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

SCARSE, an AI framework combining ESM2 embeddings with Gaussian process regression, outperforms baselines in low-data peptide engineering tasks with up to 500 samples.

• Why it matters: Peptide engineering faces a major bottleneck due to limited experimental data, making it challenging to predict effective models before costly wet-lab experiments.
• What they did: The authors developed SCARSE and benchmarked it across 24 diverse datasets, demonstrating superior performance in substitution and indel tasks and effective active learning in low-data regimes.
• The result: SCARSE can identify high-performing peptides up to 7 times more efficiently than random sampling, with CV R2 serving as a practical criterion for model suitability, enabling more data-efficient discovery.