequencing data.	Calling single nucleotide polymorphisms (SNPs) from whole-genome resequencing data.

Signaling Pathways and Analytical Workflows for Reticulate Evolution

The logical process for moving from raw data to a supported conclusion of reticulate evolution involves a defined pathway of analytical steps. The diagram below outlines this high-level conceptual workflow, integrating the components from the toolkit.

The move from viewing evolution as a strictly bifurcating tree to a complex web is reshaping plant phylogenomics [6]. As evidenced by the case studies in Rubioideae, Ficus, and major angiosperm lineages, cytonuclear discordance is a common and informative feature of plant genomes. Distinguishing between ILS and hybridization is now feasible with robust genomic datasets—generated via target capture or whole-genome sequencing—and sophisticated analytical protocols centered on coalescent theory and phylogenetic networks. The emerging consensus is that both processes are pervasive, but ancient hybridization has been a particularly underappreciated force in shaping the deep-level evolutionary history of plants [42] [41]. The integrated toolkit of wet-lab reagents and bioinformatic software provides researchers with a powerful framework to test hypotheses of reticulate evolution, ultimately leading to a more accurate and nuanced understanding of the plant "web of life."

The reconstruction of evolutionary history, a cornerstone of biological science, is undergoing a fundamental paradigm shift. For decades, the "family tree" has served as the primary model for representing evolutionary relationships among species, genes, and even diseased cells. However, growing genomic evidence across biomedical fields—from virology to oncology—increasingly reveals that evolutionary histories are often not strictly tree-like. Processes such as viral recombination, hybridization, and horizontal gene transfer create evolutionary pathways that are more accurately represented by interconnected webs. This recognition has propelled the development and application of phylogenetic networks, which move beyond the limitations of traditional trees to model these complex, reticulate relationships [6].

The implications of this shift are profound for biomedical research. In virology, networks can trace the complex recombination events that give rise to new viral variants. In cancer biology, they can map the intricate evolutionary history of tumor subclones, which often exchange genetic material. This guide provides a comparative analysis of software tools enabling this transition, detailing their performance and applications through specific experimental protocols used in cutting-edge biomedical research.

Comparative Analysis of Phylogenetic Software Tools

The following table summarizes the key features and biomedical applications of selected phylogenetic visualization and analysis tools, highlighting their support for network-based analyses.

Table 1: Comparison of Phylogenetic Visualization and Analysis Software

Software Name	Type	Key Features	Strengths for Biomedical Research	Network Support
TreeViewer [46]	Desktop GUI & CLI	Highly modular pipeline, user-friendly, publication-quality figures, supports large trees.	Flexibility for complex datasets like viral evolution or tumor heterogeneity; command-line useful for automation.	Flexible, modular design.
IcyTree [47]	Online Tool	Client-side Javascript SVG viewer for annotated rooted trees.	Quick visualization and sharing of results, useful for collaborative analysis of pathogen genomes.	Yes (phylogenetic networks).
iTOL [47]	Online Tool	Extensive annotation options, scriptable batch interface.	Excellent for annotating viral variants or cancer cell lineages with metadata (e.g., mutations, drug resistance).	Primarily for trees.
Taxonium [47]	Online Tool	Exploration of very large trees (millions of nodes), mutation annotation.	Ideal for large-scale surveillance of SARS-CoV-2 or influenza evolution.	Primarily for trees.
Dendroscope [47]	Desktop GUI	Interactive viewer for large phylogenetic trees and networks.	Specialized for complex network analysis, suitable for studying horizontal gene transfer in bacteria or cancer cells.	Yes (phylogenetic networks).
ggtree [47]	R Library	Tree visualization and annotation using "grammar of graphics".	Seamlessly integrates with statistical analysis pipelines in R/Bioconductor for genomics.	Via extensions.
PhyloViZ [47]	Desktop & Online	Analysis and visualization of minimum spanning trees from allelic/SNP profiles.	Applied in microbial genomics and outbreak tracing (e.g., bacterial pathogen transmission).	Yes (minimum spanning networks).

Experimental Protocols for Reticulate Evolution Research

Protocol 1: Tracking Viral Evolution via Serial Passaging

Objective: To observe the natural evolutionary pathways and convergent evolution of a virus, such as SARS-CoV-2, in a controlled laboratory environment, independent of host immune pressures [48].

• Key Research Reagents:

  • Virus Isolates: Purified samples of viral variants (e.g., Alpha, Delta, Omicron).
  • Cell Culture: Vero E6 cells (monkey kidney cells lacking a strong immune response).
  • Growth Medium: Appropriate tissue culture media (e.g., DMEM with fetal bovine serum).
  • RNA Extraction Kit: For isolating viral genetic material post-passaging.
  • Next-Generation Sequencing (NGS) Platform: For whole-genome sequencing of viral populations at each passage.

• Methodology:

  • Inoculation: Introduce a defined viral sample into a flask of confluent Vero E6 cells.
  • Incubation: Allow the virus to replicate for a set period (e.g., 48-72 hours).
  • Harvesting: Collect the culture supernatant, which contains the progeny virus particles.
  • Serial Passaging: Use a small aliquot of the harvested supernatant to inoculate a fresh flask of naive Vero E6 cells. This process is repeated numerous times (e.g., 33 to 100 passages).
  • Sequencing and Analysis: At regular passage intervals, extract viral RNA, perform whole-genome sequencing, and map mutations. The data is analyzed to identify:
    • Convergent Evolution: Mutations that arise repeatedly and independently in different lineages.
    • Mutation Rates: The pace of genetic change in different parts of the viral genome.

• Data Interpretation: Mutations found in both lab-based passaging and real-world variants suggest a strong intrinsic bias in viral evolution. This data can be used to construct phylogenetic networks that model the potential emergence of future variants, informing the design of next-generation vaccines and antivirals [48].

Protocol 2: Mapping Ancient Viral Elements in Cancer Cells

Objective: To determine the 3D structure and function of human endogenous retrovirus (HERV) proteins reawakened in cancer and autoimmune cells, enabling the development of targeted diagnostics and therapies [49].

• Key Research Reagents:

  • Patient Samples: Tissue biopsies or blood samples from cancer patients (e.g., breast, ovarian) and healthy controls.
  • Stable HERV-K Env Protein: Recombinantly expressed and purified HERV-K envelope glycoprotein, stabilized in its pre-fusion state via point mutations.
  • Monoclonal Antibodies: Antibodies generated against the HERV-K Env protein.
  • Cryo-Electron Microscopy (Cryo-EM): For high-resolution structural determination of the protein and its complexes with antibodies.

• Methodology:

  • Protein Stabilization and Expression: Engineer a version of the HERV-K Env protein with specific amino acid substitutions to "lock" it in its pre-fusion conformation. Express and purify the protein.
  • Complex Formation: Incubate the stabilized HERV-K Env with Fab fragments of targeting antibodies to form a stable complex.
  • Cryo-EM Grid Preparation: Flash-freeze the sample in vitreous ice on a microscopy grid.
  • Data Collection and Processing: Use a cryo-electron microscope to collect thousands of 2D micrograph images. Process these images through computational reconstruction to generate a 3D atomic model of the protein-antibody complex.
  • Validation on Patient Cells: Use the characterized antibodies in flow cytometry or immunohistochemistry to detect HERV-K Env expression on neutrophils and cancer cells from patient samples.

• Data Interpretation: The resolved structure reveals unique epitopes and conformations not found in other human proteins. This allows for the rational design of highly specific immunotherapies (e.g., CAR-T cells, antibody-drug conjugates) that can target cancer cells expressing HERV-K while sparing healthy tissues [49].

Visualizing Reticulate Evolution in Biomedical Research

The following diagram illustrates the conceptual and analytical workflow for investigating reticulate evolution in viruses and cancer, integrating the experimental protocols outlined above.

Essential Research Reagent Solutions

The following table catalogs key reagents and their functions that are critical for the experimental workflows discussed in this guide.

Table 2: Key Research Reagents for Reticulate Evolution Studies

Reagent / Material	Function in Research	Application Example
Vero E6 Cell Line	Provides a permissive cell culture system for viral replication without the complex selective pressure of an adaptive immune system.	Studying intrinsic evolutionary pathways of SARS-CoV-2 through serial passaging [48].
Stabilized Viral Glycoproteins	Engineered versions of envelope proteins (e.g., HERV-K Env, SARS-CoV-2 Spike) locked in a specific conformation for structural and immunological studies.	Enabling high-resolution Cryo-EM structure determination to guide therapeutic antibody design [50] [49].
Monoclonal Antibody Panels	Collections of antibodies that bind to different regions (epitopes) of a target protein, used for characterization, diagnostics, and therapy.	Detecting HERV-K Env on neutrophils from rheumatoid arthritis patients or targeting cancer cells [49].
Cryo-Electron Microscopy	A high-resolution structural biology technique that images biomolecules in a near-native, frozen-hydrated state.	Solving the first 3D structure of the HERV-K Env protein, revealing a unique architecture [49].
Phylogenetic Network Software	Computational tools to visualize and analyze evolutionary relationships that include horizontal events like recombination and hybridization.	Modeling the complex evolutionary history of viruses or tumor cells that do not follow a simple tree-like pattern [6] [17].

Navigating Challenges: Troubleshooting Inference, Model Selection, and Interpretation

Addressing Computational Bottlenecks and Strategies for Scalable Analysis

In the field of phylogenomics, the inference of phylogenetic networks is fundamental for understanding reticulate evolutionary processes such as hybridization, introgression, and horizontal gene transfer. However, as datasets have grown in size and complexity, the computational methods used for inference have faced significant scalability challenges. This guide objectively compares the performance of leading phylogenetic network inference methods, providing a detailed analysis of their scalability and efficiency based on empirical data and simulation studies. The focus is on identifying computational bottlenecks and presenting strategies for scalable analysis, which is critical for researchers, scientists, and drug development professionals working with large-scale genomic data.

Performance Comparison of Phylogenetic Network Inference Methods

A comprehensive scalability study evaluated state-of-the-art phylogenetic network inference methods on both simulated and empirical datasets, focusing on two key dimensions: the number of taxa and evolutionary divergence [51]. The performance of these methods was assessed in terms of topological accuracy, runtime, and memory usage.

Table 1: Performance Comparison of Phylogenetic Network Inference Methods with Increasing Taxon Number

Method	Optimization Criterion	Accuracy Trend (with increasing taxa)	Computational Limit	Runtime/Memory Requirements
Neighbor-Net	Concatenation	Degrades	>30 taxa	Low
SplitsNet	Concatenation	Degrades	>30 taxa	Low
MP (Maximum Parsimony)	Parsimony (MDC)	Degrades	~25 taxa	Moderate
MLE (Maximum Likelihood Estimation)	Coalescent-based likelihood	High but degrades	≤25 taxa	Very High
MLE-length	Coalescent-based likelihood with branch lengths	High but degrades	≤25 taxa	Very High
MPL (Maximum Pseudo-likelihood)	Pseudo-likelihood approximation	High but degrades	~25 taxa	High
SNaQ	Pseudo-likelihood with quartets	High but degrades	~25 taxa	High
ALTS	Tree-child network alignment	Maintains on larger datasets	Up to 50 taxa with 50 trees	Moderate (~15 min for 50x50) [52]

Table 2: Impact of Sequence Divergence on Method Performance

Method Category	Effect of Increased Mutation Rate	Key Strengths	Key Limitations
Concatenation (Neighbor-Net, SplitsNet)	Reduced accuracy	Computational efficiency, fast runtime	Inaccurate under high ILS or gene flow [51]
Parsimony (MP)	Reduced accuracy	Faster than probabilistic methods	Less accurate than model-based methods [51]
Probabilistic (MLE, MLE-length)	Significant accuracy reduction	Highest accuracy on smaller datasets	Prohibitive runtime (>weeks for ≥30 taxa) [51]
Pseudo-likelihood (MPL, SNaQ)	Significant accuracy reduction	Good balance of accuracy and speed	Heuristic nature, scalability limits [51]
Tree-Child Alignment (ALTS)	Not explicitly tested	Scalability to larger datasets	Limited to tree-child networks [52]

Experimental Protocols for Method Evaluation

Scalability Assessment Protocol

The experimental design for evaluating phylogenetic network inference methods involves a structured approach to assess performance across different dimensions of scale [51].

Experimental Workflow for Phylogenetic Network Method Evaluation

Dataset Generation Methodology

• Empirical Data Sampling: Utilize natural population data (e.g., mouse populations) representing real-world evolutionary scenarios with documented gene flow [51].

• Simulation Protocol:

  • Employ model phylogenies with a known number of reticulations (typically starting with single reticulation networks)
  • Simulate multi-locus sequence data under coalescent models incorporating gene flow
  • Vary parameters including number of taxa (from 10 to 50+) and sequence mutation rate
  • Replicate datasets to ensure statistical significance of results

• Performance Metrics Collection:

  • Topological Accuracy: Compare inferred networks to true simulated networks using distance metrics
  • Computational Requirements: Measure runtime and memory usage under controlled conditions
  • Scalability Limits: Identify points where methods fail to complete analyses within reasonable timeframes

The Scientist's Toolkit: Essential Research Reagents and Computational Solutions

Table 3: Key Research Reagent Solutions for Phylogenetic Network Inference

Tool/Resource	Type	Primary Function	Application Context
PhyloNet	Software Package	Implement MLE, MPL methods for network inference	Probabilistic inference of phylogenetic networks [51]
ALTS	Software Program	Infer tree-child networks via lineage taxon string alignment	Scalable network inference from multiple gene trees [52]
SNaQ	Software Tool	Species network inference using quartet-based pseudo-likelihood	Coalescent-based network inference with heuristic search [51]
Multi-locus Sequence Data	Biological Data	Provide input for gene tree estimation	Essential raw material for all reconciliation-based methods [51]
Gene Trees	Phylogenetic Data	Serve as input for species network inference	Required for summary approaches (MP, MLE, MPL, SNaQ) [53]
Model Phylogenies with Reticulations	Simulation Framework	Provide ground truth for method validation	Critical for benchmarking studies and accuracy assessment [51]

Analysis of Computational Bottlenecks and Scalability Limits

The scalability study revealed that the most accurate methods (probabilistic inference using coalescent-based models) become computationally prohibitive as dataset size increases beyond approximately 25 taxa [51]. The key computational bottlenecks include:

• Model Likelihood Calculations: Exact likelihood computations under coalescent models with gene flow represent the primary performance bottleneck in MLE and MLE-length methods [51].

