From sequences and structures to fitness, folding and function: challenges in the age of AI
- 1Centre for Genomic Regulation (CRG), The Barcelona Institute of Science and Technology, Dr. Aiguader 88, Barcelona, Spain, 08003
- 2University Pompeu Fabra (UPF)
Abstract
Deep-learning models have transformed our ability to predict the phenotypic effects of sequence perturbations. Protein language models (pLMs) score the evolutionary propensity of any amino-acid substitution, emerging as fast and accurate predictors of fitness changes. Inverse-folding (IF) models add structural information and are powerful alternatives for estimating folding-stability changes. But these likelihoods are imperfect and sometimes biased predictors of the biological quantities they are used to measure. Scaling pLMs has shown that bigger models do not always outperform smaller ones. IF models sometimes read sequence conservation as a stability constraint, which leads to biases when this signal is used to isolate functional contributions beyond folding. We review here the strategies emerging to mitigate, and where possible overcome, these limitations. Highlights:- Scaling protein language models does not close the likelihood–fitness gap - Inverse-folding models can read functional conservation as lost stability - Disentangling function from stability inherits both biases - Emerging strategies are attempting to address these issues
Related
- Structures of protein folding intermediates on the ribosome
- Predicting protein structures at any folding stage from sequence
- CLEAN Database: providing enzyme function coverage for over 55 million sequences
- Sequence design for three-dimensional genome folding using Akita Semifreddo
- An inverse-protein-folding pre-trained model for predicting the mutational effects in protein-protein interactions
BibTeX
@misc{galpern2026from,
title = {{From sequences and structures to fitness, folding and function: challenges in the age of AI}},
author = {Ezequiel A. Galpern and Charles W. J. Pugh and Mafalda Dias and Jonathan Frazer},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009409/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009409/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - From sequences and structures to fitness, folding and function: challenges in the age of AI
AU - Ezequiel A. Galpern
AU - Charles W. J. Pugh
AU - Mafalda Dias
AU - Jonathan Frazer
PY - 2026
DA - 2026/09/24
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009409/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009409/v1
AB - Deep-learning models have transformed our ability to predict the phenotypic effects of sequence perturbations. Protein language models (pLMs) score the evolutionary propensity of any amino-acid substitution, emerging as fast and accurate predictors of fitness changes. Inverse-folding (IF) models add structural information and are powerful alternatives for estimating folding-stability changes. But these likelihoods are imperfect and sometimes biased predictors of the biological quantities they are used to measure. Scaling pLMs has shown that bigger models do not always outperform smaller ones. IF models sometimes read sequence conservation as a stability constraint, which leads to biases when this signal is used to isolate functional contributions beyond folding. We review here the strategies emerging to mitigate, and where possible overcome, these limitations. Highlights:- Scaling protein language models does not close the likelihood–fitness gap - Inverse-folding models can read functional conservation as lost stability - Disentangling function from stability inherits both biases - Emerging strategies are attempting to address these issues
ER -