PhyDL

PhyDL reconstructs phylogenetic topologies from protein sequence data using deep residual neural networks (ResNets) to infer evolutionary relationships while addressing amino acid substitution heterogeneities across sites and lineages.

Key Features:

• Deep residual neural networks (ResNets): Uses ResNets to model complex non-linear relationships in sequence evolution without requiring explicit substitution models.
• Training on simulated data: Trained on simulated protein sequence data characterized by extensive amino acid substitution heterogeneities.
• Mitigation of model misspecification: Designed to reduce errors arising from substitution model misspecification and insufficiency.
• Phylogenetic topology inference: Infers tree topologies directly from protein sequence alignments.
• Comparative performance: Demonstrates superior performance to maximum likelihood methods in simulation scenarios with significant substitution heterogeneities.
• Empirical concordance: Produces predictions that generally align with established topologies on real phylogenetic datasets.

Scientific Applications:

• Phylogenetic reconstruction under substitution heterogeneity: Reconstructs evolutionary relationships for datasets with heterogeneous amino acid substitution processes.
• Benchmarking and method comparison: Provides a machine-learning-based comparator to maximum likelihood phylogenetic inference in simulation and empirical studies.
• Protein sequence evolution analysis: Applicable to studies of molecular evolution using protein sequence alignments exhibiting site- and lineage-specific heterogeneities.

Methodology:

Employs deep residual neural networks (ResNets) trained on simulated protein sequence datasets with extensive amino acid substitution heterogeneities, with performance evaluated against maximum likelihood methods and on empirical phylogenetic datasets.