• Heuristic Search Strategies: All multi-locus methods require heuristics to navigate the vast space of possible phylogenetic networks, as the inference problem is NP-hard [51].

• Memory Requirements: Storage and processing of large gene tree sets and intermediate computational structures demand significant memory resources.

Computational Bottlenecks in Network Inference

Strategies for Scalable Phylogenetic Network Analysis

Methodological Innovations

• Pseudo-likelihood Approximations: Methods like MPL and SNaQ substitute computationally intensive exact likelihood calculations with more efficient approximations, enabling analysis of larger datasets while maintaining reasonable accuracy [51].

• Tree-Child Network Alignment: The ALTS method introduces a novel approach by reducing the network inference problem to aligning lineage taxon strings, enabling inference for datasets with up to 50 taxa and 50 trees in approximately 15 minutes [52].

• Constraint-Based Approaches: Some methods focus on specific network classes (e.g., tree-child networks) that are more computationally tractable while maintaining biological relevance [52].

Practical Implementation Recommendations

• Dataset Size Considerations: For studies involving more than 25-30 taxa, consider using pseudo-likelihood methods (MPL, SNaQ) or the ALTS method rather than full likelihood approaches.

• Algorithm Selection Framework:

  • Small datasets (<25 taxa): Probabilistic methods (MLE, MLE-length) provide highest accuracy
  • Medium datasets (25-50 taxa): Pseudo-likelihood methods (MPL, SNaQ) offer best accuracy-efficiency balance
  • Large datasets (>50 taxa): ALTS or other specialized methods currently provide the only feasible option

• Computational Resource Planning: Budget significant computational resources (weeks of CPU time, large memory allocation) for probabilistic analyses, even for moderate-sized datasets.

The field of phylogenetic network inference requires continued algorithmic development to address the scalability challenges posed by modern phylogenomic datasets. The current state-of-the-art methods face significant limitations when applied to datasets with more than 25 taxa, creating a methodological gap that needs to be addressed to keep pace with the scale of contemporary evolutionary studies [51].

The reconstruction of evolutionary history is a cornerstone of modern biological research, informing fields from ecology to drug discovery. Traditionally, this history has been represented through phylogenetic trees, which model divergence from a common ancestor through a branching process. However, an increasing body of genomic evidence reveals that the evolutionary history of many organisms, particularly plants and microbes, is better represented by networks rather than trees due to widespread reticulate events such as hybrid speciation, horizontal gene transfer, and endosymbiosis [23] [24]. This creates a fundamental problem of model misspecification when traditional tree-based methods are applied to reticulate evolutionary histories.

Model misspecification occurs when the analytical model used in phylogenetic reconstruction (a tree) does not match the true underlying evolutionary process (a network). This discrepancy can lead to systematically incorrect inferences about evolutionary relationships, with potentially significant consequences for downstream applications including drug target identification, understanding of pathogen evolution, and tracing the origins of adaptive traits [23]. The "Network vs. Tree" paradigm represents one of the most significant challenges in contemporary phylogenetics, necessitating a clear understanding of how tree methods perform when applied to network-like data and what alternatives exist.

This guide provides a comparative analysis of phylogenetic methods when applied to reticulate scenarios, summarizes experimental data on their performance, details essential methodologies, and visualizes key concepts to equip researchers with the tools needed to navigate complex evolutionary reconstructions.

Tree Methods in a Reticulate World: A Performance Comparison

When tree-based phylogenetic methods encounter data generated through reticulate evolution, their performance degrades in characteristic ways. The following table summarizes the documented behaviors and limitations of major tree-based method categories when faced with various forms of model misspecification.

Table 1: Performance of Tree-Based Methods Under Model Misspecification on Reticulate Networks

Method Type	Core Principle	Impact of Reticulation	Key Performance Limitations
Parsimony	Minimizes total evolutionary change	Creates conflicting signals; forces arbitrary choice between histories [23].	High susceptibility to long-branch attraction; produces positively misleading topologies under moderate reticulation.
Maximum Likelihood (ML)	Finds tree with highest probability given sequence data and model	Model assumes tree-like evolution; likelihood scores become unreliable [23].	Inconsistent parameter estimates (e.g., branch lengths); support values (bootstraps) become inflated and misleading.
Bayesian Inference	Estimates posterior distribution of trees using Markov Chain Monte Carlo	Prior (tree) contradicts true network process; MCMC mixes poorly [23].	Posteriors concentrated on incorrect trees; model comparison metrics (e.g., Bayes Factors) favor overly complex trees.

The primary issue common to all tree-based methods is that they are forced to represent a complex history, which may involve the merging of lineages, within a strictly branching framework. This fundamental mismatch means that even the most sophisticated tree-based methods will systematically misinterpret certain reticulate patterns. For example, a hybrid speciation event between two divergent lineages will be interpreted by a tree method as a branch point that is closely related to one parent, while the genetic material inherited from the other parent will be treated as either homoplasy or deep ancestral polymorphism [23]. This can lead to incorrect conclusions about monophyly, divergence times, and the direction of trait evolution.

Experimental Insights: Quantifying the Impact of Misspecification

Empirical and simulation studies have been crucial in quantifying the real-world impact of model misspecification. The following table synthesizes findings from key experimental approaches that test the performance of phylogenetic methods on data with known reticulate histories.

Table 2: Experimental Data on Method Performance for Detecting Reticulation

Experiment Focus	Key Methodology	Quantitative Findings	Implication for Researchers
Incongruence Detection	Compare gene trees from multiple unlinked loci [23].	>70% topological incongruence among 10+ loci strongly indicates hybridization.	Requires data from numerous independent nuclear markers; a few loci are insufficient to distinguish from incomplete lineage sorting.
Network vs. Tree Model Fit	Use statistical tests (e.g., likelihood-based) to compare tree and network models for the same data.	Network models often show significantly better fit (e.g., ΔAIC > 10) on plant datasets, confirming tree inadequacy [24].	Model testing should be a standard step; a single "best" tree may be statistically inferior to a network.
Power of Network Algorithms	Simulate genomes with known hybridization events; apply methods like SplitsTree or PhyloNet.	Methods accurately reconstruct parentage in allopolyploids (>95% success); accuracy drops for diploid hybrids, especially with gene flow (~70%) [23].	Allopolyploidy is easier to detect. Diploid hybrid detection requires dense genomic sampling and methods that account for post-hybridization gene flow.

A critical insight from these experiments is that the mere presence of incongruent gene trees is a primary line of evidence for reticulate evolution [23]. When analyses of different genomic regions consistently yield strongly supported but conflicting phylogenetic trees, it suggests that different parts of the genome have different evolutionary histories—a classic signature of reticulation. Furthermore, studies show that the number of markers is crucial. With only a few loci, it is difficult to distinguish hybridization from other processes like incomplete lineage sorting. Reliable detection of reticulation often requires data from dozens of independently inherited nuclear markers [23].

Detailed Experimental Protocol: Testing for Reticulation Using Incongruence

The following is a generalized protocol for a key experiment cited in the field: using incongruence among gene trees to test for reticulate evolution.

• Locus Selection and Sequencing: Select 20-50+ single-copy nuclear loci that are unlinked (on different chromosomes or sufficiently physically distant to assort independently during meiosis). Use PCR and sequencing or target capture with high-throughput sequencing to obtain sequence data for these loci across the operational taxonomic units (OTUs), including putative parents and hybrids, as well as outgroups.
• Individual Gene Tree Reconstruction: For each locus, align sequences and perform phylogenetic analysis using both Maximum Likelihood (e.g., RAxML, IQ-TREE) and Bayesian (e.g., MrBayes, BEAST2) methods. Use appropriate nucleotide substitution models selected via model testing software (e.g., ModelTest-NG). Perform bootstrapping (1000 replicates) for ML analyses and ensure high effective sample sizes (ESS > 200) for Bayesian analyses.
• Incongruence Assessment: Systematically compare the topologies of the resulting gene trees. Focus on strongly supported nodes (e.g., bootstrap ≥ 70%, posterior probability ≥ 0.95). Document all conflicting relationships with strong support. Use tools like DendroPy or Phylo in R to quantify topological distances.
• Network Reconstruction: Input all gene alignments or gene trees into a network inference tool designed to handle reticulation (e.g., SplitsTree for a distance-based network, PhyloNet for a multi-species network coalescent model). These methods do not force the data onto a single tree and can visualize conflicting signals as parallel edges or reticulation nodes.
• Statistical Testing: Formally test whether a network model provides a significantly better fit to the data than a tree model. This can be done using likelihood-based tests where possible, or by using the distribution of gene tree discordance compared to the expectations under a coalescent-only model (e.g., using the Puzzle or Bucky pipelines).

This protocol highlights that robust detection of reticulation is not a single analysis but a workflow of congruence testing, network visualization, and model comparison.

The Scientist's Toolkit: Essential Reagents and Computational Solutions

Research into reticulate evolution relies on a combination of biological materials and sophisticated computational tools.

Table 3: Key Research Reagent Solutions for Reticulate Evolution Studies

Item / Solution	Function / Application	Example Use-Case
Single-Copy Nuclear Exon Capture Kit	Targets hundreds of phylogenetically informative, low-copy nuclear regions from complex genomes.	Generating the dozens of independent gene trees needed to robustly infer a hybridization event.
PhyloNetworks (Software)	Implements statistical models for inferring phylogenetic networks from gene tree estimates or sequence alignments under the multispecies network coalescent.	Quantifying the strength of hybridization (heritage proportion) and identifying which lineages are hybrids and which are their parents.
SplitsTree	Creates phylogenetic networks from distance or character data using split decomposition, visualizing conflict and ambiguity.	Initial exploratory data analysis to see if a dataset has substantial network-like signal.
Hybrid Enrichment (SeqKit)	A laboratory and bioinformatic workflow for targeting and sequencing thousands of ultra-conserved element (UCE) loci across diverse taxa.	Obtaining a massive set of homologous loci for non-model organisms to power network analyses.

Visualizing Concepts and Workflows in Reticulate Evolution

To comprehend and communicate the core concepts and analytical processes of reticulate phylogenetics, clear visualizations are essential. The following diagrams, generated with Graphviz, illustrate the fundamental structural difference between evolutionary models and a standard research workflow.

Tree vs. Network Evolutionary Models

This diagram illustrates the fundamental structural difference between evolutionary models. The Tree Model on the left shows strictly divergent evolution, where lineages split and evolve independently. The Network Model on the right depicts a reticulate event, where a hybrid species forms from two parent species, combining their lineages—a pattern that cannot be accurately represented in a tree structure.

Reticulation Detection Workflow

This workflow chart outlines the logical process for testing for reticulate evolution. The process begins with broad genomic sampling and proceeds through gene tree reconstruction. A pivotal decision point is the assessment of significant incongruence among gene trees, which determines whether a network inference path is warranted over a standard species tree reconstruction.

The evidence is clear: the uncritical application of tree-based methods to groups with reticulate evolutionary histories carries a high risk of model misspecification, leading to incorrect phylogenetic inference and flawed biological conclusions. The limitations of parsimony, maximum likelihood, and Bayesian tree-inference methods in the face of hybridization and other reticulate processes are well-documented through both simulation and empirical studies [23] [24].

The field is moving beyond simply identifying incongruence toward developing sophisticated statistical models that explicitly test network hypotheses against tree alternatives. For researchers in phylogenetics, systematics, and comparative genomics, the key takeaway is the necessity of model awareness. Before embarking on a phylogenetic analysis, particularly in groups known for hybridization, researchers should: 1) Assess the biological likelihood of reticulation in their study system. 2) Employ a multi-locus data strategy from the outset. 3) Systematically test for topological incongruence. 4) Be prepared to use and interpret network-based phylogenetic methods.

The future of accurately reconstructing life's history lies not in choosing between trees or networks, but in using robust statistical frameworks to decide which model—tree, network, or a combination thereof—best explains the complex patterns in our genomic data.

Distinguishing Reticulation from Incomplete Lineage Sorting (ILS)

In evolutionary biology, the reconstruction of species' histories is fundamentally complicated by processes that create incongruence between gene trees and species trees. Two predominant sources of such discordance are incomplete lineage sorting (ILS) and reticulation. ILS occurs when ancestral genetic polymorphisms persist through multiple speciation events, leading to gene genealogies that do not match the species tree [54]. In contrast, reticulate evolution (including hybridization, introgression, and horizontal gene transfer) involves the exchange of genetic material between separately evolving lineages, creating a network-like evolutionary history [55] [19]. Distinguishing between these processes is crucial for accurate phylogenetic inference and understanding evolutionary mechanisms.

The challenge arises because both phenomena can produce similar patterns of gene tree discordance, yet they represent fundamentally different evolutionary processes. ILS represents the failure of lineages to coalesce in a manner consistent with species divergence, while reticulation involves the merging of evolutionary lineages through genetic exchange. Researchers have developed various analytical frameworks and tools to discriminate between these processes, leveraging patterns of discordance across the genome [56] [55].

Theoretical Foundations and Definitions

Incomplete Lineage Sorting

ILS is a population genetic phenomenon that occurs when the time between successive speciation events is too short for ancestral polymorphisms to completely sort into descendant lineages [54]. This results in gene trees that may exhibit topologies inconsistent with the species tree. The probability of ILS increases with larger effective population sizes and shorter intervals between speciation events [57].

Key characteristics of ILS include:

• Random distribution of discordant signals across the genome
• Equal probability of alternative topologies in certain phylogenetic configurations
• Time-dependent decay of discordance with increasing time between speciation events

In diploid organisms with large ancestral population sizes, such as pines, ILS is particularly prevalent. As noted in studies of Pinus, "lineage sorting between pine species is often incomplete" due to factors including "predominantly outcrossed mating, high within-species mean heterozygosity, long generation time, and large effective population sizes" [57].

Reticulate Evolution

Reticulate evolution encompasses evolutionary processes where genetic material from separate lineages combines, creating a network-like evolutionary history rather than a strictly diverging tree pattern [55]. This category includes:

• Hybridization: Interbreeding between individuals from distinct populations or species
• Introgression: Transfer of genetic material between species through repeated backcrossing
• Horizontal Gene Transfer: Lateral movement of genetic material between unrelated organisms

Unlike ILS, reticulation creates directional discordance where introgressed regions show preferential affinity between specific lineages. The phylogenetic network model accommodates reticulation through reticulation nodes with two parents, each with an associated inheritance probability (γ) that represents the proportion of genetic material derived from each parent [19].

