Peptide lead optimization stops enumerating sequences: a linear model plus dynamic programming returns the in-model optimum across 10⁵⁰ candidates

Peptide lead optimization can now deterministically return the in-model optimal sequence from a space of more than 10⁵⁰ candidates without scoring every sequence one by one. Conventional QSAR models had to enumerate and score full sequences to identify the best candidates, and noncanonical amino acid design spaces make exhaustive search infeasible. On the A-2 series of 14-mers, dynamic programming found sequences with predicted scores of 11.576 and 12.143 in two design spaces, above the 10.928 and 10.517 reached by a genetic algorithm.

The model splits each amino acid into position-specific substructure fingerprints, so it can score noncanonical amino acids that never appeared in the training set. One-hot-encoded Free-Wilson models can only recognize amino acids already present in the training data and cannot extrapolate to new monomers. Across a library of about 4000 candidate monomers, the model gave RbMeBip4CO2H a score of 2.73, higher than the scores of synthesized monomers.

Coefficients of the linear model read directly as each substructure's contribution to activity, with the sign giving direction and the magnitude giving strength. Nonlinear models such as random forests rely on permutation or node purity to generate importances and may not produce explicit signs. In the A-2 series, the model attributed the potency contribution of RbMeY at P9 to its R-beta-methyl substructure, and that attribution was later confirmed by experimental data.

The framework was evaluated on 11 peptide series from 5 proprietary drug discovery projects, with peptide lengths from 14 to 17 and 415 to 2013 peptides per series, and activity data collected from the same assay type within each series. It targets lead optimization of noncanonical-amino-acid-containing peptides and returns candidate sequences ranked by predicted potency for downstream synthesis and experimental validation. The code is open-sourced on GitHub with an example dataset extracted from patent WO2024215820, but the full proprietary data used to reproduce the benchmarking results cannot be shared because of IP restrictions.

Linear model weights reflect statistical correlation rather than causal effect, and when the dataset carries bias, high-performing monomers repeatedly used as substrates for mutating other positions can receive underestimated scores. The additive linear model has no mechanism to detect a saturation boundary, so a substructure that was favorable within the training range keeps receiving a favorable weight when extrapolated, which can over-predict the benefit of ever-larger or ever-more-decorated monomers. Predictions are reliable only within the substructure-level applicability domain, becoming less reliable when a candidate monomer introduces substructures the model has never observed; the authors recommend treating such scores as hypothesis-generating rather than quantitative and combining them with orthogonal signals such as 3D structural modeling, chemist judgment, or confirmatory assays.