Language-guided expertise evolution for protein optimization

This project studies how frozen language models can be adapted to protein optimization by improving explicit domain expertise rather than model parameters. The system maintains an evolving pool of executable expertise blocks, lets different blocks specialize to different parts of the protein sequence space, and uses language feedback as a dense signal for refinement.

The early version of this work won the Best Creativity Award in the 玻尔+SciMaster AI4S 智能体开发竞赛, reached the interview round of MiraclePlus (top 10%), and led to the workshop paper Language-guided expertise evolution for protein optimization.