Disentangling folding stability and function in deep mutational scanning fitness

Abstract

This repository accompanies a computational analysis decomposing protein variant fitness into folding-stability and function components across 151 deep mutational scanning (DMS) assays from ProteinGym. For each assay, the agreement of every predictor with measured fitness was quantified, structure- and sequence-based stability predictors (RaSP and DDGun) were benchmarked against a graded ladder of trivial baselines (BLOSUM62, hydrophobicity change, side-chain volume change, and relative solvent accessibility) and against five function-aware models (ESM2, ESM-1v, ESM-C, ESM3, and EVE), and partial-correlation and cross-validated variance-partition analyses were used to separate signal shared with, versus independent of, folding stability. A clinical arm evaluated pathogenic-versus-benign separation on expert-annotated ClinVar variants in disease genes, using predictors that had not been trained on clinical labels. Contents: all analysis scripts (numbered in execution order), derived per-assay result tables from which every figure and reported value is computed, the six manuscript figures, and the software-environment specifications for both the data-preparation and analysis pipelines. Raw input data (ProteinGym assays and zero-shot scores, the Tsuboyama et al. 2023 mega-scale stability dataset, and the ClinVar variant summary) are not included; they are available from their original sources, with MD5 checksums listed in the manuscript Methods, and are reproduced by the deposited scripts. Computation was performed on the PARAM Shakti supercomputing facility at the Indian Institute of Technology Madras.