Few-shot active learning for de novo dual-target peptide design with high bio-activity

Despite the urgent need for high bio-activity peptides in novel biomedical therapies, the de novo design of such peptides, especially those with dual targets, remains an unsolved challenge. Here, we introduce ORIDTP, a few-shot active learning pipeline that integrates in silico peptide generation with in vitro experimental feedback for de novo design of both single-target and dual-target peptides with high bio-activity. ORIDTP involves single-target or dual-target oriented peptide de novo generation, binding affinity maturation, and iterative reinforcement of bio-activity based on wet-laboratory feedback. Using ORIDTP, we successfully designed high bio-activity peptides targeting GLP-1R after four iterative rounds, achieving $\rm EC_{50}$ (half maximal effective concentration) values ranging from 35.1 pM to 8.1 pM, which outperform the natural peptide with the highest known bio-activity of 40.8 pM. Furthermore, ORIDTP successfully designed de novo dual-target peptides for activating GLP-1R ($\rm EC_{50}$ values ranging from 53.4 pM to 8.2 pM) and GCGR ($\rm EC_{50}$ values ranging from 0.82 nM to 0.21 nM) after four iterative rounds. The best dual-target peptide outperformed two natural peptides with the highest known bio-activity for their respective target proteins (8.2 pM versus 40.8 pM for GLP-1R, and 0.24 nM versus 1.4 nM for GCGR). ORIDTP represents a significant advancement in the rapid and effective design of dual-target peptides for therapeutic applications.