Leveraging Unified Sequence-Structure Representations for Enhanced Protein Stability Prediction

Abstract Protein thermal stability, quantified by the change in Gibbs free energy (ΔΔG) upon mutation, is critical for drug design and enzyme engineering. Current multi-modal deep learning models, despite integrating sequence, often struggle with indirect information fusion and incomplete capture of sequence-structure interactions. We introduce ProStab-Former, addressing these limitations by establishing a unified sequence-structure representation space for protein stability prediction. It leverages a frozen, multi-modal protein foundation encoder for residue-level feature extraction. Fine-tuned modules include Stability-Aware Attention Layers (SAAL) with structural prior bias and mutation-aware gating, and an Epistatic Interaction Module for multi-point mutation prediction. The model achieves superior or competitive performance, surpassing a state-of-the-art baseline. Ablation confirms SAAL’s critical role; strong generalization is shown across tasks including melting temperature prediction and pathogenic mutation classification. Its exceptional efficiency, predicting numerous single-point mutations in a single pass, positions it as a practical tool for high-throughput protein engineering and variant effect analysis.