A knowledge-aware, constraint-guaranteed decoder for cyclic peptides

Cyclic peptides reach a prized middle ground between small molecules and biologics: large enough to grip the flat protein–protein interfaces that defeat conventional drugs, yet sometimes lipophilic enough to be taken as a pill. Deep generative models have transformed small-molecule and protein design, but cyclic peptides are a harder frontier, with a design space of non-canonical residues, N-methylation, D-amino acids and macrocyclic rings that is far too large to search by hand. The models built for it are also hard to evaluate: they are often scored by distribution metrics, judged by the predictor that trained them, or compared with untuned baselines at unequal budgets — tests on which a weak method can look strong. Here we present IH-LM, a cyclic-peptide generator that compiles the design rules directly into the decoder, so every molecule it completes satisfies them by construction; across a 900-molecule test set it also discarded no rule-satisfying string. The preliminary experimental results are promising. Benchmarked against tuned baselines at a matched sampling budget, it produces about three times as many valid, novel molecules as enumerating a monomer library (773 versus 246 per 1,000 samples) and spans about three times as many distinct scaffolds for the same effort. Because validity is guaranteed rather than scored, the full strength of evaluation can be devoted to the properties that decide a candidate’s fate — chemical diversity, permeability and, ultimately, target activity — which is where real progress in cyclic-peptide design will be won.