Learning Constraints in Protein Evolution beyond Known Substitution Frequencies
Abstract
During the evolution of protein sequences, amino acid substitutions follow broad patterns that are well established, such as those induced by the physico-chemical properties of residues or the position of sites along the sequence. While these deviations from random changes have been exploited for practical applications, such as to simulate protein evolution or to score protein alignments, they are too general to predict the precise evolution of individual protein sequences. To investigate whether protein sequences contain contextual information about their subsequent evolution beyond that captured by conventional substitution statistics and evolutionary conservation, we trained deep learning models based on an Evolutionary Scale Modelling (ESM2) encoder to predict the rate, the position and the nature of the amino-acid substitutions in the evolution of ancestral sequences and compared their performance to simpler and explainable heuristics. Our results show that predicting the rate of amino acid changes is achieved equally well by deep learning and by using evolutionary information in outgroup sequences. This suggests that the protein sequence does not contain information that significantly improves the predictions for this task. On the opposite, we show that deep learning models predict the position of these substitutions more accurately than statistics-based baselines. Predictions of the nature of substitutions largely outperform baselines, with over 37% accuracy overall. We observe that rapidly evolving sequences are harder to predict and substitutions preferentially occur between chemically similar amino acids. Interestingly, predictive information is predominantly contained within a local sequence context of 16 residues. These findings indicate that the nature of substitutions in protein evolution is partly governed by predictive rules that remain to be discovered through a more accurate modelling that could describe evolutionary trajectories based on sequence context alone.
Related
- Protein Language Models: Learning From Evolution, Designing Beyond It
- Cytogenetic constraints on hybridization: a meta-analysis investigating the role of chromosome number in monocot hybrid evolution using a newly developed tool, the ploidy deviation index (PDI)
- Limit-pushing overexpression reveals constraints on protein abundance
- Environment-mediated metabolic constraints shape antimicrobial resistance evolution in Pseudomonas aeruginosa
- Cldn18 evolution reveals pleiotropic constraints in vertebrate stomach loss
BibTeX
@article{bricout2026learning,
title = {{Learning Constraints in Protein Evolution beyond Known Substitution Frequencies}},
author = {Raphaël Bricout and Hugues Roest Crollius and Auguste Genovesio},
journal = {Molecular Biology and Evolution},
year = {2026},
month = nov,
doi = {10.1093/molbev/msag270},
url = {https://hal.science/hal-05855588v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - Learning Constraints in Protein Evolution beyond Known Substitution Frequencies
AU - Raphaël Bricout
AU - Hugues Roest Crollius
AU - Auguste Genovesio
PY - 2026
DA - 2026/11/04
JO - Molecular Biology and Evolution
DO - 10.1093/molbev/msag270
UR - https://hal.science/hal-05855588v1
AB - During the evolution of protein sequences, amino acid substitutions follow broad patterns that are well established, such as those induced by the physico-chemical properties of residues or the position of sites along the sequence. While these deviations from random changes have been exploited for practical applications, such as to simulate protein evolution or to score protein alignments, they are too general to predict the precise evolution of individual protein sequences. To investigate whether protein sequences contain contextual information about their subsequent evolution beyond that captured by conventional substitution statistics and evolutionary conservation, we trained deep learning models based on an Evolutionary Scale Modelling (ESM2) encoder to predict the rate, the position and the nature of the amino-acid substitutions in the evolution of ancestral sequences and compared their performance to simpler and explainable heuristics. Our results show that predicting the rate of amino acid changes is achieved equally well by deep learning and by using evolutionary information in outgroup sequences. This suggests that the protein sequence does not contain information that significantly improves the predictions for this task. On the opposite, we show that deep learning models predict the position of these substitutions more accurately than statistics-based baselines. Predictions of the nature of substitutions largely outperform baselines, with over 37% accuracy overall. We observe that rapidly evolving sequences are harder to predict and substitutions preferentially occur between chemically similar amino acids. Interestingly, predictive information is predominantly contained within a local sequence context of 16 residues. These findings indicate that the nature of substitutions in protein evolution is partly governed by predictive rules that remain to be discovered through a more accurate modelling that could describe evolutionary trajectories based on sequence context alone.
SN - 0737-4038
SN - 1537-1719
ER -