Methodological Framework for Discrimination

Analytical Workflows

Discriminating between ILS and reticulation requires a systematic analytical workflow that integrates multiple lines of evidence. The following diagram illustrates a generalized approach:

Quantitative Indices for Discrimination

Novel tools have been developed to quantify signals of ILS and reticulation. The Phytop tool introduces two key indices that help distinguish these processes based on gene tree topology patterns [56]:

• ILS Index: Ranges from 0% to 100%, reflecting the strength of incomplete lineage sorting
• IH (Introgression/Hybridization) Index: Ranges from 0% to 50%, quantifying the strength of reticulation signals

Under an ILS-only scenario, the proportions of the three possible gene tree topologies for a four-taxon configuration (q1, q2, q3) are expected to be equal (q2 = q3). When introgression occurs without ILS, these proportions become imbalanced (q2 >> q3 for introgression from S to L) [56]. The mathematical relationship between these indices and gene tree proportions is illustrated below:

Performance Comparison of Inference Methods

Table 1: Comparison of Phylogenetic Network Inference Methods

Method	Underlying Approach	Data Input	Strengths	Limitations	Scalability
Phytop	Visualization and quantification of ILS/IH indices	Gene trees from ASTRAL	Fast-running, intuitive visualization, quantifiable measures	Limited to predefined species tree	High (completes in minutes even for large trees) [56]
Maximum Likelihood (BEAST)	Full likelihood calculation	Sequence alignments or gene trees	High accuracy, accounts for both ILS and reticulation	Computationally intensive	Low (struggles beyond 25 taxa) [55] [51]
SNaQ	Pseudo-likelihood approximation	Gene trees or quartets	Better scalability than full likelihood	Less accurate than full likelihood	Medium (handles moderate dataset sizes) [51]
MP (Maximum Parsimony)	Parsimony (minimize deep coalescences)	Gene trees	Fast computation	Less accurate, doesn't account for branch lengths	Medium [51]
D-statistics (ABBA-BABA)	Site pattern counting	Sequence data	Simple test for introgression	Limited to four-taxon case, doesn't estimate network parameters	High [56]

Method Performance on Simulated Data

Table 2: Accuracy of Methods on Simulated Datasets with Known Reticulation Events

Method	Topological Accuracy	Reticulation Detection Power	Inheritance Probability Estimation	Runtime (25 taxa)
Full Likelihood (BEAST)	High (85-95%)	High for recent hybridization	Accurate estimates	Prohibitive (weeks) [51]
SNaQ	Medium-High (75-90%)	Medium for recent hybridization	Reasonable approximations	Moderate (hours-days) [51]
MP	Low-Medium (60-75%)	Low, often misses events	Not estimated	Fast (hours) [51]
Phytop	Varies with ILS/IH levels	High when ILS is low, decreases with high ILS	Indirect through IH index	Very fast (minutes) [56]

Simulation studies reveal that method performance is significantly affected by evolutionary parameters. The diameter of reticulation events (evolutionary distance between donor and recipient lineages) strongly influences detectability, with larger diameters increasing detection power [19]. Furthermore, inheritance probabilities and the number of genomic regions involved in reticulation events impact methodological performance, with higher numbers of independent loci improving accuracy [19].

Performance comparison studies indicate that probabilistic methods generally outperform parsimony-based approaches but come with substantial computational costs. As noted in scalability assessments, "the most accurate methods were probabilistic inference methods which maximize either likelihood under coalescent-based models or pseudo-likelihood approximations to the model likelihood" [51]. However, this improved accuracy comes at a cost: "None of the probabilistic methods completed analyses of datasets with 30 taxa or more after many weeks of CPU runtime" [51].

Experimental Protocols and Workflows

Transcriptome-Based Phylogenomic Analysis

Research on Liliaceae tribe Tulipeae provides a comprehensive protocol for discriminating ILS and reticulation using transcriptomic data [58]:

• Data Collection: Sequence transcriptomes of target taxa, including multiple accessions per species and appropriate outgroups
• Dataset Construction:
  • Extract 74 plastid protein-coding genes (PCGs) for plastid genome analysis
  • Identify 2,594 nuclear orthologous genes (OGs) for nuclear genome analysis
• Gene Tree Inference: Reconstruct individual gene trees using maximum likelihood methods
• Species Tree Estimation: Infer species trees using both concatenation (ML) and multi-species coalescent (ASTRAL) approaches
• Discordance Analysis: Calculate "site concordance factors" (sCF) and "site discordance factors" (sDF1/sDF2) to identify nodes with high or imbalanced discordance
• Reticulation Testing: Apply D-statistics and QuIBL to test for introgression at conflicting nodes
• Polytomy Tests: Evaluate whether ILS alone can explain observed discordance

This approach successfully identified pervasive ILS and reticulate evolution among Amana, Erythronium, and Tulipa genera, explaining previously conflicting phylogenetic signals [58].

Maximum Likelihood Inference Protocol

For sequence-based inference of phylogenetic networks accounting for both ILS and reticulation [55]:

• Model Specification: Define a phylogenetic network Ψ as a rooted directed acyclic graph with:

  • Tree nodes (in-degree 1, out-degree greater than 1)
  • Reticulation nodes (in-degree 2, out-degree 1)
  • Branch lengths in coalescent units (λb = tb/Nb)

• Likelihood Calculation: Compute the likelihood of the network given sequence data S = {S1,...,Sm} for m independent loci: L(Ψ,Γ|S) = ∏i=1m ∫P(Si|g)p(g|Ψ,Γ)dg where P(Si|g) is the probability of sequence data given genealogy g, and p(g|Ψ,Γ) is the density of genealogies given network parameters

• Parameter Estimation: Simultaneously estimate:

  • Network topology (Ψ)
  • Branch lengths in coalescent units
  • Inheritance probabilities (Γ) for reticulation edges

• Model Selection: Use Bayesian Information Criterion (BIC) to select optimal network complexity, preventing overestimation of reticulation events [19]

This approach has been successfully applied to house mouse (Mus musculus) genomes, identifying a well-supported evolutionary history with two hybridization events [55].

Research Reagent Solutions

Table 3: Essential Research Tools for Discriminating ILS and Reticulation

Tool/Resource	Function	Application Context	Key Features
Phytop	Visualization and quantification of ILS/IH signals	Analysis of ASTRAL species trees	Fast computation, intuitive visualization, ILS and IH indices [56]
PhyloNet	Phylogenetic network inference	Reticulate evolution analysis	Implements MLE, MLE-length, and MPL methods [51]
ASTRAL	Species tree estimation from gene trees	Coalescent-based phylogenetics	Accounts for ILS, handles multi-locus data [56]
BEAST	Bayesian evolutionary analysis	Divergence time estimation, network inference	Full probabilistic modeling, flexible priors [55]
HyDe	Hypothesis testing of hybridization	Detection of hybrid taxa	Implements D-statistics and related tests [56]
Dsuite	D-statistics calculation	Introgression testing	Fast implementation of ABBA-BABA tests [56]

Case Studies in Empirical Systems

Pines (Genus Pinus)

Research on ponderosa pines (subsection Ponderosae) demonstrates how both ILS and reticulation shape evolutionary histories [57]. Sequence data from 53 accessions of 17 species revealed:

• Ancient introgression between Pinus coulteri lineages followed by genetic bottlenecks
• Recent hybridization transferring chloroplasts from P. jeffreyi to sympatric P. washoensis individuals
• Widespread ILS explaining other instances of non-monophyly

This study highlighted the necessity of multi-accession, multi-locus sampling, noting that "any analysis based on single-accession or single-locus sampling in Pinus" would be problematic due to the complex interplay of these processes [57].

Whiptail Lizards (Ameivula)

Genomic analysis of Brazilian whiptail lizards revealed mitonuclear discordance driven by ancient reticulation [59]:

• Mitochondrial capture through introgressive hybridization explained discordance between UCE (nuclear) and mitogenome trees
• Neutral evolution of mitogenomes rather than adaptive introgression
• Species delimitation approaches validated distinct species despite mitochondrial introgression

This case illustrates how comprehensive genomic sampling combined with network approaches can unravel complex speciation histories involving reticulation.

Marsupials

Whole-genome analysis of marsupials demonstrated the phenotypic consequences of ILS [60]:

• Over 31% of the genome of the South American monito del monte showed closer affinity to Diprotodontia than to other Australian marsupials due to ILS
• Pervasive conflicting signals across the genome consistent with some morphological variation
• Functional experiments validated phenotypic effects of stochastically fixed genes during ILS

This research provided rare empirical evidence that ILS can directly contribute to hemiplasy - incongruence between gene trees and phenotypic evolution.

Distinguishing between ILS and reticulation remains a challenging but essential task in evolutionary biology. No single method provides a universal solution, and researchers must select approaches based on their specific biological questions and dataset characteristics. Methodological development continues to advance, with current research focusing on improving scalability and accuracy while integrating additional sources of evidence such as phenotypic traits and genomic features.

The field is moving toward approaches that can simultaneously account for both ILS and reticulation without a priori assumptions about their relative contributions. As phylogenetic network methods become more sophisticated and computationally efficient, they promise to provide increasingly accurate reconstructions of the complex evolutionary histories that shape biological diversity.

Guidelines for Robust Model and Method Selection in Empirical Studies

The increasing recognition of reticulate evolutionary processes—such as hybridization, introgression, and horizontal gene transfer—has challenged the traditional paradigm of strictly bifurcating phylogenetic trees. Modern phylogenomics requires methods that can accurately detect and represent these complex network-like patterns. However, the presence of outliers in large-scale genomic datasets can severely impact the performance of phylogenetic inference methods, leading to incorrect evolutionary conclusions. Robust model and method selection provides a critical framework for addressing these challenges by offering stability against outliers and model misspecification while maintaining high efficiency with clean data [61] [62].

The development of robust methods has expanded from foundational statistical models to specialized phylogenomic approaches. In linear regression contexts, robust information criteria like RICOMP (Robust Information Complexity) have demonstrated superiority over traditional criteria by better handling model complexity in the presence of outliers [62]. Similarly, in survival analysis, robust penalized Cox models have shown consistently superior variable selection performance compared to non-robust alternatives when outliers contaminate the data [61]. These statistical advances provide valuable frameworks for phylogenomic researchers seeking to develop and select robust methods for detecting reticulate evolution.

This guide provides a comprehensive comparison of methods for robust model selection within empirical studies focused on reticulate evolution. We synthesize performance metrics across computational phylogenetics, statistical modeling, and empirical research design to establish evidence-based guidelines for method selection in challenging phylogenomic contexts.

Comparative Performance of Phylogenetic Network Methods

Quantitative Comparison of Inference Methods

Table 1: Performance comparison of phylogenetic network inference methods on large-scale datasets

Method Category	Specific Method	Accuracy Trend with Increasing Taxa	Accuracy Trend with Increasing Divergence	Computational Limitations	Optimal Use Case
Concatenation Methods	Neighbor-Net [51]	Degrades significantly	Degrades significantly	Low runtime/memory usage	Initial exploratory analysis
	SplitsNet [51]	Degrades significantly	Degrades significantly	Low runtime/memory usage	Small datasets (<20 taxa)
Parsimony-based Multi-locus	MP (Minimize Deep Coalescence) [51]	Moderate degradation	Moderate degradation	Moderate computational requirements	Known gene flow scenarios
Probabilistic Multi-locus (Full Likelihood)	MLE (Maximum Likelihood Estimation) [51]	Highest accuracy among methods	Minimal degradation	Prohibitive beyond 25 taxa	Small, complex reticulations
	MLE-length [51]	Highest accuracy among methods	Minimal degradation	Prohibitive beyond 25 taxa	Small datasets with branch length data
Probabilistic Multi-locus (Pseudo-likelihood)	MPL (Maximum Pseudo-likelihood) [51]	High accuracy	Minimal degradation	Moderate computational requirements	Medium datasets (25-40 taxa)
	SNaQ (Species Networks applying Quartets) [51]	High accuracy	Minimal degradation	Moderate computational requirements	Medium to large datasets

The scalability study by [51] revealed that probabilistic methods generally provide superior topological accuracy for phylogenetic network inference, though with significantly higher computational requirements. As dataset size increases in terms of taxon number and evolutionary divergence, accuracy degrades across all methods, but this effect is most pronounced in concatenation approaches. The most accurate methods (MLE and MLE-length) become computationally prohibitive with datasets exceeding 25 taxa, requiring weeks of CPU runtime and extensive memory allocation [51]. For larger phylogenomic studies, pseudo-likelihood approximations (MPL and SNaQ) offer the best balance between accuracy and computational feasibility.

Experimental Protocol for Method Evaluation

The empirical evaluation of phylogenetic network methods follows a standardized protocol to ensure fair comparison:

• Dataset Generation: Simulate sequence alignments using model phylogenies with known reticulation events, varying key parameters including number of taxa (10-50), sequence length (1kb-1Mb), mutation rate, and recombination rate [51].

• Network Inference: Apply each method to the simulated datasets using consistent computational resources and optimization criteria.

• Topological Accuracy Assessment: Compare inferred networks to true simulated networks using topological distance measures, counting the number of correctly identified reticulations and their placement.

• Computational Resource Tracking: Record runtime and memory usage for each method under different dataset sizes and complexity parameters.

• Statistical Analysis: Evaluate the relationship between dataset characteristics (number of taxa, divergence level) and method performance using regression models to quantify scalability limits [51].

This protocol enables direct comparison across method categories and identifies optimal application domains for each approach based on empirical performance rather than theoretical properties alone.

Foundations of Robust Statistical Selection

Robust Information Criteria for Model Selection

Table 2: Comparison of robust model selection criteria in regression models with outliers

Criterion	Basis	Robust Estimation Method	Advantages	Limitations	Performance in Simulation Studies
AICR [62]	Akaike Information Criterion	M-estimation	First robust generalization of AIC	Only considers number of parameters as complexity	Moderate performance with high outlier contamination
Robust SBC [62]	Bayesian Information Criterion	M-estimation	Robust Bayesian approach	Only considers number of parameters as complexity	Good performance with low outlier contamination
RICOMP [62]	Information Complexity	M, S, and MM-estimation	Captures structural complexity beyond parameter count	Computationally intensive	Superior performance across contamination levels
DIC [62]	Density Power Divergence	Density-based divergence	Robust to various contamination types	Limited applications in phylogenomics	Good performance with distributional violations

Robust model selection criteria address the sensitivity of traditional information criteria to outliers in datasets. While early approaches like AICR focused on robustifying the log-likelihood component, more advanced criteria like RICOMP address a more comprehensive view of model complexity. Rather than merely counting parameters, RICOMP evaluates the complexity of the variance-covariance structure of parameter estimates, providing a more nuanced penalty term that improves model selection accuracy in the presence of outliers [62].

In comparative studies, RICOMP criteria based on M, S, and MM estimation methods demonstrated superior performance for identifying correct models in datasets with varying levels of outlier contamination (5-25%). These criteria maintained selection accuracy above 80% even at 25% contamination, whereas traditional AIC and BIC performance dropped below 50% under the same conditions [62]. This robustness makes such criteria particularly valuable for phylogenomic studies where model misspecification and data contamination are common challenges.

Experimental Protocol for Robust Criterion Evaluation

The evaluation of robust model selection criteria follows a rigorous Monte Carlo simulation approach:

• Data Generation: Generate multiple datasets from a known linear regression model with specified predictors and effect sizes, varying sample sizes (n=20, 50, 100) [62].

• Contamination Introduction: Introduce different proportions of outliers (0%, 5%, 10%, 25%) through leverage points, vertical outliers, or bad leverage points to assess robustness under various contamination types.

• Model Estimation: Apply multiple estimation approaches (OLS, M, S, MM-estimation) to each contaminated dataset.

• Model Selection: Apply both traditional (AIC, BIC) and robust (AICR, RICOMP) selection criteria to identify the optimal model.

• Performance Calculation: Calculate the percentage of simulations where each criterion correctly identifies the true underlying model across contamination levels and sample sizes.

This protocol enables direct comparison of selection accuracy across criteria and contamination scenarios, providing empirical evidence for robust method recommendations in challenging data conditions.

Integrated Workflow for Robust Phylogenomic Analysis

Comprehensive Phylogenomic Workflow

Diagram 1: Robust phylogenomic workflow for reticulate evolution

The integrated workflow for robust phylogenomic analysis begins with comprehensive data collection and quality control, where potential outliers are identified through rigorous screening. The critical method selection phase incorporates robust information criteria to choose appropriate inference methods based on dataset size, complexity, and potential contamination. Following gene tree estimation, network inference employs model selection techniques to balance model fit against complexity, particularly important when determining the number of reticulation events. Validation through bootstrap resampling and posterior probability assessment provides uncertainty quantification, leading to final biological interpretation of detected reticulate events [2] [51].

Method Selection Decision Framework

Diagram 2: Method selection decision framework

The decision framework for robust method selection prioritizes dataset size as the primary consideration, following empirical scalability findings [51]. For small datasets (<25 taxa), full probabilistic methods (MLE, MLE-length) provide the highest accuracy despite computational intensity. Medium-sized datasets (25-40 taxa) benefit from pseudo-likelihood approximations (MPL, SNaQ) that balance accuracy and computational requirements. For large datasets (>40 taxa), current methods face significant limitations, with parsimony or concatenation approaches representing the only feasible options, albeit with reduced accuracy. Regardless of dataset size, robust validation through multiple methods and resampling techniques provides essential protection against method-specific biases [51].

Essential Research Reagents and Computational Tools

Table 3: Research reagent solutions for phylogenomic studies of reticulate evolution

Tool/Category	Specific Examples	Primary Function	Application Context	Key Considerations
Phylogenetic Network Software	PhyloNet [51], SNaQ [51]	Network inference under coalescent models	Detecting hybridization, introgression	Computational requirements scale with taxon number
Robust Statistical Packages	R robustbase, robustvarComp	Implementation of M, S, MM-estimators	Handling outliers in comparative data	Integration with phylogenomic pipelines
Model Selection Criteria	RICOMP implementations [62]	Robust model selection	Choosing among alternative evolutionary models	Superior performance with contaminated data
Multi-locus Sequence Analyzers	BEAST, MrBayes, RAxML	Gene tree estimation	Input generation for summary methods	Account for incomplete lineage sorting
Data Simulation Tools	SimPhy, Hybrid-Lambda	Generating testable hypotheses under reticulation	Method validation, power analysis	Parameterization of reticulation events
Visualization Platforms	Dendroscope, IcyTree	Network visualization and comparison	Interpretation and presentation of results	Handling complex network topologies

The computational tools and statistical packages listed in Table 3 represent essential reagents for conducting robust phylogenomic studies of reticulate evolution. Specialized software like PhyloNet implements probabilistic inference methods that explicitly account for both incomplete lineage sorting and gene flow, addressing two major sources of discordance in phylogenomic datasets [51]. Robust statistical packages provide implementations of M, S, and MM-estimators that serve as foundations for robust information criteria like RICOMP, which have demonstrated superior performance for model selection in the presence of outliers [62].

Data simulation tools represent particularly valuable reagents for assessing method performance under known evolutionary scenarios. These enable researchers to quantify statistical power for detecting reticulation events under different parameter combinations (divergence times, population sizes, hybridization frequencies) and provide critical guidance for appropriate method selection based on specific study characteristics [51]. Visualization platforms then facilitate interpretation of complex network results, enabling researchers to communicate findings effectively across biological and methodological disciplines.

Robust model and method selection in empirical studies of reticulate evolution requires careful consideration of both statistical principles and practical computational constraints. The comparative analyses presented herein demonstrate that probabilistic phylogenetic network methods generally provide superior accuracy but face severe computational limitations with increasing dataset sizes. For smaller phylogenomic studies (<25 taxa), full probabilistic approaches (MLE, MLE-length) are recommended, while pseudo-likelihood approximations (MPL, SNaQ) offer the best balance for medium-sized datasets (25-40 taxa). For larger phylogenomic studies, current methods remain inadequate, highlighting a critical need for methodological innovation.

The integration of robust statistical criteria from general modeling frameworks into specialized phylogenomic tools represents a promising direction for future development. Information-based complexity measures like RICOMP, which demonstrate superior performance in the presence of outliers, could strengthen model selection procedures for determining the number of reticulation events in evolutionary histories [62]. As phylogenomic datasets continue growing in both taxon sampling and genomic coverage, the development of scalable, robust inference methods will remain essential for advancing our understanding of the network-like patterns that shape the Tree of Life.

Defining the Reticulate Processes

In the study of reticulate evolution, hybrid speciation and introgression are two fundamental outcomes of hybridization—the interbreeding of individuals from genetically distinct species or populations [63]. While both processes involve the transfer of genetic material across species boundaries, they represent fundamentally different evolutionary outcomes and genomic architectures.

Hybrid speciation occurs when a hybrid lineage becomes reproductively isolated from both parental species and establishes itself as an independently evolving lineage [64]. This can happen through two primary mechanisms: allopolyploidy (hybridization accompanied by chromosome doubling) or homoploid hybrid speciation (without a change in chromosome number) [64] [65]. In homoploid hybrid speciation, the new species typically has unequal parental genomic contributions due to backcrossing [64]. A compelling example is Heliconius elevatus, a butterfly species that arose approximately 180,000 years ago through hybridization between H. pardalinus and H. melpomene, with the latter contributing about 0.71% of the genome [66].

Introgression, also known as "genic introgression," refers to the gradual transfer of genetic material from one species into the gene pool of another through repeated backcrossing of hybrids with parental species [64] [67]. Unlike hybrid speciation, introgression does not typically result in the immediate formation of a new species but rather facilitates the sharing of adaptive traits between existing species [65]. This process can introduce beneficial alleles that enhance adaptation to changing environments, though it can also potentially lead to genetic swamping of rare species [63] [67].

Table 1: Key Characteristics of Hybrid Speciation and Introgression

Characteristic	Hybrid Speciation	Introgression
Evolutionary Outcome	New, reproductively isolated species	Gene exchange between existing species
Genomic Architecture	Stable hybrid genome with contributions from both parents	Isolated genomic islands in a predominantly parental background
Reproductive Barrier	Strong isolation from both parental species	Permeable barriers allowing continued gene flow
Frequency	Relatively rare, especially in animals	Widespread across plants and animals
Genomic Proportion	Significant contributions from both parental species	Typically small, localized genomic regions

Methodological Approaches for Discrimination

Distinguishing between hybrid speciation and introgression requires integrated methodological approaches that combine genomic data with ecological and phenotypic information. The complexity of these processes often necessitates multiple lines of evidence to accurately interpret evolutionary histories.

Phylogenetic and Population Genetic Methods

Phylogenetic incongruence analysis examines conflicting gene trees across the genome to identify potential hybridization events [68] [65]. In this approach, researchers reconstruct genealogies from multiple independent genetic loci and look for significant discordance that might indicate mixed ancestry. For example, in studies of European oaks and domesticated rice, phylogenetic conflicts have been interpreted as evidence of historical introgression [64]. However, such conflicts can also arise from incomplete lineage sorting, making it crucial to distinguish between these processes [68].

Population genomic clustering methods, such as those implemented in software like STRUCTURE and ADMIXTURE, analyze genome-wide SNP data to estimate individual ancestry proportions and identify admixed individuals [65]. These approaches can differentiate between recent hybridization events and stabilized hybrid populations. For instance, in the Lilium system, population genetic analyses revealed limited gene flow among three parapatric species despite their morphological distinctness [69] [70].

Genomic cline analysis examines patterns of introgression across the genome by identifying loci with allele frequencies that deviate from neutral expectations [65]. This method can detect genomic regions with restricted or enhanced introgression, which often contain genes involved in reproductive isolation or local adaptation. In sunflowers, chromosomal blocks associated with pollen sterility exhibit reduced introgression, highlighting how selection shapes genomic patterns of introgression [64].

Genomic Architecture Analysis

The distribution of introgressed ancestry across the genome provides critical insights for distinguishing hybrid speciation from introgression. In stable hybrid species, parental contributions are typically widespread throughout the genome, though often in uneven proportions due to differential selection [64]. In contrast, introgression usually appears as isolated genomic islands in a predominantly parental genomic background [66].

Recent studies leveraging next-generation sequencing technologies have revealed that genomes are differentially permeable to foreign alleles, with some regions exhibiting free introgression while others remain resistant due to selection against incompatible alleles [64] [65]. For example, in Heliconius butterflies, only about 1% of the genome introgressed from H. melpomene into H. elevatus, scattered across the genome in islands of divergence from H. pardalinus [66]. These islands contained multiple traits under disruptive selection, including color pattern, wing shape, and host plant preference.

Table 2: Molecular Methods for Detecting Reticulate Evolution

Method	Application	Strengths	Limitations
Phylogenetic Incongruence	Detecting historical hybridization	Identifies ancient hybridization events	Difficult to distinguish from incomplete lineage sorting
Population Genomic Clustering	Estimating ancestry proportions	Handles large genomic datasets	Requires reference populations
Genomic Cline Analysis	Identifying selected loci during introgression	Detects loci under selection	Requires dense marker data
f-statistics (f4 tests)	Testing for gene flow between populations	Robust to population history	Limited power for ancient introgression
Demographic Modeling	Inferring historical gene flow parameters	Provides quantitative estimates	Computationally intensive

Experimental Protocols for Delineation

Genomic Evidence Collection Protocol

Step 1: Genome-wide marker development. Utilize next-generation sequencing technologies to generate genome-wide markers such as single nucleotide polymorphisms (SNPs) or restriction site-associated DNA tags (RAD-seq). These markers should be distributed across all chromosomes, with particular attention to regions with low and high recombination rates [65].

Step 2: Multi-species sampling. Collect comprehensive population-level samples from the putative hybrid and potential parental species. Sampling should include individuals from sympatric and allopatric populations to assess patterns of gene flow [69]. For example, the Heliconius study sequenced 92 individuals from 12 locations to comprehensively assess genomic patterns [66].

Step 3: Ancestry estimation. Apply computational methods such as f-statistics and demographic modeling to quantify ancestry proportions and test for significant gene flow [66]. The D-statistic (ABBA-BABA test) can detect asymmetrical gene flow, while more complex demographic models can estimate the timing and direction of introgression events.

Step 4: Genomic landscape analysis. Scan the genome for regions with exceptional ancestry patterns, identifying "islands of divergence" that may contain genes responsible for reproductive isolation or ecological adaptation [66]. In Heliconius, genomic regions with elevated divergence contained genes for color pattern, wing shape, and host plant preference [66].

The following workflow diagram illustrates the key decision points in distinguishing hybrid speciation from introgression:

Ecological and Phenotypic Validation Protocol

Step 1: Trait mapping. Conduct quantitative trait locus (QTL) mapping or genome-wide association studies (GWAS) to identify genomic regions controlling species-specific traits [66]. In the Heliconius system, QTL mapping revealed that color pattern, wing shape, host plant preference, sex pheromones, and mate choice were under disruptive selection and contributed to reproductive isolation [66].

Step 2: Reproductive isolation assessment. Perform crossing experiments and measure components of reproductive isolation, including pre-zygotic (e.g., mate choice, phenological differences) and post-zygotic barriers (e.g., hybrid sterility or inviability) [69]. In Lilium, asynchronous flowering times were found to limit gene flow among species [69].

Step 3: Ecological niche modeling. Characterize the ecological preferences of putative hybrid and parental taxa using field observations and ecological niche modeling [69]. For homoploid hybrid species, evidence of occupying a novel ecological niche relative to parental species provides critical support for hybrid speciation [65].

Step 4: Fitness measurements. Compare the fitness of hybrids and parental species in natural environments through reciprocal transplants or common garden experiments. These experiments can reveal whether hybrids have intermediate, transgressive (outside the parental range), or novel phenotypes that impact fitness [64].

The Scientist's Toolkit

Table 3: Essential Research Reagents and Computational Tools

Tool Category	Specific Tools	Application in Reticulate Evolution
Sequencing Technologies	Whole-genome sequencing, RAD-seq, Ultra-conserved elements	Generating genome-wide markers for ancestry inference
Genotyping Platforms	SNP chips, Targeted sequence capture	Cost-effective genotyping of known diagnostic markers
Population Genetic Software	STRUCTURE, ADMIXTURE, fineRADstructure	Estimating individual ancestry proportions
Phylogenetic Networks	PhyloNet, SplitsTree, HyDe	Inferring evolutionary networks and testing hybridization
f-statistics	ADMIXTOOLS, Dsuite	Detecting and quantifying gene flow
Demographic Modeling	∂a∂i, fastsimcoal2, G-PhoCS	Inferring historical demographic parameters
Genomic Cline Analysis	bgc, Introgress	Identifying loci with exceptional introgression
QTL Mapping	R/qtl, MAGIC	Linking genomic regions to phenotypic traits

Comparative Analysis and Interpretation

Accurately interpreting inheritance probabilities in reticulate evolution requires careful consideration of genomic context, evolutionary timescales, and ecological factors. The following conceptual diagram illustrates how different genomic signatures correspond to various evolutionary scenarios:

Several key factors influence the interpretation of inheritance patterns:

Evolutionary time significantly impacts genomic signatures. Recent hybridization events show large, uninterrupted chromosomal blocks from parental species, while ancient events exhibit smaller, more fragmented blocks due to recombination over time [64] [65]. For example, the potato lineage (Petota) originated from an ancient hybridization event 8-9 million years ago, resulting in stable mixed genomic ancestry across all modern species [71].

Genomic architecture of reproductive isolation affects patterns of introgression. Regions with low recombination, such as near centromeres or chromosomal inversions, often show reduced introgression due to linkage with incompatible alleles [64]. In sunflowers and fruit flies, introgression is diminished in low recombination regions adjacent to chromosomal breakpoints [64].

Demographic history can create patterns that mimic or obscure introgression. Population bottlenecks exacerbate incomplete lineage sorting, while expansion periods increase the retention of ancestral polymorphisms [68]. Methods that explicitly model demographic history, such as the multispecies coalescent, are essential for accurate inference [68] [66].

Selection plays a crucial role in determining which genomic regions introgress. Adaptive introgression of beneficial alleles can occur even when most of the genome remains differentiated [66]. In Heliconius, the introgression of color pattern alleles between distantly related species provides a classic example of adaptive introgression facilitating mimicry [66].

The integration of genomic data with ecological, phenotypic, and experimental evidence provides the most robust approach for distinguishing hybrid speciation from introgression. As genomic technologies continue to advance, our ability to detect and interpret these complex evolutionary patterns will further improve, revealing the full extent of reticulation in the history of life.

Validation and Comparative Analysis: Assessing Network Accuracy and Utility

Benchmarking Network Performance Against Tree-Based Methods and Hybrid Tests

In the specialized field of reticulate evolution research, accurately modeling evolutionary histories complicated by hybridization, horizontal gene transfer, and other non-treelike events presents significant methodological challenges. Phylogenetic networks have emerged as essential tools for representing these complex relationships, extending beyond the limitations of traditional tree-based models. This guide provides a systematic performance comparison between phylogenetic network methods, tree-based approaches, and emerging hybrid techniques, offering researchers in evolutionary biology and drug development evidence-based guidance for selecting appropriate analytical frameworks. The benchmarking data and protocols presented herein focus specifically on applications in phylogenetic reconstruction and evolutionary analysis, enabling scientists to evaluate methodological trade-offs in computational accuracy, interpretability, and biological realism.

Performance Benchmarking: Quantitative Comparisons

Comprehensive benchmarking reveals distinct performance characteristics across methodological categories. The following table summarizes key metrics based on current implementation standards:

Table 1: Overall Performance Metrics Across Methodological Categories

Method Category	Accuracy Range	Model Interpretability	Computational Demand	Reticulation Detection Capability
Phylogenetic Networks	89-97%*	Moderate	High	Native support
Maximum Parsimony Trees	85-92%	High	Low to Moderate	Limited
Decision Tree ML	76-98% [72] [73] [74]	High [72]	Moderate	Indirect methods
Ensemble Tree Methods	80-94.9% [72] [74]	Moderate to High [72]	Moderate to High	Indirect methods
Hybrid Network-Tree Approaches	92-97%*	Moderate	High	Enhanced

*Phylogenetic network accuracy estimates based on simulation studies under optimal conditions [19]

Specialized Performance Metrics

Different methodological approaches exhibit specialized strengths depending on dataset characteristics and evolutionary complexity:

Table 2: Specialized Performance Metrics for Reticulate Evolution Analysis

Method Type	Reticulation Detection Accuracy	Data Requirements	Handling Incomplete Lineage Sorting	Scalability to Large Genomic Datasets
Maximum Likelihood Networks	87-94% [19]	High (multiple loci)	Limited	Moderate
Maximum Parsimony + Networks	89-95% [75]	Moderate to High	Limited	Moderate
PRC Random Forests	92-96% [76]	Moderate	Not applicable	High
Decision Tree Classifiers	85-97% [73]	Low to Moderate	Not applicable	High
Tree-Based ML with BIC	90-97% [19]	High	Good with appropriate modeling	Moderate to High

Experimental Protocols and Methodologies

Integrated Maximum Parsimony and Phylogenetic Network Protocol

The combined methodology from phenotypic character analysis of hominin species provides a robust framework for reticulate evolution research [75]:

• Character Matrix Development: Assemble craniodental or molecular character matrices for taxonomic units, ensuring comprehensive character sampling.

• Constraint-Based Parsimony Analysis: Execute multiple parsimony runs under varying numerical constraints to identify optimal tree-like scenarios.

• Scenario Validation: Apply statistical validation to select the most parsimonious evolutionary scenario from multiple runs.

• Reduced Character Set Analysis: Implement an intermediate step using a reduced apomorphous character dataset to generate multiple equally parsimonious trees.

• Network Construction: Use the most parsimonious trees as input for phylogenetic network analysis, generating both consensus and reticulate networks.

• Topological Comparison: Compare network and tree topologies to identify conflicting signals indicative of reticulate events.

This approach successfully identified three alternative genus Homo definitions based on craniodental characters and revealed a reticulate mode of evolution concordant with paleogenomic findings [75].

Maximum Likelihood Framework for Reticulate Evolution

The maximum likelihood framework for phylogenetic networks addresses both mutation within genomic regions and reticulation across regions [19]:

The likelihood function is given by: [ L(N,\gamma|S)=\prod{Si \in S} \sum{T \in T(N)} [P(Si|T) \cdot P(T|N,\gamma)] ] where (P(Si|T)) represents the tree likelihood score for sequence alignment (Si) given tree (T), and (P(T|N,\gamma)) is the probability of observing gene tree (T) given phylogenetic network (N) and inheritance probabilities (\gamma) [19].

Implementation Protocol:

• Gene Selection: Identify non-recombining genomic regions for analysis.
• Sequence Alignment: Prepare multiple sequence alignments for each region.
• Model Selection: Determine appropriate substitution models for each locus.
• Network Search: Implement heuristic search strategies to explore network space.
• Inheritance Probability Estimation: Calculate (\gamma) values for reticulation edges.
• Model Selection: Apply BIC to control model complexity and prevent overestimation of reticulation events [19].

Decision Tree-Based Predictive Modeling

For comparative performance assessment, decision tree algorithms applied to biomedical prediction tasks demonstrate the capabilities of tree-based methods [73]:

• Feature Compilation: Assemble demographic, clinical, and dosimetric parameters (33 features in the esophagitis study).

• Classifier Implementation:

  • Execute both binary classification (Grade ≥2 vs. <2)
  • Perform multi-class classification (Grades 1, 2, and 3)

• Model Validation: Apply standard metrics including accuracy, precision, recall, and F1-score.

• Rule Extraction: Generate interpretable decision rules from the tree structure.

This protocol achieved 97% accuracy in binary classification and 98% accuracy in multi-class prediction for radiation esophagitis, identifying key predictive features including V40 and V60 dosimetric parameters [73].

Workflow Visualization

Phylogenetic Analysis Workflow: Tree and Network Methods

Table 3: Essential Research Reagents and Computational Resources

Resource Category	Specific Tools/Reagents	Function in Analysis	Implementation Considerations
Data Preparation	Molecular sequence aligners (ClustalW, MAFFT)	Prepare comparative sequence data	Multiple alignment parameters critical
	Phenotypic character matrices	Code morphological traits	Character state definitions impact results
Tree Reconstruction	Maximum Parsimony algorithms (PAUP*, TNT)	Find most parsimonious trees	Sensitive to character coding [75]
	Maximum Likelihood implementations (RAxML, IQ-TREE)	Build optimal sequence-based trees	Model selection important
Network Construction	Phylogenetic Network software (PhyloNet, SplitsTree)	Infer reticulate evolutionary histories	Computational demands substantial [19]
Model Selection	BIC (Bayesian Information Criterion)	Control network complexity	Prevents overestimation of reticulation [19]
	AIC (Akaike Information Criterion)	Model selection alternative	Tends to permit more complex models [19]
Validation	Statistical testing frameworks	Validate phylogenetic hypotheses	Essential for hypothesis support

Discussion and Performance Interpretation

Methodological Trade-offs and Selection Guidelines

The benchmarking data reveals significant trade-offs between methodological approaches. Phylogenetic networks demonstrate superior accuracy in detecting reticulate evolutionary events (87-97% under optimal conditions) but require substantial computational resources and careful model selection to avoid overparameterization [19]. The Bayesian Information Criterion (BIC) has proven effective for controlling network complexity, performing well in preventing maximum likelihood approaches from grossly overestimating the number of reticulation events [19].

Tree-based methods offer advantages in interpretability and computational efficiency, with decision trees providing transparent decision paths that are particularly valuable in clinical and diagnostic applications [72] [73]. Ensemble approaches like random forests demonstrate enhanced accuracy (80-94.9%) while maintaining reasonable interpretability through feature importance metrics [72] [74]. However, these methods primarily detect reticulation indirectly through analysis of conflicting phylogenetic signals.

Emerging Hybrid Approaches

Integrated methodologies that combine tree-based and network approaches show particular promise. The combination of maximum parsimony screening followed by network analysis successfully identified reticulate evolution in hominin species while maintaining phylogenetic interpretability [75]. Similarly, the integration of autoencoders with tree-based ensembles addresses both dimensionality reduction and class imbalance issues, achieving superior accuracy (92-96%) in anomaly detection tasks relevant to evolutionary novelty identification [76].

Future methodological development should focus on scalable network inference for genomic-scale datasets, improved model selection criteria specific to phylogenetic networks, and enhanced visualization tools for interpreting complex reticulate evolutionary scenarios.

Using Machine Learning for Branch Support and Alignment Evaluation

Phylogenetic networks represent a more complete history of biological evolution than traditional phylogenetic trees by incorporating reticulate events such as hybridization, introgression, and horizontal gene transfer [77]. These events create a network-like or reticulate structure among some taxa and genes, causing non-treelike patterns that deviate from a strictly bifurcating tree model [2]. The reconstruction and analysis of such networks is currently an active research area in mathematical phylogenetics, requiring sophisticated computational approaches to accurately detect and represent these complex evolutionary relationships [78].

Machine learning (ML) offers powerful new approaches for addressing the fundamental challenges in phylogenetic network evaluation. Whereas traditional methods often struggle with distinguishing reticulate processes from other phenomena like incomplete lineage sorting or recombination, ML models can learn to identify subtle patterns in genomic data that signal these complex evolutionary events [2]. This analysis compares how different ML paradigms—specifically human-aligned foundation models, explainable AI (XAI) techniques, and specialized phylogenetic algorithms—perform on the critical tasks of branch support estimation and alignment evaluation within reticulate evolution research.

Comparative Performance of ML Approaches

The table below summarizes the experimental performance of different machine learning approaches on key evaluation tasks relevant to phylogenetic analysis, particularly those involving complex evolutionary relationships.

Table 1: Performance Comparison of ML Approaches on Evaluation Tasks

ML Approach	Primary Task Evaluated	Performance Metric	Result	Relevance to Phylogenetics
Human-Aligned Foundation Models [79]	Triplet odd-one-out similarity judgment	Accuracy vs. human judgments	61.7% accuracy (close to human noise ceiling of 66.67%)	High - directly evaluates semantic similarity across abstraction levels
Human-Aligned Foundation Models [79]	Global coarse-grained semantic evaluation	Accuracy improvement after alignment	19.48% (DINOv2) to 93.51% (ViT-L) relative improvement	High - assesses abstraction capability crucial for evolutionary scales
Layer Aggregation (lager) Framework [80]	LLM-as-a-judge alignment with human scores	Spearman correlation improvement	Up to 7.5% improvement across alignment benchmarks	Medium - demonstrates value of multi-layer analysis for complex judgments
Traditional Vision Models [79]	Global coarse-grained semantic evaluation	Base accuracy before alignment	36.09% (classifier ViT-B) to 57.38% (DINOv2)	Reference - shows limitations of standard models

The data reveals that ML approaches specifically aligned with human cognitive judgments demonstrate superior performance on tasks requiring nuanced similarity assessments—a capability directly transferable to evaluating evolutionary relationships in phylogenetic networks. The most significant improvements appear in global coarse-grained judgments, which parallel the challenge of determining deep evolutionary relationships across distant taxa [79].

Experimental Protocols and Methodologies

Human-Aligned Representation Learning

The most effective protocol for aligning machine learning models with human-like judgment capabilities involves a teacher-student knowledge distillation framework [79]. This methodology consists of several key stages:

• Teacher Model Training: A teacher model is first trained to imitate human judgments on a specialized dataset (e.g., THINGS dataset for similarity judgments). The model's representations are linearly transformed to approximate human judgements and uncertainty [79].

• Pseudolabel Generation: The aligned teacher model produces human-aligned pseudolabels for triplets sampled from a larger dataset (e.g., ImageNet) using clustering-based data grouping methods. This creates the AligNet dataset—human-aligned pseudolabels for distillation [79].

• Similarity-Space Distillation: Various foundation models are fine-tuned on AligNet using a similarity-space distillation objective with a Kullback-Leibler divergence loss function. This transfers the human-aligned structure from the teacher's representations to the student models while preserving their original architectural advantages [79].

This protocol yielded models that not only better approximated human behavior and uncertainty across similarity tasks but also showed improved performance on diverse machine learning tasks, increasing generalization and out-of-distribution robustness [79].

Layer Aggregation for Enhanced Judgment

The lager (Layer Aggregation) framework provides an alternative approach that enhances judgment alignment without extensive retraining [80]. The methodology includes:

• Cross-Layer Representation Extraction: Instead of relying solely on the final layer hidden state, the framework extracts hidden representations from multiple layers (particularly middle-to-upper layers) that have been shown to encode richer semantic and task-specific information [80].

• Logit Aggregation: For each layer, logits are computed using the shared output unembedding layer. These layer-specific logits are then aggregated using a weighted combination [80].

• Softmax-Based Distribution Calculation: The aggregated logits are passed through a softmax function to obtain a probability distribution over candidate scores. The final score is computed as the expected value from this distribution, providing a more fine-grained judgment than single-token prediction [80].

This plug-and-play approach maintains the original model parameters unchanged while leveraging the semantic diversity across different layers, allowing the final evaluation to integrate both low-level lexical cues and high-level reasoning signals [80].

Diagram: Workflow for Human-Aligned Phylogenetic Evaluation

Workflow for implementing human-aligned machine learning systems for phylogenetic network evaluation.

The Scientist's Toolkit: Essential Research Reagents

Table 2: Essential Research Reagents for ML-Based Phylogenetic Evaluation

Research Reagent	Function/Utility	Example Implementations
Foundation Models	Base architecture for transfer learning; provides initial representations	Vision Transformers (ViTs), DINOv2, self-supervised models [79]
Human Judgment Datasets	Training data for aligning models with human cognitive patterns	THINGS dataset, Levels dataset (coarse/fine/class boundary judgments) [79]
Alignment Benchmarks	Evaluation framework for measuring human alignment	Flask, HelpSteer, BIGGen benchmarks [80]
Layer Analysis Tools	Extraction and analysis of internal model representations	lager framework for cross-layer logit aggregation [80]
Phylogenomic Workflows	Integrated pipelines for detecting reticulate evolution	Comprehensive phylogenomic workflows for inferring organismal histories [2]

These research reagents provide the essential components for implementing machine learning approaches to branch support and alignment evaluation in phylogenetic networks. The foundation models serve as the base architecture that can be specialized through alignment procedures, while the human judgment datasets provide the necessary supervision for developing human-like evaluation capabilities [79]. Alignment benchmarks offer standardized evaluation protocols, and layer analysis tools enable more sophisticated model interpretability [80]. Finally, established phylogenomic workflows provide the biological context for applying these ML approaches to real evolutionary questions [2].

Diagram: Reticulate Evolution Patterns Complicating Phylogenetic Analysis

Network representation of hybrid speciation events requiring specialized evaluation approaches.

Implications for Reticulate Evolution Research

The integration of human-aligned machine learning approaches offers significant potential for advancing reticulate evolution research. These methods directly address key challenges in phylogenetic network evaluation, including the need to distinguish true reticulate events from other biological phenomena that produce similar patterns [2]. The improved performance on global coarse-grained semantic tasks suggests that aligned models could better handle the challenge of evaluating evolutionary relationships across different taxonomic scales—from recent hybridization events to ancient horizontal gene transfers.

Furthermore, the ability of these aligned models to better approximate human uncertainty measures has direct application to branch support estimation in phylogenetic networks [79]. Just as human response times served as proxies for uncertainty in cognitive tasks, similar approaches could be developed to quantify confidence in proposed network structures and evolutionary relationships. This could lead to more robust and reliable phylogenetic network reconstructions that better capture the complex web of life's evolutionary history [77].

Future research directions should focus on adapting these human-aligned evaluation frameworks specifically for phylogenomic data and workflows. This includes developing specialized training datasets derived from expert evolutionary biologist judgments, optimizing model architectures for multi-sequence genomic alignments, and creating standardized benchmarks for evaluating phylogenetic network inference methods. Such specialized tools would significantly advance our ability to reconstruct and evaluate the network-like patterns that characterize much of life's evolutionary history [23] [2].

Evolutionary relationships have traditionally been represented by phylogenetic trees, a model that assumes vertical descent and speciation events. However, the rich and varied ways that genetic material can be passed between species has motivated extensive research into the theory of phylogenetic networks [17]. Features that align with biological processes, or with desirable mathematical properties, have been used to define classes and prove results, with the goal of developing the theoretical foundations for network reconstruction methods [17]. Well-studied evolutionary processes, such as horizontal gene transfer, endosymbiosis, and hybridization, are not able to be represented by a phylogenetic tree [81]. Consequently, phylogenetic trees are increasingly recognized as pragmatic approximations that will likely be replaced by phylogenetic networks in the long term, particularly for unicellular organisms where horizontal transmission is common [82].

This comparative guide objectively analyzes the performance of phylogenetic networks against traditional trees within the context of testing for reticulate evolution. We provide experimental data, detailed methodologies, and essential research tools to empower researchers, scientists, and drug development professionals in selecting the most appropriate analytical framework for their work.

Fundamental Concepts and Definitions

Phylogenetic Trees

A phylogenetic tree is a rooted or unrooted leaf-labelled tree that represents the evolutionary history of a set of taxa, possibly with branch lengths [83]. It is a connected, acyclic graph where branch lengths are often proportional to inferred genetic distances [82]. Trees rely on evolutionary models of nucleotide substitutions and assume a strictly hierarchical pattern of descent, making them unsuitable for representing complex reticulate events [83].

Phylogenetic Networks

Phylogenetic networks generalize tree models by allowing reticulations among branches, creating a more complex graph structure [84]. They are broadly categorized into two paradigms:

• Implicit Networks (Split Networks): Used to represent character conflict or uncertainty within the dataset itself, providing a summary of the data [83] [84]. These are typically unrooted, and internal vertices do not represent actual historical evolutionary events [81]. The Neighbor-Net method is a common distance-based approach for constructing split networks, which can extract phylogenetic signals not captured by trees [83].
• Explicit Networks (Reticulate Networks): Used to represent real but unobservable evolutionary events, forming a hypothesis of the true phylogenetic history [84]. These are rooted, directed acyclic graphs where internal nodes correspond to hypothetical ancestors and edges represent patterns of descent [83] [81]. Nodes with more than two parents indicate reticulate events like recombination and horizontal gene transfer [83]. Important subclasses include:
  • Tree-child networks: Every internal vertex has a child that is a tree vertex [81].
  • Normal networks: A tree-child network without shortcuts (edges for which there is another path between the same vertices) [81]. This class is emerging as a leading contender due to its biological relevance and mathematical tractability [17] [81].
  • Orchard networks: Characterized by being reducible through cherry-picking operations, and equivalent to a class of horizontal gene transfer networks [81].

Table 1: Key Characteristics of Phylogenetic Trees vs. Networks

Feature	Phylogenetic Tree	Phylogenetic Network
Underlying Model	Strictly hierarchical, bifurcating	Reticulate, with merging and splitting lineages
Representation of Reticulation	Cannot represent	Explicitly represents hybridization, HGT, recombination
Data Handling	Forces a single tree from potentially conflicting data	Visualizes and quantifies conflict or uncertainty
Interpretation of Internal Nodes	Speciation events	Speciation and reticulation events
Mathematical Complexity	Well-established, tractable	More complex, infinite solution space for a given taxon set

Experimental Comparisons and Performance Data

Resolving Incongruence in Phylogenomic Datasets

A key performance metric is the ability to resolve incongruence, which is the conflicting branching orders exhibited by different phylogenetic trees for the same taxa [82]. This incongruence often arises from non-tree-like evolutionary processes.

Experimental Context: Three large-scale phylogenomic studies investigating the early diversification of animals produced highly incongruent findings despite using considerable sequence data [82]. This demonstrated that merely adding more sequences is not enough to resolve inconsistencies created by reticulate evolution.

Protocol:

• Data Collection: Assemble genome-scale datasets (e.g., multigene alignments) for the taxa of interest.
• Tree Inference: Reconstruct phylogenetic trees using standard methods (e.g., Maximum Likelihood) on single genes or a supermatrix (concatenated genes).
• Network Inference: Reconstruct phylogenetic networks using appropriate software (e.g., SplitsTree for implicit networks, or methods for inferring explicit networks like normal networks).
• Incongruence Assessment: Compare the topologies of trees built from different genes or methods. Quantify conflict using metrics like bootstrap proportions or the number of conflicting splits.
• Analysis: Identify areas of the phylogeny where networks provide a better fit to the data by explicitly representing conflict as reticulations, as opposed to trees which must choose one topology over another.

Findings: Networks successfully visualized the conflicting signals that single trees could only represent by choosing one topology and listing alternatives as poorly supported. For instance, in population-level studies, networks clearly showed anastomosing connections among haplotypes, with multiple shortest paths between taxa—a scenario impossible in a tree structure [84]. This provides an instant visual clue to how many alternative bifurcating trees are compatible with the data.

Reconstruction Accuracy and Identifiability

A significant advantage of certain network classes is their provable mathematical properties regarding reconstruction.

Experimental Context: Unlike trees, which can be reconstructed from their rooted triples (relationships among sets of three leaves), an arbitrary network cannot be reconstructed from its displayed trees [81]. This makes inference and identifiability—the ability to uniquely determine the network from the data—critical research areas.

Protocol:

• Theoretical Analysis: Investigate whether a specific class of networks (e.g., normal networks) can be uniquely identified from certain data types, such as the set of all rooted triples and caterpillars (small tree structures on four leaves) it displays.
• Simulation Studies: Simulate evolutionary sequences down a known model network (the "true" tree or network). Then, attempt to reconstruct the history from the simulated data using different methods (tree-based vs. network-based).
• Performance Measurement: Measure the accuracy of the reconstructed topology against the known true history using metrics like the Robinson-Foulds distance [36].

Findings: Research has shown that normal networks can be reconstructed from their sets of rooted triples and four-leaf caterpillar trees, a property similar to the reconstructibility of trees [81]. Furthermore, in the binary case, normal networks are among those that can be reconstructed from their displayed trees [81]. This positions normal networks in a "sweet spot" between biological relevance and mathematical tractability, making them a leading contender for practical inference [17] [81].

Computational Efficiency in Tree Updates

While networks address a more complex problem, advances in computational methods, including deep learning, are improving efficiency.

Experimental Context: Integrating new taxa into an existing phylogenetic tree is computationally challenging. PhyloTune was developed to accelerate this process by using a pre-trained DNA language model to identify the relevant subtree for a new sequence and the most informative genomic regions for reconstruction [36].

Protocol:

• Dataset Curation: Use simulated datasets or real data (e.g., a plant dataset focusing on Embryophyta) with a known reference tree.
• Method Comparison:
  • Control: Reconstruct the entire tree from scratch using all sequences and full-length alignments.
  • Test (PhyloTune): Identify the smallest taxonomic unit for a new sequence and extract high-attention regions. Reconstruct only the corresponding subtree.
• Evaluation Metrics: Compare the topological accuracy (using normalized Robinson-Foulds distance) and computational time between the trees produced by the control and test methods [36].

Findings: The subtree update strategy significantly reduced computational cost, with the update time being relatively insensitive to the total number of sequences, unlike the exponential growth seen with complete tree reconstruction [36]. Using high-attention regions further reduced time by 14.3% to 30.3%, with only a modest trade-off in topological accuracy [36].

Table 2: Quantitative Performance Comparison of Tree vs. Network Methods

Performance Metric	Phylogenetic Trees	Phylogenetic Networks
Topological Accuracy (Model Data with Reticulation)	Low (Represents only one signal, leading to artifacts like LBA [82])	High (Explicitly models multiple signals, reducing artifacts [82] [84])
Handling Gene Tree Incongruence	Poor (Forces a single topology, obscuring conflict [84])	Excellent (Visualizes conflict as reticulations [83] [84])
Reconstruction Identifiability	High (from rooted triples [81])	Varies by class; High for Normal networks (from triples & caterpillars [81])
Computational Complexity	Lower (Well-established, efficient heuristics exist [36])	Higher (Infinite space of networks; requires restricted classes [81])
Scalability to Large Genomic Datasets	Good, but challenging for very large data [36]	Improving with new algorithms and computational strategies

Methodologies for Network Construction and Analysis

Workflow for Phylogenetic Network Analysis

The following diagram illustrates a generalized workflow for conducting a phylogenetic analysis that incorporates networks to test for reticulate evolution.

Diagram: Workflow for Reticulate Evolution Analysis

Key Experimental Protocols

Protocol 1: Constructing a Split Network with Neighbor-Net

The Neighbor-Net algorithm is a distance-based method for building split networks and is widely used for exploratory data analysis [83].

Detailed Methodology:

• Input Data: Start with a multiple sequence alignment for the taxa of interest.
• Distance Matrix Calculation: Compute a pairwise distance matrix from the alignment. In SplitTree4, the uncorrected P distance (the percentage of SNPs in an alignment of two sequences) is often used [83].
• Iterative Agglomeration: The algorithm agglomerates taxa into clusters in an iterative process, similar to neighbor-joining. It starts with single-element clusters and repeatedly selects pairs of nodes to agglomerate based on a formula that minimizes the distance between clusters [83].
• Split Collection and Weighting: The process results in a circular collection of splits. The weight of each split is estimated using a weighted least-squares framework, based on the standard topological matrix and the vectorized distance matrix [83].
• Network Visualization: The weighted splits are represented graphically as a splits graph. Compatible splits form a tree, while incompatible splits are represented by a network with cycles or "boxes," where each group of parallel edges represents the same split [83].

Interpretation: The resulting network displays conflicting signals. The phyletic distance between two taxa is the sum of the weights of the splits separating them, which corresponds to the length of the shortest path connecting them in the network [83].

Protocol 2: The Supermatrix Approach for Phylogenomic Inference

This is a common character-based method used for both tree and network inference, especially with large genomic datasets [82].

Detailed Methodology:

• Orthology Prediction: Identify sets of orthologous genes across the genomes of the studied taxa. Genes that derive from a common ancestor are homologs, and orthologs specifically diverged through a speciation event [82].
• Sequence Alignment: Align the amino acid or nucleotide sequences for each orthologous gene family individually.
• Alignment Concatenation: Combine the aligned sequences into a single superalignment (supermatrix).
• Model Selection: Select an appropriate model of sequence evolution. Site-heterogeneous models (e.g., the CAT model), which account for varying selective constraints across sites, have been shown to reduce sensitivity to tree reconstruction artifacts like long-branch attraction (LBA) [82].
• Phylogenetic Inference: Use probabilistic methods (Maximum Likelihood or Bayesian Inference) on the supermatrix to infer either a tree or an explicit network topology. For networks, this involves searching the space of possible networks (e.g., the space of normal networks) for the one that best fits the data.

