Pairwise learning of cyclic peptide membrane permeability on a matched-molecular-pair graph

Cyclic peptides are a promising drug scaffold that can target proteins considered undruggable by small molecules, but their poor membrane permeability limits further application. Deep learning has been used to lower the wet-lab burden of finding cell-permeable peptides. Most existing models are trained on measurements curated from many publications and predict a single peptide’s permeability directly. That formulation ignores systematic measurement differences between publications, and it does not answer the actual question drug development asks: not whether a peptide is permeable, but which modification would make a given lead peptide more permeable. Here we introduce CycPeptMMP, which learns the permeability difference within a matched molecular pair (denoted as pair mode) instead of a single peptide’s value (denoted as single mode). Conventional molecular pairing approaches do not apply to cyclic peptides, so we constructed a graph of matched molecular pairs over the four modification types commonly used to improve permeability, organizing 5,674 peptides across seven publication sources. We trained both single and pair modes on five graph neural network backbones across 10 random splits. Pair mode performs better on sources with fewer than 1,000 peptides except Wang, whose peptides often lack a close analogue. Pair mode’s advantage is greatest for the closest analogues differing by a single modification. Single mode is preferable for sources with more than 1,000 peptides. We also implemented a censored loss to retain peptides with clamped measurements in the regression. For pairs involving those peptides, pair mode is more accurate when predicting which one in the pair is more permeable. Taken together, we provide a framework for comparing single and pair mode within each source, and show where pair mode is worth using.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-18
DOI
https://doi.org/10.26434/chemrxiv.15012450/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

Pairwise learning of cyclic peptide membrane permeability on a matched-molecular-pair graph

Wei Liu, Hwee Kuan Lee, Jianguo Li, James Lu
7 links
ChemRxiv
Computational Drug Discovery Methods
preprint

Pairwise learning of cyclic peptide membrane permeability on a matched-molecular-pair graph

Wei Liu, Hwee Kuan Lee, Jianguo Li, James Lu
preprint en

Abstract

Cyclic peptides are a promising drug scaffold that can target proteins considered undruggable by small molecules, but their poor membrane permeability limits further application. Deep learning has been used to lower the wet-lab burden of finding cell-permeable peptides. Most existing models are trained on measurements curated from many publications and predict a single peptide’s permeability directly. That formulation ignores systematic measurement differences between publications, and it does not answer the actual question drug development asks: not whether a peptide is permeable, but which modification would make a given lead peptide more permeable. Here we introduce CycPeptMMP, which learns the permeability difference within a matched molecular pair (denoted as pair mode) instead of a single peptide’s value (denoted as single mode). Conventional molecular pairing approaches do not apply to cyclic peptides, so we constructed a graph of matched molecular pairs over the four modification types commonly used to improve permeability, organizing 5,674 peptides across seven publication sources. We trained both single and pair modes on five graph neural network backbones across 10 random splits. Pair mode performs better on sources with fewer than 1,000 peptides except Wang, whose peptides often lack a close analogue. Pair mode’s advantage is greatest for the closest analogues differing by a single modification. Single mode is preferable for sources with more than 1,000 peptides. We also implemented a censored loss to retain peptides with clamped measurements in the regression. For pairs involving those peptides, pair mode is more accurate when predicting which one in the pair is more permeable. Taken together, we provide a framework for comparing single and pair mode within each source, and show where pair mode is worth using.

ChemRxiv
Agency for Science, Technology and Research (SG), National University of Singapore (SG), Singapore Eye Research Institute (SG), Institute of Materials Research and Engineering (SG), Bioinformatics Institute (SG)
Agency for Science, Technology and Research, National Supercomputing Centre Singapore, Biomedical Research Council
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