Semantical and geometrical protein encoding toward enhanced bioactivity and thermostability

• DOI: 10.7554/elife.98033.4
• Publication Date: 2025-05-02T11:43:04Z
• Authors: Yang Tan; Bingxin Zhou; Lirong Zheng; Guisheng Fan; Liang Hong
• Keywords: Thermostability

Protein engineering is a pivotal aspect of synthetic biology, involving the modification amino acids within existing protein sequences to achieve novel or enhanced functionalities and physical properties. Accurate prediction variant effects requires thorough understanding sequence, structure, function. Deep learning methods have demonstrated remarkable performance in guiding for improved functionality. However, approaches predominantly rely on sequences, which face challenges efficiently encoding geometric aspects acids’ local environment often fall short capturing crucial details related folding stability, internal molecular interactions, bio-functions. Furthermore, there lacks fundamental evaluation developed predicting thermostability, although it key property that frequently investigated practice. To address these challenges, this article introduces pre-training framework integrates sequential encoders primary tertiary structures. This guides mutation directions toward desired traits by simulating natural selection wild-type proteins evaluates based their fitness perform specific functions. We assess proposed approach using three benchmarks comprising over 300 deep mutational scanning assays. The results showcase exceptional across extensive experiments compared other zero-shot methods, all while maintaining minimal cost terms trainable parameters. study not only proposes an effective more accurate comprehensive predictions facilitate efficient engineering, but also enhances silico assessment system future models better align with empirical requirements. PyTorch implementation available at https://github.com/ai4protein/ProtSSN .