DeepTherm is an end-to-end Deep-learning method for prediction of Protein Thermostability changes (\(\Delta\Delta G\) and \(\Delta T_m\)) induced by single residue mutations. Given a protein sequence and the mutation information, DeepTherm first deduces the evolutionary information and predicts corresponding contact maps for wild-type and mutated protein sequences with a protein large language model. Based on these two types of information, DeepTherm then introduces a neighboring encoder to capture the local structural changes around the mutation site. Finally, the difference of local structure between wild-type and mutated proteins is concatenated with environmental factors (pH and temperature) to predict the stability changes (\(\Delta\Delta G\)) and melting temperature changes (\(\Delta T_m\)). Notably, DeepTherm employs a self-distillation inference strategy under the supervision of an antisymmetric constraint to address data imbalance and the antisymmetric nature of mutation effects. Benchmarking across multiple large-scale datasets shows that DeepTherm achieves state-of-the-art accuracy in both \(\Delta\Delta G\) and \(\Delta T_m\) predictions. For more details about the method and online server, please check the [Help] page.
References
- Wenkang Wang, Yihang Zhou, Xiaoqiang Huang, Yifan Wu, Fanding Xu, Ruiqing Zheng, Min Li, Yang Zhang. Accurate prediction of protein stability changes from single mutations using self-distillation and antisymmetric constraint strategies. Submitted (2025).