AI-guided discovery for low-resource peptide engineering using evolutionary scale modeling

Data scarcity is a major bottleneck in artificial intelligence (AI)-guided peptide engineering, where experimental training sets often contain only tens to a few hundred sequences, and it is difficult to judge in advance whether a protein language model (PLM) will effectively guide iterative discovery before committing to wet-lab work. Here we present SCARSE, a machine learning framework that couples ESM2 protein language model embeddings with Gaussian process regression. Benchmarked across 24 diverse peptide and small-protein datasets spanning substitution and indel variants, antimicrobial peptides, and hemolytic peptides, SCARSE significantly outperformed hand-engineered descriptor baselines on substitution and indel tasks in low-data regimes (20-500 training samples) and performed comparably on short non-mutant peptides. In active learning simulations it consistently prioritized high-performing candidates, discovering up to 7-fold more top-tier peptides than random sampling. Cross-validation (CV) R2 correlated strongly with final active learning performance (Pearson r = 0.80-0.84), and evaluating on as few as 50 labeled peptides reduced CV R2 variance enough for it to serve as a practical, data-efficient criterion for assessing model suitability before iterative screening. SCARSE thus provides an early diagnostic indicator for active learning success and is freely available as an open-source Python package (pip) and through an interactive Hugging Face Spaces web interface.