Interpretable Structural and Pilot Evolutionary Diagnostics for Channel-Specific Mutation Effects in Protein Folding

This repository contains the data, analysis outputs, and supporting materials for the study “Interpretable Structural and Evolutionary Diagnostics for Channel-Specific Mutation Effects in Protein Folding.” The work develops an interpretable, channel-resolved framework for analyzing mutation effects, separating them into three mechanistically distinct components: folding slowdown, native-state destabilization, and unfolding acceleration. Rather than collapsing these outcomes into a single damage metric, the study demonstrates that they are empirically separable and exhibit different dependencies on structural and mutation-local features. The empirical backbone is constructed from the K-Pro kinetics dataset combined with a completed wild-type structure mapping pass, enabling near-complete residue-level localization across mutation entries. The modeling framework compares weak-local mutation descriptors, structure-only descriptors, and their combination across both regression and classification tasks under grouped validation, including leave-one-protein-out transfer. Results show that structure-aware features most strongly organize destabilization and unfolding behavior, while slowdown remains more dependent on local mutation context. Severe-event classification further supports this asymmetry, with the strongest performance observed for destabilization. To extend beyond structure alone, a pilot evolutionary layer is included using a multiple sequence alignment of alpha-globin homologs. This analysis reveals a strongly conserved structural core alongside a smaller set of flexible, higher-entropy positions, supporting a decomposition in which global stability is associated with conserved residues while local variability corresponds to sites of tolerance and fragility. This provides a concrete direction for future integration of evolutionary constraints into the predictive framework. The repository includes processed mutation datasets, structure-mapped descriptors, model evaluation outputs across multiple validation regimes, surrogate-collapse validation results, evolutionary alignment data, entropy profiles, and all figures used in the manuscript. The goal of this work is to provide a compact, interpretable, and empirically grounded diagnostic framework that can be extended with richer evolutionary and thermodynamic layers in future studies.