Dynamics-aware geometric learning predicts disease-associated molecular perturbations

Abstract

Missense mutations and post-translational modifications (PTMs) are major molecular perturbations that reshape protein function but are traditionally studied independently. Current computational approaches largely rely on sequence conservation or static structural features, limiting our understanding of how perturbations alter intrinsic protein dynamics. We present DynGeo-Pheno, a unified geometric deep learning framework that integrates protein language model representations with anisotropic network model-derived dynamics to jointly capture evolutionary, structural, and biophysical information. DynGeo-Pheno predicts disease-associated phosphosites and pathogenic missense mutations with high accuracy on independent test datasets. Ablation analyses indicate that protein dynamics provide complementary information beyond sequence evolution and structural topology for pathogenicity prediction. Beyond predictive performance, DynGeo-Pheno reveals that disease-associated perturbations preferentially localize to functional structural regions, including ligand-binding pockets and PPI interfaces. Mechanistically, phosphosites and missense mutations appear to exhibit distinct yet convergent dynamic signatures. Phosphosites preferentially occur in flexible regulatory regions, whereas pathogenic mutations are enriched in ordered structural elements. Despite these differences, both perturbation types display enhanced long-range coupling, increased perturbation responsiveness, and elevated mechanical stability, indicating that pathogenic residues preferentially occupy mechanically constrained and allosteric regulatory sites. This study provides compelling evidence that intrinsic protein dynamics is an important complementary determinant of pathogenicity and establishes a unified framework for interpretable AI predictions and mechanistic understanding of how genetic and regulatory perturbations may shape protein function.