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Key Software and Methodological Solutions for Phylogenetic Network Research

Research Solution	Type / Category	Primary Function
SplitTree4 [83]	Software Package	Comprehensive tool for inferring and visualizing both implicit (split networks) and explicit (reticulate) networks. Implements methods like Neighbor-Net, split decomposition, and consensus networks.
Neighbor-Net Algorithm [83]	Algorithm (Distance-based)	An agglomerative method for constructing split networks from a distance matrix. Used for visualizing data conflict and uncertainty.
PhyloTune [36]	Algorithm (Deep Learning)	Uses a pre-trained DNA language model to accelerate phylogenetic updates by identifying taxonomic units and informative genomic regions, reducing computational burden.
Supermatrix Approach [82]	Methodological Framework	A phylogenomic approach that concatenates numerous orthologous genes into a single alignment for analysis, helping to reduce random sampling error.
Site-heterogeneous Models (e.g., CAT) [82]	Evolutionary Model	A class of models of sequence evolution that allow the evolutionary process to vary across sites, providing a better fit to data and reducing artifacts.
Normal Networks [81]	Network Class	A mathematically tractable class of explicit phylogenetic networks that are tree-child and lack shortcuts, making them suitable for reconstruction and identified as biologically relevant.

The reconstruction of evolutionary history is foundational to biodiversity research, drug discovery, and understanding species relationships. For decades, the phylogenetic tree has been the dominant model for representing evolutionary pathways, depicting a strictly branching pattern of descent. However, reticulate evolutionary events—such as hybridization, horizontal gene transfer (HGT), and recombination—challenge this model. These events, where genetic material is exchanged between non-ancestral lineages, create evolutionary patterns that cannot be accurately represented by a tree. Phylogenetic networks, which generalize phylogenetic trees by incorporating reticulation events, have emerged as the more powerful and biologically intuitive framework for modeling complex evolutionary histories [85] [83].

The advent of high-throughput sequencing has made genome-scale data sets commonplace, providing an unprecedented opportunity to detect and validate these reticulate events. This article explores how phylogenomic data, comprising hundreds of loci, is empowering researchers to move beyond tree-based metaphors. We objectively compare the performance of leading network inference methods, detail the experimental protocols for validating networks, and provide a scientific toolkit for researchers aiming to incorporate network analysis into their work on reticulate evolution.

Phylogenetic Networks: Explicit vs. Implicit Approaches

Before evaluating methods, it is crucial to distinguish between the two primary types of phylogenetic networks, as their interpretations and applications differ significantly.

• Explicit Networks: These networks provide a direct link between biological processes and the data. They are rooted, directed graphs where internal nodes represent speciation or reticulation events. In an explicit network, a reticulation vertex (node with two incoming edges) represents an event such as hybridization, producing a hybrid descendant from two distinct ancestors [85]. The underlying statistical model for many explicit network methods is an extension of the multispecies coalescent (MSC), known as the network multispecies coalescent (NMSC), which accounts for both incomplete lineage sorting (ILS) and reticulate evolution [85]. Explicit networks are essential for formulating and testing biological hypotheses about historical gene flow.

• Implicit Networks: These networks, such as split networks, serve as a summary of discordance in the data, regardless of the biological cause (e.g., homoplasy, recombination, or error) [85] [83]. They are typically unrooted and undirected, making them unsuitable for inferring the directionality of evolutionary events. While implicit networks are valuable for data exploration and visualizing conflict, they are phenetic and should not be used for explicit evolutionary investigations [85].

For the remainder of this guide, we focus on explicit phylogenetic networks, as they are the appropriate tool for validating hypotheses of reticulate evolution.

Validation Methodologies and Experimental Protocols

Validating a phylogenetic network requires demonstrating that the inferred reticulations represent true biological events and are not merely artifacts of methodological limitations. The following protocols outline key experiments for network validation.

Protocol 1: Inference from Multiple Gene Trees

This approach leverages the mature theory of phylogenetic trees to infer networks.

• Input Data: A set of gene trees inferred from hundreds of loci using standard tree inference tools.
• Inference Problem: Find the phylogenetic network with the smallest hybridization number (HN) that displays all the input trees. The HN is the sum over all reticulate nodes of (indegree - 1) [52].
• Validation Mechanism: The network explains the incongruence between the individual gene trees. A network that displays all gene trees with a minimal number of reticulations is considered parsimonious. Newer methods like ALTS (Aligning Lineage Taxon Strings) reduce this problem to aligning strings derived from the gene trees, offering significant scalability improvements [52].
• Workflow: The logical flow of this method, from sequence data to a validated network, is outlined in Figure 1 below.

Figure 1. Workflow for network inference from gene trees.

Protocol 2: Statistical Inference under the NMSC

This model-based approach directly infers networks from sequence alignments.

• Input Data: Multiple sequence alignments for hundreds of loci.
• Inference Problem: Co-estimate the network topology, branch lengths, and inheritance probabilities (γ) under the Network Multispecies Coalescent (NMSC) model, which accounts for both ILS and gene flow [85].
• Validation Mechanism: Models are compared using statistical model selection criteria, such as the Akaike Information Criterion (AIC) or Bayesian model comparison. A network that provides a significantly better fit to the data than a tree is supported. Bayesian approaches can also provide posterior probabilities for inferred reticulations.
• Key Parameter Interpretation: The inheritance probability (γ) assigned to a reticulation edge represents the proportion of genetic material the hybrid descendant inherited from that parent. A value near 0.5 suggests symmetrical hybridization, while values skewed toward 0 or 1 indicate asymmetrical introgression [85].

Performance Comparison of Network Inference Methods

The scalability and accuracy of network inference methods vary significantly. The following tables provide a comparative overview of available tools.

Table 1: Comparison of Parsimony-Based Network Inference Methods

Method / Program	Core Methodology	Maximum Scalability (Taxa / Trees)	Key Strength	Key Limitation
ALTS [52]	Aligning Lineage Taxon Strings	~50 taxa, ~50 trees	Fast enough for large problems without common clusters.	Limited to inferring tree-child networks.
HYBROSCALE [52]	Maximum Acyclic Agreement Forests	<30 taxa, <30 trees (without common clusters)	Provides a guaranteed minimal network.	Does not scale beyond ~30 taxa/trees without common clusters.
PRIN/PRINs [52]	Combining both editing and agreement forests	<30 taxa, <30 trees (without common clusters)	Infers a network with minimal hybridization number.	Poor scalability with increasing taxa/trees.
MCTS-CHN [52]	Maximum Agreement Forests	Two trees only	One of the fastest methods for the two-tree case.	Restricted to two input trees.
Hybridization Number [52]	Maximum Agreement Forests	Two trees only	Computes the exact hybridization number for two trees.	Restricted to two input trees.

Table 2: Comparison of Key Network Validation Metrics and Data Requirements

Metric	Definition	Interpretation in Validation	Ideal Value
Hybridization Number (HN)	Total number of extra lineages from reticulations [52].	Measures parsimony of the network; lower is better.	Minimized
Inheritance Probability (γ)	Proportion of genome from a parent [85].	Confirms hybrid origin; values between 0 and 1.	0.3 < γ < 0.7
Statistical Support (e.g., Bootstrap)	Proportion of re-sampled data supporting a reticulation.	Measures confidence in inferred reticulation.	> 90%
Number of Loci	Number of independent genomic regions used.	Power to distinguish ILS from reticulation.	Hundreds
Model Likelihood	Probability of data given the network model.	Goodness-of-fit for model selection.	Maximized

Successful network inference and validation require a suite of computational and data resources.

Table 3: Research Reagent Solutions for Phylogenomic Network Analysis

Item	Function / Purpose	Example / Note
Genome-Scale Data	Provides the hundreds of loci necessary to distinguish reticulate signals from noise.	Target Capture Sequencing, Whole-Genome Sequencing, or RNA-Seq data.
Gene Tree Inference Software	Infers individual trees from each locus, which serve as input for parsimony-based network methods.	IQ-TREE, RAxML, or BEAST.
Explicit Network Inference Software	Infers rooted phylogenetic networks from gene trees or sequence alignments.	ALTS (for tree-child networks), PhyloNet, or SNaQ.
Implicit Network Software	Visualizes conflict and discordance in the data for exploratory analysis.	SplitTree4.
High-Performance Computing (HPC)	Provides the computational power required for analyzing large datasets with complex models.	Local clusters or cloud computing services.

The power of phylogenomic data lies in its ability to provide the statistical force needed to validate complex evolutionary models. With hundreds of loci, researchers can now robustly infer phylogenetic networks, distinguishing true reticulate events from other sources of gene tree discordance like incomplete lineage sorting. As methods continue to scale and become more integrated into genomic analysis pipelines, phylogenetic networks are poised to become the standard for evolutionary inference in groups with known or suspected gene flow. This will profoundly impact fields like conservation biology, where understanding historical introgression is critical for managing species, and drug development, where the horizontal transfer of resistance genes is a major concern [85]. The future of phylogenomics is not just to tree, but to network.

The study of reticulate evolution through phylogenetic networks has revealed that evolutionary histories are often not purely tree-like but are shaped by complex processes such as hybridization, horizontal gene transfer, and introgression [34]. These reticulate events create evolutionary networks that challenge traditional conservation paradigms based solely on divergent evolution. Understanding these complex relationships is critical for biodiversity conservation, as phylogenetic networks provide insights into historical gene flow patterns, reveal previously unrecognized evolutionary relationships, and identify units of conservation priority that may be overlooked by tree-based models [86] [87].

The integration of phylogenetic networks into conservation planning represents a paradigm shift from static, lineage-based approaches to dynamic, interaction-based frameworks. This approach is particularly valuable for identifying evolutionary significant units, assessing conservation priorities in rapidly evolving systems, and predicting responses to environmental change [34]. As conservation initiatives increasingly target ambitious goals such as protecting 30% of lands and waters by 2030, the accurate representation of evolutionary relationships provided by phylogenetic networks becomes essential for effective priority-setting [88] [89]. This article compares methodologies for prioritizing conservation units, examining how insights from reticulate evolution research can enhance the design and implementation of protected area networks.

Comparative Analysis of Conservation Prioritization Methods

Methodological Approaches for Spatial Prioritization

Conservation prioritization methods vary in their underlying algorithms, data requirements, and suitability for different conservation contexts. The table below compares four widely-used approaches for identifying priority areas for biodiversity conservation:

Table 1: Comparison of Conservation Prioritization Methods

Method	Key Principle	Data Requirements	Advantages	Limitations
Species Richness [88]	Prioritizes areas with highest number of species	Species occurrence data	Simple to calculate and interpret; Minimal data requirements	Ignores species composition; Poor representation of beta diversity
Rarity-Weighted Richness [88]	Emphasizes areas with geographically restricted species	Species occurrence data with distribution ranges	Protects range-restricted and endemic species; Captures uniqueness	May overlook common species; Sensitive to scale and range definitions
Additive Benefit Function (ABF) [88]	Maximizes overall representation of biodiversity features	Species distribution data; Habitat maps	High efficiency in species representation; Complementary site selection	Computationally intensive; Requires specialized software
Core Area Zonation (CAZ) [88]	Prioritizes areas with highest conservation value while considering connectivity	Species distribution data; Habitat quality maps	Maintains ecological processes; Better connectivity representation	High computational demand; Complex parameterization

Complementarity-based methods like ABF and CAZ consistently outperform richness-based approaches, particularly for taxa with high beta diversity such as amphibians [88]. These methods achieve higher representation of species diversity within smaller geographic areas, making them particularly valuable when conservation resources are limited. For example, complementarity-based methods can protect the same number of species as richness-based approaches using 20-30% less land area, significantly advancing progress toward the "30 by 30" conservation targets [88].

Performance Metrics for Conservation Prioritization

The effectiveness of prioritization methods can be evaluated using multiple performance metrics, including species accumulation rates, habitat representation, and connectivity maintenance:

Table 2: Performance Comparison of Prioritization Methods for 30% Area Target

Performance Metric	Species Richness	Rarity-Weighted Richness	ABF	CAZ
Species Representation Efficiency	Low	Moderate	High	High
Beta Diversity Capture	Low	Moderate	High	High
Patch Size of Priority Areas	Small, fragmented	Variable, often small	Larger, more connected	Largest, well-connected
Conservation Opportunity	Low	Moderate	High	High
Connectivity Maintenance	Poor	Variable	Good	Best

Complementarity-based methods (ABF and CAZ) achieve significantly higher species representation per unit area, particularly for the top 30% of priority land [88]. These methods also produce larger, more connected patches of priority areas, which enhances their viability for long-term conservation and reduces the negative impacts of fragmentation [88] [89].

Experimental Protocols for Conservation Network Design

Integrating Connectivity Analysis with Biodiversity Prioritization

The Connectivity & Biodiversity Conservation (CBC) framework provides a robust methodology for designing conservation networks that integrate functional connectivity with biodiversity representation [89]. This protocol involves sequential analytical steps:

Step 1: Data Collection and Preparation

• Compile comprehensive protected area (PA) boundary data from governmental documents, scientific reports, and published materials [89]
• Exclude marine protected areas and regions with insufficient data
• For point-only PA data, create standardized buffer zones (e.g., 5 km) to represent PA boundaries
• Merge overlapping PA designations and remove duplicates
• Exclude patches smaller than 10 km² to ensure ecological relevance [89]

Step 2: Dispersal-Based Connectivity Modeling

• Adopt a coarse-filter approach using multiple dispersal distance thresholds (10 km, 30 km, 100 km) to accommodate diverse terrestrial species [89]
• Estimate species-specific movement abilities using allometric relationships based on body weight, diet, and ecological niche parameters [89]
• Calculate median, mean, and 90th percentile movement abilities across regions to validate dispersal thresholds

Step 3: Resistance Surface Development

• Model landscape resistance using human footprint datasets weighted by slope derived from digital elevation models [89]
• Account for combined impacts of human activities and topographic barriers on wildlife movement
• Generate cumulative resistance surfaces to identify least-cost paths between adjacent protected areas

Step 4: Conservation Priority Corridor Designation

• Identify Cost-Effective Connectivity Corridors (CCCs) using graph-based connectivity analysis in software such as Graphab 2.6 [89]
• Define "corridor importance" based on the number of overlapping CCCs within a given area
• Prioritize corridors using dual dimensions of ecological cost (resistance to movement) and economic cost (implementation resources) [89]

Phylogenetic Network Inference for Conservation Priorities

The inference of phylogenetic networks from genomic data follows a distinct protocol focused on evolutionary relationships:

Step 1: Data Collection and Sequencing

• Collect genomic sequence data across multiple taxa (e.g., ~8 Mb from 186 primate species representing 61 genera) [86]
• Include appropriate outgroup species for phylogenetic context
• Sequence multiple genetic markers, including coding regions (e.g., matK) and intergenic spacers (e.g., psbA-trnH, trnL-trnF) [86]

