Interpretable machine learning framework for peptide lead optimization in ultralarge design spaces.

In modern peptide drug discovery, the incorporation of noncanonical amino acids (ncAAs) has become a powerful strategy to improve potency and key developability properties, including stability, solubility, and oral bioavailability. This shift has expanded peptide chemical diversity and design spaces well beyond natural residues, creating fundamental challenges for predictive methods (e.g., quantitative structure-activity relationship modeling) used in lead optimization: limited transferability from natural residues to novel ncAAs, limited interpretability, and poor scalability to ultralarge peptide spaces (>1050). We present PepOpt, an interpretable machine learning framework that unifies molecule-based and sequence-based peptide representations in linear models and exploits dynamic programming to efficiently explore ultralarge peptide spaces. PepOpt achieves high predictive accuracy and interpretability, supports training and inference across diverse ncAAs, including ncAAs absent from the training set whose constituent substructures are represented in the training data. Its linear structure decomposes predictions into residue-specific contributions, enabling systematic optimization without exhaustive sequence enumeration. Notably, PepOpt is inherently capable of multi-parameter optimization without specialized modifications. Together, these capabilities position PepOpt as a valuable platform for prediction, interpretation, ideation, and rapid exploration in ncAA-rich ultralarge peptide chemical spaces, accelerating the rational design of novel peptide therapeutics.