ERIS2: A deep-learning platform for predicting mutation-induced protein stability changes.

Abstract

Accurately predicting changes in protein thermodynamic stability upon mutation is a fundamental challenge in protein engineering and therapeutics. Deep learning has improved prediction, but many models remain limited by small, imbalanced training data and poor generalization across mutation types. Here, we present ERIS2, an attention-convolutional deep learning framework that integrates contextual sequence embeddings with atomistic and residue-level structural features from experimental protein structures to predict mutation-induced stability changes. Trained on large-scale experimental measurements combined with curated ProTherm data, a 5-fold ERIS2 ensemble achieved Pearson correlation coefficients of 0.69 on S461, 0.51 on S669, and 0.60 on S96, three blind-test datasets sequence-disjoint from all training configurations, and showed negligible anti-symmetry bias on Ssym forward and reverse pairs. ERIS2 is freely available as a structure-based web server without registration, providing a robust framework for protein stability prediction and advancing computational tools for protein engineering and therapeutic design.