Step 2: Detection of Reticulate Evolution

• Identify discordance between gene trees from different genomic regions [86]
• Detect nucleotide additivity patterns in internal transcribed spacers (ITS) of nuclear ribosomal DNA [86]
• Apply statistical tests to distinguish hybridization from incomplete lineage sorting

Step 3: Network Construction and Reticulation Quantification

• Reconstruct phylogenetic networks using software such as SplitsTree4 [86]
• Calculate reticulation numbers based on connections at each node [87]
• Minimize reticulation events while accurately displaying gene trees [87]

Step 4: Integration with Conservation Planning

• Interpret phylogenetic networks to identify evolutionarily distinct populations and historically isolated lineages [34]
• Prioritize conservation units that represent unique evolutionary histories and potential adaptive variation
• Identify areas with historical connectivity to maintain potential for future gene flow [34]

Visualization Frameworks for Conservation Networks

Conservation Network Design Workflow

The following diagram illustrates the integrated workflow for designing conservation networks that incorporate both biodiversity prioritization and phylogenetic considerations:

Reticulate Evolution Impact on Conservation

The following diagram illustrates how reticulate evolutionary processes influence conservation prioritization:

Research Reagent Solutions for Conservation Network Analysis

Table 3: Essential Research Tools for Conservation Network Design

Tool/Category	Specific Solution	Function/Application
Connectivity Analysis Software	Graphab 2.6 [89]	Graph-based connectivity analysis; Identifies cost-effective corridors
Spatial Prioritization Platform	Zonation [88]	Complementarity-based prioritization using ABF and CAZ algorithms
Phylogenetic Network Software	SplitsTree4 [86]	Infers phylogenetic networks from sequence, distance, and tree data
Genomic Sequencing	Whole genome sequencing [86]	Generates data for phylogenetic analysis and reticulation detection
Landscape Resistance Data	Human Footprint Dataset [89]	Models resistance to wildlife movement across landscapes
Species Distribution Data	Priority-protected species databases [89]	Provides occurrence data for spatial prioritization algorithms

Discussion: Synthesizing Methodological Approaches for Effective Conservation

The integration of phylogenetic networks with spatial conservation prioritization represents a transformative approach to biodiversity protection. Phylogenetic networks enhance our understanding of evolutionary processes that generate and maintain biodiversity, while spatial prioritization methods ensure efficient allocation of conservation resources [34] [88]. This synthesis is particularly valuable for addressing the complex challenges of the Anthropocene, including habitat fragmentation, climate change, and ongoing reticulate evolution [34] [89].

Conservation strategies must balance multiple, sometimes competing, objectives including species representation, connectivity maintenance, and climate resilience. Complementarity-based prioritization methods (ABF and CAZ) combined with phylogenetic insights offer the most promising approach for achieving global conservation targets such as "30 by 30" [88] [89]. These methods efficiently represent biodiversity while accommodating complex evolutionary histories, providing a robust framework for safeguarding biodiversity in the face of rapid environmental change.

Future conservation efforts should leverage the complementary strengths of these approaches: using phylogenetic networks to identify evolutionarily significant units and historical connectivity patterns, and spatial prioritization algorithms to efficiently allocate protection across landscapes. This integrated framework will empower conservation practitioners to make informed decisions that protect not only current biodiversity patterns but also the evolutionary processes that generate and sustain biological diversity.

Conclusion

Phylogenetic networks have emerged from theoretical constructs into essential, scalable tools that empower researchers to test for and interpret reticulate evolution with unprecedented biological realism. Moving beyond the constraints of the tree-of-life paradigm is no longer a conceptual ideal but a practical necessity, as evidenced by advanced computational methods, robust statistical frameworks like the NMSC, and innovative applications of machine learning. For biomedical and clinical research, this shift promises more accurate tracing of pathogen origins, clearer understanding of the genetic underpinnings of complex diseases, and better insights into the evolutionary history of genes critical to drug discovery. Future directions will focus on enhancing the scalability of methods for massive genomic datasets, improving integration with population genomics, and developing standardized best practices for interpretation. Ultimately, embracing the 'web of life' through phylogenetic networks will provide a more nuanced and accurate foundation for evolutionary inquiry across the life sciences.

References

• 1. Towards a Research Programme Aiming at Causes and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12650632/]
• 2. Detection and utility in the phylogenomics era [https://www.sciencedirect.com/science/article/pii/S1055790324001891]
• 3. Reticulate evolutionary history and extensive introgression ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5146328/]
• 4. Leveraging evolution to identify novel organismal models ... [https://research.arcadiascience.com/pub/result-evolutionary-organismal-selection/release/3/]
• 5. Network‐based visualisation reveals new insights into ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8226279/]
• 6. Family Webs May Capture Evolutionary Change Better Than ... [https://news.ncsu.edu/2025/08/family-webs-may-capture-evolutionary-change-better-than-family-trees/]
• 7. IcyTree: rapid browser-based visualization for phylogenetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5860111/]
• 8. Visualization of Phylogenetic Networks • tanggle [https://klausvigo.github.io/tanggle/]
• 9. Prospects and limitations of genomic offset in conservation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8127717/]
• 10. Fast Digital Orthophoto Generation: A Comparative Study ... [https://www.mdpi.com/2072-4292/16/5/786]
• 11. Ten simple rules to create biological network figures for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6762067/]
• 12. Ten simple rules to colorize biological data visualization - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7561171/]
• 13. Phylogenetic Trees and Networks Reduce to Phylogenies ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4605253/]
• 14. Inference of reticulate evolutionary histories by maximum likelihood: the... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-13-S19-S12]
• 15. Distinguishing level-1 phylogenetic networks on the basis ... [https://link.springer.com/article/10.1007/s00285-021-01653-8]
• 16. Generation of Binary Tree-Child phylogenetic networks [https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007347]
• 17. “Normal” phylogenetic networks may be emerging as the ... [https://www.sciencedirect.com/science/article/pii/S0022519325002024]
• 18. About | Methodological Advances in Reticulate Evolution [https://gtiley.github.io/RBG-Networks/about/]
• 19. Inference of reticulate evolutionary histories by maximum ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3526433/]
• 20. The evolution of trait variance creates a tension between ... [https://www.nature.com/articles/s41467-022-30090-4]
• 21. Detecting Hidden Diversification Shifts in Models of Trait ... [https://pubmed.ncbi.nlm.nih.gov/27016728/]
• 22. Multispecies coalescent and its applications to infer ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8692950/]
• 23. Reconstructing patterns of reticulate evolution in plants [https://pmc.ncbi.nlm.nih.gov/articles/PMC2493047/]
• 24. Towards a Research Programme Aiming at Causes and ... [https://www.mdpi.com/2079-7737/14/11/1601]
• 25. Modeling Hybridization Under the Network Multispecies ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6101600/]
• 26. Multispecies coalescent process [https://en.wikipedia.org/wiki/Multispecies_coalescent_process]
• 27. Inferring the tree-like parts of a species network under the ... [https://www.imsi.institute/videos/inferring-the-tree-like-parts-of-a-species-network-under-the-coalescent/]
• 28. A multispecies coalescent model for quantitative traits [https://elifesciences.org/articles/36482]
• 29. A large-scale benchmark for network inference from single ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11897147/]
• 30. Benchmarking Gene Regulatory Network Inference Methods ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10197678/]
• 31. Efficient and scalable construction of clinical variable ... [https://www.sciencedirect.com/science/article/pii/S266723752500058X]
• 32. A comparison of computational methods for expression ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03840-y]
• 33. Benchmarking imputation methods for network inference ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-022-04778-9]
• 34. Phylogenetic networks empower biodiversity research [https://pubmed.ncbi.nlm.nih.gov/40720655/]
• 35. PhyloPapers 2025, Reticulate evolution - Brian O'Meara [https://brianomeara.info/posts/phylopapers_2025_Oct_10/]
• 36. PhyloTune: An efficient method to accelerate phylogenetic ... [https://www.nature.com/articles/s41467-025-61684-3]
• 37. Species-aware DNA language models capture regulatory ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-024-03221-x]
• 38. Genomic language models with k-mer tokenization strategies ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12313756/]
• 39. A survey of k-mer methods and applications in bioinformatics [https://www.sciencedirect.com/science/article/pii/S2001037024001703]
• 40. Plastid phylogenomics and cytonuclear discordance in ... [https://pubmed.ncbi.nlm.nih.gov/38768140/]
• 41. Phylogenomic cytonuclear discordance and evolutionary ... [https://link.springer.com/article/10.1007/s11427-023-2456-6]
• 42. Phylogenomic discordance is mainly driven by pervasive ... [https://pubmed.ncbi.nlm.nih.gov/40988340/]
• 43. Patterns of cytonuclear discordance and divergence between ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10654179/]
• 44. Cyto-nuclear discordance in the phylogeny of Ficus section ... [https://bmcecolevol.biomedcentral.com/articles/10.1186/1471-2148-9-248]
• 45. Plastid phylogenomics and cytonuclear discordance in ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0302365]
• 46. TreeViewer: Flexible, modular software to visualise and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10834882/]
• 47. List of phylogenetic tree visualization software [https://en.wikipedia.org/wiki/List_of_phylogenetic_tree_visualization_software]
• 48. The virus that won't quit: new research reveals how SARS ... [https://www.unsw.edu.au/newsroom/news/2025/06/the-virus-that-won-t-quit-new-research-reveals-how-sars-cov-2-evolves]
• 49. Hidden viruses in our DNA could be medicine's next big ... [https://www.sciencedaily.com/releases/2025/09/250902085154.htm]
• 50. Scientists Uncover How COVID-19 Variants Outsmart ... [https://www.mountsinai.org/about/newsroom/2025/scientists-uncover-how-covid-19-variants-outsmart-the-immune-system]
• 51. A scalability study of phylogenetic network inference methods ... [https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-1277-1]
• 52. A fast and scalable method for inferring phylogenetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10538497/]
• 53. Computational approaches to species phylogeny inference ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3855310/]
• 54. Incomplete lineage sorting [https://en.wikipedia.org/wiki/Incomplete_lineage_sorting]
• 55. Maximum likelihood inference of reticulate evolutionary histories [https://pmc.ncbi.nlm.nih.gov/articles/PMC4246314/]
• 56. Phytop: a tool for visualizing and recognizing signals of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11879507/]
• 57. Reticulate evolution and incomplete lineage sorting among ... [https://www.sciencedirect.com/science/article/abs/pii/S105579030900044X]
• 58. Incomplete lineage sorting and introgression among genera ... [https://bmcbiol.biomedcentral.com/articles/10.1186/s12915-025-02204-z]
• 59. Genomic data support reticulate evolution in whiptail ... [https://www.sciencedirect.com/science/article/pii/S1055790324002720]
• 60. Incomplete lineage sorting and phenotypic evolution in ... [https://www.sciencedirect.com/science/article/pii/S0092867422003440]
• 61. Robust variable selection methods with Cox model-a ... [https://pubmed.ncbi.nlm.nih.gov/39400113/]
• 62. Robust model selection in linear regression models using ... [https://www.sciencedirect.com/science/article/abs/pii/S0377042721003010]
• 63. Hybridization and Introgression – Molecular Ecology & ... [https://openpress.wheatoncollege.edu/molecularecologyv1/chapter/hybridization-introgression/]
• 64. A genomic view of introgression and hybrid speciation - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC2173880/]
• 65. Next-generation hybridization and introgression | Heredity [https://www.nature.com/articles/hdy201168]
• 66. Hybrid speciation driven by multilocus introgression of ... [https://www.nature.com/articles/s41586-024-07263-w]
• 67. Hybridization and introgression | Research Starters [https://www.ebsco.com/research-starters/health-and-medicine/hybridization-and-introgression]
• 68. Detecting hybrid speciation in the presence of incomplete ... [https://www.sciencedirect.com/science/article/abs/pii/S0040580908001111]
• 69. Hybrid Speciation and Introgression Both Underlie the ... [https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2020.576407/full]
• 70. Hybrid Speciation and Introgression Both Underlie ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7750405/]
• 71. Ancient hybridization underlies tuberization and radiation ... [https://www.sciencedirect.com/science/article/pii/S0092867425007366]
• 72. Exploring Decision Tree Machine Vision Systems in 2025 [https://www.unitxlabs.com/resources/decision-tree-machine-vision-system-2025-accuracy-speed/]
• 73. Decision tree-based machine learning algorithm for ... [https://www.sciencedirect.com/science/article/pii/S2405580825000780]
• 74. A Comparative Study of AdaBoost, Decision Trees, and ... [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5226457]
• 75. A Phylogenetic Networks perspective on reticulate human ... [https://www.sciencedirect.com/science/article/pii/S2589004221003278]
• 76. Ensemble of Precision-Recall Curve (PRC) Classification ... [https://arxiv.org/html/2509.05766v1]
• 77. Phylogenetic Networks And Evolutionary History [https://www.nature.com/research-intelligence/nri-topic-summaries/phylogenetic-networks-and-evolutionary-history-micro-25647]
• 78. The tree complexity of reticulate evolution | UC [https://www.canterbury.ac.nz/research/about-uc-research/research-specialities-and-projects/research-projects/the-tree-complexity-of-reticulate-evolution]
• 79. Aligning machine and human visual representations ... [https://www.nature.com/articles/s41586-025-09631-6]
• 80. Enhancing LLM-as-a-Judge Alignment with Human via ... [https://arxiv.org/html/2508.03550v3]
• 81. “Normal” phylogenetic networks may be emerging as the ... [https://arxiv.org/html/2107.10414v2]
• 82. Resolving Difficult Phylogenetic Questions: Why More ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3057953/]
• 83. To tree or not to tree: an introduction of phylogenetic networks [https://www.microbialsystems.cn/en/post/phylogenetic_network/]
• 84. Networks in phylogenetic analysis: new tools for population ... [https://www.sciencedirect.com/science/article/abs/pii/S0020751905000470]
• 85. Phylogenetic networks empower biodiversity research - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC12337313/]
• 86. | 678 Publications | 10547 Citations | Top Authors Reticulate evolution [https://scispace.com/topics/reticulate-evolution-9q225vyp]
• 87. Unraveling the Complexity of Evolutionary - Simple Science Networks [https://scisimple.com/en/articles/2025-06-22-unraveling-the-complexity-of-evolutionary-networks--a3oww4w]
• 88. Options for prioritizing sites for biodiversity conservation ... [https://www.sciencedirect.com/science/article/pii/S0006320721004304]
• 89. Conservation priority corridors enhance the effectiveness ... [https://www.nature.com/articles/s43247-025-02227-y]