Computational & AI chemRxiv
Pairwise learning on a matched-pair graph predicts which modification boosts cyclic peptide permeability
- CycPeptMMP predicts the permeability change between two peptides linked in a matched-pair graph, rather than one peptide's absolute value, sidestepping differences between publications.
- The graph was built for cyclic peptides using four common modification types, covering several thousand peptides from seven literature sources; five GNN backbones and repeated random splits compared pair and single modes.
- Evaluation was purely retrospective on curated data: pair mode helps most for smaller sources and for single-modification analogues, while single mode wins on larger sources.
- A censored loss handles clamped measurements, and pair mode ranks such pairs more accurately.
Predicting the delta between two close analogues is how med chemists actually reason, and it cancels much of the assay and lab offset that wrecks absolute permeability models across papers. The catch: random splits keep near-identical analogues on both sides, so gains may be optimistic. Worth testing with scaffold or source-held-out splits, then a prospective round.
Paper: chemRxiv ↗ Pairwise learning of cyclic peptide membrane permeability on a matched-molecular-pair graph