TL;DR
This paper investigates how common simplifications in phylogenetic inference models (model misspecification) can lead to inaccurate conclusions about evolutionary history, particularly regarding gene flow. The authors demonstrate that accounting for gene tree estimation error and variation in evolutionary rates across the genome significantly improves the reliability of inferring complex evolutionary histories, like those involving reticulations (gene flow).
What Problem Does This Paper Address?
Inferring the evolutionary history of species, especially when gene flow (reticulations) occurs between them, is a complex challenge. Traditional phylogenetic methods often assume a simple tree-like history, ignoring gene flow. While newer methods can model these reticulations (species networks), they are susceptible to "model misspecification." This means that if the model used for inference doesn't perfectly match the true evolutionary process, the results can be misleading. Two key sources of misspecification are:
- Gene Tree Estimation Error (GTEE): When inferring gene trees from DNA sequences, errors can arise. If these errors are treated as genuine evolutionary signals, they can be misinterpreted as evidence of reticulations.
- Substitution Rate Heterogeneity: The rate at which DNA sequences evolve can vary across different parts of the genome (loci). Assuming a single, constant rate for all loci when this isn't true can also lead to incorrect inferences.
This paper aims to quantify the impact of these two misspecification issues on phylogenetic network inference and to propose solutions.
Methods
The authors employed a multi-pronged approach:
- Simulations: They generated synthetic genomic data under various scenarios, manipulating the level of incomplete lineage sorting (ILS) and gene tree estimation error (GTEE). This allowed them to control the "ground truth" and assess how well different methods performed.
- Statistical Tests for Reticulation: They evaluated existing statistical tests designed to distinguish between tree-like and network-like evolutionary histories. A key innovation here was developing a "triplet multitest" approach, which applies tests to smaller three-taxon subsets of the data and then corrects for multiple testing. This aims to improve robustness against GTEE.
- Phylogenetic Network Inference Methods: They tested two main types of methods:
- Full Bayesian Methods (e.g., MCMC_SEQ): These methods jointly infer the species phylogeny and gene trees, integrating over uncertainty. They explored how assuming a single substitution rate versus allowing for rate variation (using a Dirichlet distribution, implemented via a Delta Exchange Operator - DEO) affected accuracy.
- Summary Statistic Methods (e.g., InferNetwork_ML): These methods use pre-estimated gene trees as input. They assessed the robustness of these methods to GTEE and rate heterogeneity.
- Quantitative Accuracy Measures: They used metrics like Normalized Robinson-Foulds (nrRF) distance for topology and Normalized Branch Score (nrBS) for topology and branch lengths to quantify the accuracy of inferred gene trees and species networks.
Datasets
The study primarily used simulated datasets. They also reanalyzed two empirical datasets:
- Anopheles mosquito genomes: Used to test statistical approaches for reticulate evolution.
- Heliconius butterfly genomes: Used to investigate the impact of rate heterogeneity and to re-evaluate previous findings of introgression.
Key Insights and Ideas
- GTEE is a major problem for reticulation tests: Existing statistical tests designed to detect gene flow are highly susceptible to errors in gene tree estimation, leading to many false positives (incorrectly inferring reticulations when the history is actually a tree).
- Triplets save the day (for tests): Applying statistical tests to smaller three-taxon subsets and correcting for multiple testing significantly mitigates the false positive problem caused by GTEE. This "divide-and-conquer" strategy is more robust.
- Rate heterogeneity fools full Bayesian methods: Assuming a single substitution rate across all genes when rates actually vary leads full Bayesian methods to incorrectly infer spurious reticulations.
- Accounting for rate variation is crucial: Implementing models that allow for per-locus rate variation (like the Dirichlet rates model) dramatically improves the accuracy of full Bayesian phylogenetic network inference when such variation is present.
- Summary methods are more robust: Methods that use pre-estimated gene trees (summary methods) are less affected by GTEE and rate heterogeneity, though they might still require manual inspection to determine the correct network complexity.
Key Results
- Statistical Tests: The triplet-based approach with multiple testing correction significantly reduced false positive rates for detecting reticulations, especially under high GTEE. For instance, when the true history was a tree, the false positive rate was kept around 5% with the triplet method, whereas other methods showed much higher rates.
- Full Bayesian Inference: When data were simulated with rate heterogeneity, inferring with a single-rate model led to a significant decrease in accuracy compared to using a Dirichlet rates model. For example, in one scenario, accuracy dropped substantially when the SR model was used for DR data. However, using the DR model for SR data had only a minor impact on accuracy.
- Summary Methods: These methods showed good robustness to GTEE and rate heterogeneity, with only minor decreases in accuracy. However, they did not automatically determine the number of reticulations as effectively as the full Bayesian methods when rate heterogeneity was present.
- Empirical Data: Reanalysis of Anopheles data supported the presence of introgression, while some previously reported introgression in Heliconius species was not supported, potentially due to model misspecification in earlier analyses.
Limitations and Future Work
- Triplet Test Limitations: The proposed triplet-based statistical test is limited to testing against a tree null hypothesis and has identifiability issues for three-taxon networks (trinets). Future work could explore methods to address this.
- Computational Cost: Full Bayesian methods, especially with the inclusion of rate heterogeneity, can be computationally demanding, leading to longer convergence times. The paper notes that ESS accumulation was significantly slower with the DR model.
- Summary Method Complexity: While robust, summary methods may still require manual inspection to determine the correct number of reticulations, especially when dealing with complex evolutionary histories.
- Future Directions: The authors recommend using models that account for rate heterogeneity whenever possible for phylogenetic network inference. They have also implemented their DR model in MCMC_SEQ to facilitate this. Further research could explore more sophisticated methods for determining network complexity in summary methods.
Summary updated