Self-distillation and antisymmetric constraints enable accurate prediction of protein stability changes from single mutations

Accurately predicting protein stability changes upon residue-level mutations is essential for protein engineering and design. However, previous computational methods are typically designed for specific tasks and struggle to generalize across different stability systems, preventing the accurate guidance of protein mutagenesis. Sequence-based methods are easier to apply to large-scale protein datasets since they do not rely on high-quality structures. However, existing sequence-based approaches struggle to capture structural changes, resulting in lower performance compared to structure-based methods. In this study, we propose DeepTherm, a sequence-based deep learning solution to model changes in folding free energy (∆∆G) and melting temperature (∆T_m) resulting from single-point mutations. DeepTherm transfers a protein large language model as a core component and incorporates a cross-attention mechanism to capture the contact-map changes around mutated positions. To address data imbalance and the antisymmetric nature of mutation effects, DeepTherm employs a self-distillation inference strategy under the supervision of an antisymmetric constraint. Benchmarking across multiple datasets shows that DeepTherm achieves state-of-the-art accuracy in both ∆∆G and ∆T_m predictions. Furthermore, practical evaluations confirm DeepTherm's capability in accurately ranking protein stability on large-scale datasets and effectively identifying critical structural contacts impacting stability. Detailed analysis on cDNA display proteolysis data further highlights the critical importance of self-distillation and antisymmetric constraints in enhancing predictive performance.