Machine learning for peptide discovery

Peptides play important roles in physiological processes and serve as promising starting points for drug development. However, discovering peptides remains challenging due to limited annotated data, experimental noise and complex follow-up experiments to determine function. Moreover, the vast sequence space also makes subsequent peptide optimization data-intensive. We introduce machine learning methods to address various aspects of peptide discovery and design. These include identifying bioactive peptides from mass spectrometry data, predicting peptide cleavage events in proteins, and computationally deorphanizing peptide receptors. Additionally, we investigate DNA language models, which may become relevant in the future as potential tools for computationally identifying novel peptide-coding sequences. For peptide optimization, we introduce a novel acquisition function for batch Bayesian optimization. This enables efficient, controllable parallel exploration of design spaces, addressing the need for data-efficient parallel methods in peptide engineering. Our work demonstrates how machine learning can enhance peptide research across the discovery pipeline from identification to functional characterization and molecule optimization. While ML methods show promise, integrating computational approaches with experimental efforts remains crucial for realizing their full potential in peptide drug development.