Harnessing machine learning to boost heuristic strategies for phylogenetic-tree search

Abstract

Inferring a phylogenetic tree, which describes the evolutionary relationships among a set of organisms, genes, or genomes, is a fundamental step in numerous evolutionary studies. With the aim of making tree inference feasible for problems involving more than a handful of sequences, current algorithms for phylogenetic tree reconstruction utilize various heuristic approaches. Such approaches rely on performing costly likelihood optimizations, and thus evaluate only a subset of all potential trees. Consequently, all existing methods suffer from the known tradeoff between accuracy and running time. Here, we train a machine-learning algorithm over an extensive cohort of empirical data to predict the neighboring trees that increase the likelihood, without actually computing their likelihood. This provides means to safely discard a large set of the search space, thus avoiding numerous expensive likelihood computations. Our analyses suggest that machine-learning approaches can make heuristic tree searches substantially faster without losing accuracy and thus could be incorporated for narrowing down the examined neighboring trees of each intermediate tree in any tree search methodology.

Dissecting Factors Underlying Phylogenetic Uncertainty Using Machine Learning Models

“…This is because the features learned from the model may not correlate with existing literature that addresses factors influencing phylogenetic uncertainty. In contrast, our study's Feature Importance analyses demonstrate clear connections with previous studies, reinforcing the notion that welldefined features can inform decision-making in heuristic search moves within the tree space (Azouri et al, 2021(Azouri et al, , 2023 and predict the difficulty of a dataset for phylogenetic reconstruction (Haag et al, 2022). However, our study also suggests, particularly after analyzing the Carangaria dataset, that current machine learning models in phylogenetics that use empirical datasets should be trained with data that is agnostic to phylogenetic model assumptions such as homogeneity, stationarity, and reversibility (Pupko and Mayrose, 2020).…”

“…The obtained insights can greatly enhance decision-making in phylogenetic analyses, aiding in the selection of appropriate DNA sequence models and data transformation methods. While machine learning has demonstrated its utility in various aspects of phylogenetic analysis, such as biogeography models (Smith et al, 2017;Fonseca et al, 2021), phylodynamic parameter estimation (Voznica et al, 2022), ancestral state reconstruction (Theobald et al, 2022), and improving heuristics for phylogenetic inference (Azouri et al, 2021(Azouri et al, , 2023, the quality of these models ultimately depends on the quality of the training data, which can inherently carry its own uncertainty, as we showed in this study. Therefore, the pursuit of identifying factors that influence the uncertainty in phylogenomic datasets and developing models that effectively capture non-linear relationships among features in phylogenetic inference is essential for advancing our understanding of the tree of life.…”

“…This presents an unprecedented opportunity to train machine learning models that can potentially learn complex non-linear relationships within the phylogenomic dataset without imposing a fixed set of probability distribution families to construct the model (i.e., non-parametric models) (Hastie et al 2009;Murphy 2012;Bokulich et al 2018). Machine learning models, capable of identifying complex non-linear relationships, have recently found applications in various phylogenetic analyses, including fast phylogenetic inference (Azouri et al 2021(Azouri et al , 2023, phylogenetic placement (Jiang, Tabaghi, and Mirarab 2022), learning phylodynamic parameters (Voznica 2021), and selecting biogeographical models from simulated data (Smith et al 2017;Burbrink and Gehara 2018). While most of these studies focus on prediction tasks, exploring the interactions of features in specific phylogenetic datasets leveraging these non-linear models remains largely unexplored.…”

Harnessing machine learning to guide phylogenetic-tree search algorithms

“…The code that supports the findings of this study was written in Python version 3.6 and has been deposited in Open Source Framework (OSF) with the identifier DOI 10.17605/ OSF.IO/B8AQJ 51 . Computation of likelihoods and parameter estimates were executed using the following application versions: PhyML 3.0 31 , RAxML-NG 0.9.0 48 .…”

“…The datasets contained within the empirical set have been deposited in Open Source Framework (OSF) with the identifier DOI 10.17605/OSF.IO/B8AQJ 51 . These datasets were assembled from the following databases: TreeBase (https://treebase.org/treebaseweb/urlAPI.html); Selectome (https://selectome.org/); protDB (https://protdb.org/); PloiDB (https://doi.org/10.3732/ajb.1500424); PANDIT (https://www.ebi.ac.uk/research/ goldman/software/pandit); OrthoMaM (https://orthomam.mbb.cnrs.fr/).…”