Accurate prediction of peptide hydrogel formation remains challenging. This study presents an interactive approach combining experiment and machine learning to predict and design tetrapeptide hydrogels. Over 160 tetrapeptides were synthesized and tested, iteratively improving prediction accuracy through machine learning-experiment loops. A score function incorporating aggregation propensity, hydrophobicity, and a gelation corrector achieved an 87.1% success rate in predicting hydrogel formation within an 8,000-sequence library. A de novo-designed peptide hydrogel enhanced the immune response to SARS-CoV-2's receptor-binding domain in mice. This approach significantly expands the scope of natural peptide hydrogels.

Tengyan Xu, Jiaqi Wang, Shuang Zhao, Dinghao Chen, Hongyue Zhang, Yu Fang, Nan Kong, Ziao Zhou, Wenbin Li, Huaimin Wang