Year: 2026
Type: preprint
Abstract: The same local modification can enhance cyclic-peptide permeability in one chemical background and reduce it in another. Molecular difference learning predicts such changes from structure pairs, yet whole-molecule descriptors can merge stereochemically distinct edits. We test whether the residue relationships changed by an edit supply additional predictive information. Our edit-relational Gaussian process combines parent descriptors and their changes with signed counts of gained and lost residue relationships, preserving stereochemistry and ring, branch and tail connectivity. Across 286 parent pools and 4,213 measured pairs, the best measured improvement over the parent after two queries is 0.588 versus 0.54 log 10 permeability units under equal component weights. Relational inputs improve this gain in all five evaluated chemical components under matched fitting and initial-mean ranking. Measured interaction panels contain 13 edit-preference reversals among 72 distinct-site comparisons, exposing the background dependence that the representation is designed to capture. Exact finite-pool analysis separates candidate resolution from measurement allocation. Controls sharing the first mean-ranked query locate the principal measured benefit in the representation and initial ranking. The findings motivate matched-edit experiments on new parents to test whether these relationships predict how permeability benefits transfer between backgrounds. An edited residue’s changing relationships provide a concrete, predictively useful description of its chemical context.
Source: ChemRxiv
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
Award: G1477135921
OpenAlex ID: W7225921401
