Predicting the Fourth Analogue: Learning Stereochemical Transfer in Cyclic-Peptide Permeability

Combining stereochemical edits offers a direct way to explore cyclic-peptide chemical space. Once a reference peptide and two single-edit analogues have been measured, the modeling question is how molecular structure improves prediction beyond those three observations. We define this conditional permeability task using 1,154 chemically certified squares from four publications and evaluate transfer across 22 complete chemical families. The three measured analogues provide an accurate local anchor: adjacent averaging reduces source–family balanced mean absolute error from 0.319 for additive completion to 0.234 log-permeability units. Structural learning improves this informed prediction. A monomer-alignment difference kernel reaches 0.211 under family holdout, a 10% reduction, and kernelneighbor calibration reaches 0.217 under family holdout and 0.227 under source holdout. Supervision offers a second route to the gain: with the peptide representation fixed, changing raw-difference squared loss to hybrid pinball loss reduces source-held-out error from 0.285 to 0.232. Matched controls connect these improvements to the combination of structural representation, learning target and calibration by measured neighbors. Shared-edit and path-consistency diagnostics distinguish the information used by a correction, while the observed edit–background coverage identifies a concrete next experiment: crossing the same edits across new families under one assay protocol. The study establishes the predictive value of structural learning beyond three measured analogues and provides an independently usable task for testing that contribution. Scientific contribution. This work turns cyclic-peptide edit prediction into a fixed-information task in which local measurements anchor the target and learned structural relationships supply additional accuracy. Chemically certified queries, complete-family holdouts and matched learning comparisons identify how that additional information can be acquired and evaluated.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-13
DOI
https://doi.org/10.26434/chemrxiv.15012172/v1
Primary Topic
Computational Drug Discovery Methods
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Predicting the Fourth Analogue: Learning Stereochemical Transfer in Cyclic-Peptide Permeability

Hiroto Saigo, Guanjie Zou, Minrui Chen, Bowen Song et al.
6 links
ChemRxiv
Computational Drug Discovery Methods
preprint

Predicting the Fourth Analogue: Learning Stereochemical Transfer in Cyclic-Peptide Permeability

Hiroto Saigo, Guanjie Zou, Minrui Chen, Bowen Song, Minqi Chen
preprint en

Abstract

Combining stereochemical edits offers a direct way to explore cyclic-peptide chemical space. Once a reference peptide and two single-edit analogues have been measured, the modeling question is how molecular structure improves prediction beyond those three observations. We define this conditional permeability task using 1,154 chemically certified squares from four publications and evaluate transfer across 22 complete chemical families. The three measured analogues provide an accurate local anchor: adjacent averaging reduces source–family balanced mean absolute error from 0.319 for additive completion to 0.234 log-permeability units. Structural learning improves this informed prediction. A monomer-alignment difference kernel reaches 0.211 under family holdout, a 10% reduction, and kernelneighbor calibration reaches 0.217 under family holdout and 0.227 under source holdout. Supervision offers a second route to the gain: with the peptide representation fixed, changing raw-difference squared loss to hybrid pinball loss reduces source-held-out error from 0.285 to 0.232. Matched controls connect these improvements to the combination of structural representation, learning target and calibration by measured neighbors. Shared-edit and path-consistency diagnostics distinguish the information used by a correction, while the observed edit–background coverage identifies a concrete next experiment: crossing the same edits across new families under one assay protocol. The study establishes the predictive value of structural learning beyond three measured analogues and provides an independently usable task for testing that contribution. Scientific contribution. This work turns cyclic-peptide edit prediction into a fixed-information task in which local measurements anchor the target and learned structural relationships supply additional accuracy. Chemically certified queries, complete-family holdouts and matched learning comparisons identify how that additional information can be acquired and evaluated.

ChemRxiv
Kyushu University (JP), Sichuan University (CN)
Japan Society for the Promotion of Science
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon