Peptide design spaces pass 1050: a linear model plus dynamic programming returns the global optimum
Synopsis
Noncanonical amino acids push peptide design spaces to the 1050 scale, where exhaustive screening is infeasible. PepOpt encodes each residue as a position-specific molecular fingerprint, predicts potency with linear regression, and uses dynamic programming under an additivity assumption to find the globally optimal sequence deterministically. Across 11 internal peptide series it matches or beats nonlinear models, and in 1020 and 1050 spaces it returns higher predicted scores than a genetic algorithm.
Interpretation
Peptide lead optimization no longer has to score a whole sequence as one large molecule; each position's residue becomes an additive term, so very large spaces can be searched directly. Earlier routes either treated peptides as oversized small molecules and lost residue position information, or used one-hot Free-Wilson models that flag which residue matters but not which part of its structure acts, and cannot extrapolate to noncanonical amino acids absent from training. On 11 peptide series from 5 proprietary drug discovery projects, each with 415 to 2013 peptides of length 14 to 17, ECFP4 fingerprints with Bayesian Ridge regression ranked best in eight of nine evaluation conditions, with an average rank of 1.1.
Predictions decompose into residue-level and substructure-level contributions, so the model can score monomers never seen in training. A one-hot Free-Wilson model cannot score amino acids outside its training set, and fragment-based vocabularies hit out-of-vocabulary problems on novel or chemically modified residues. In series A-2, the model scored about 4000 candidate monomers at position P9; RbMeBip4CO2H, RbMePhe4THP4 and RbMeBip4Cl, none present in training, scored 2.73, 2.48 and 2.37, above synthesized monomers, and the importance of the R-beta-methyl substructure at P9 was confirmed by experimental data.
Dynamic programming turns the search of ultralarge spaces from stochastic sampling into a deterministic solve that guarantees the global optimum within the model. Genetic algorithms depend on stochastic sampling and converge to different local optima depending on initialization and trajectory; exhaustive enumeration with nonlinear models is estimated at 6.3×106 to 1.3×1011 CPU years for the 1020 space and 1038 to 1041 CPU years for the 1050 space. On 14-mer peptides of series A-2, the genetic algorithm converged at generation 29 to 10.928 in the training-monomer space and was still rising at generation 50 to 10.517 in the ~4000-monomer library space, while dynamic programming returned 11.576 and 12.143 and found the global optimum regardless of k.
The same framework optimizes several properties beyond potency without a separate multi-objective pipeline. Conventional QSAR workflows output separate predictions per endpoint, so a genuinely multi-target objective cannot be posed as one tractable problem and users fall back on Pareto fronts or rule-based filters. On the E-1 series, the one with the most samples, time-split validation gave mean PCC from 0.599 to 0.679 across mean residence time, simulated intestinal fluid stability and simulated gastric fluid stability.
Perspective
The framework targets lead optimization of peptides containing noncanonical amino acids, evaluated on 11 peptide series from 5 proprietary drug discovery projects, with lengths of 14 to 17 and potency expressed as 9 − log10(EC50), where each series draws EC50 values from the same assay type. Its applicability domain is defined at the substructure level: predictions are supported for candidate monomers whose constituent substructures appear in training. Code is open-sourced, and the example dataset is extracted from patent WO2024215820.
Linear model weights reflect statistical correlation rather than causal effect, and dataset bias can underestimate the scores of favorable monomers that are repeatedly used as substrates. A linear model has no mechanism to detect boundaries such as pocket saturation, so a substructure favorable within the training range keeps receiving a favorable weight when extrapolated, potentially over-predicting ever-larger or ever-more-decorated monomers. The authors recommend treating scores for monomers rich in out-of-domain fragments as hypothesis-generating rather than quantitative, and combining them with orthogonal signals such as 3D structural modeling, chemist judgment or confirmatory assays. Replacing the linear monomer score with a nonlinear one would break the principle of optimality of dynamic programming and void the correctness guarantee of top-k enumeration, which is left as future work.