Condition-Aware and Leakage-Resistant Machine Learning of Short-Peptide Self-Assembly Across Bulk State and Supramolecular Morphology
- 1Department of Industrial Chemistry
- 2Enugu State University of Science and Technology (ESUT
Abstract
Short-peptide self-assembly is governed jointly by molecular structure and experimental environment, yet many predictive workflows treat condition-specific observations as independent samples. Here, a source-audited dataset was reconstructed from PeptideMiner and SAPdb to evaluate bulk-state and supramolecular-morphology prediction under leakage-resistant validation. Peptide identity was defined from sequence,terminal modifications,non-terminal modifications,peptide-system annotations,and conjugate partners; experimental descriptors included concentration, pH, temperature, thermal history, incubation, and explicit solvent composition. A conventional random row split was shown to be strongly non-independent: 87.3% of test molecular identities had already appeared in training and 94.2% of test rows came from molecules represented in training.Accordingly,all primary claims were based on molecular-identity-grouped validation and permanently sealed holdouts containing previously unseen molecular systems. For hydrogel versus non-gel prediction, a tuned Random Forest achieved ROCAUC 0.977, PR-AUC 0.882, balanced accuracy 0.945, F1 0.880, and MCC 0.836 on 115 conditions from 20 unseen molecular identities. Solvent identity and composition formed the dominant transferable feature block,but analyses restricted to fully aqueous conditions demonstrated that non-solvent variables,particularly concentration,pH,temperature,and related processing descriptors,retained substantial predictive information. A harder three-class non-gel/hydrogel/organogel task yielded macro ROC-AUC 0.783 and MCC 0.247 on unseen molecules, with organic non-gelling conditions constituting the dominant error mode. For broad morphology, an Extra Trees classifier achieved macro ROC-AUC 0.980, macro PR-AUC 0.925, and MCC 0.665 on 59 conditions from 14 unseen molecular identities, although spherical/particulate morphology remained comparatively difficult.Fine-grained morphology prediction was not forced when independent class support proved insufficient for stable molecule grouped cross-validation. These results establish that reliable learning of short-peptide self-assembly requires explicit representation of experimental environment, source-level curation, and validation at the level of independent molecular systems rather than condition rows.
Related
- Supramolecular Architecture of Chiral Macrocycles: Hierarchical Assembly and Confinement Effect for Enhanced Circularly Polarized Luminescence
- Cucurbit[7]uril-mediated aqueous self-assembly and supramolecular functionalisation of selenium nanoparticles
- Chemically Fueled Spatiotemporal Patterns from Charge-transfer Supramolecular Assembly
- Supramolecular modular assembly of polymer brush-grafted responsive hydrogels into biomimetic soft actuators
- Solvent heterogeneity regulates assembly pathways in supramolecular polymerization
BibTeX
@misc{orji2026conditionaware,
title = {{Condition-Aware and Leakage-Resistant Machine Learning of Short-Peptide Self-Assembly Across Bulk State and Supramolecular Morphology}},
author = {Dennis Obinna Orji},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008469/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008469/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Condition-Aware and Leakage-Resistant Machine Learning of Short-Peptide Self-Assembly Across Bulk State and Supramolecular Morphology
AU - Dennis Obinna Orji
PY - 2026
DA - 2026/09/08
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008469/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008469/v1
AB - Short-peptide self-assembly is governed jointly by molecular structure and experimental environment, yet many predictive workflows treat condition-specific observations as independent samples. Here, a source-audited dataset was reconstructed from PeptideMiner and SAPdb to evaluate bulk-state and supramolecular-morphology prediction under leakage-resistant validation. Peptide identity was defined from sequence,terminal modifications,non-terminal modifications,peptide-system annotations,and conjugate partners; experimental descriptors included concentration, pH, temperature, thermal history, incubation, and explicit solvent composition. A conventional random row split was shown to be strongly non-independent: 87.3% of test molecular identities had already appeared in training and 94.2% of test rows came from molecules represented in training.Accordingly,all primary claims were based on molecular-identity-grouped validation and permanently sealed holdouts containing previously unseen molecular systems. For hydrogel versus non-gel prediction, a tuned Random Forest achieved ROCAUC 0.977, PR-AUC 0.882, balanced accuracy 0.945, F1 0.880, and MCC 0.836 on 115 conditions from 20 unseen molecular identities. Solvent identity and composition formed the dominant transferable feature block,but analyses restricted to fully aqueous conditions demonstrated that non-solvent variables,particularly concentration,pH,temperature,and related processing descriptors,retained substantial predictive information. A harder three-class non-gel/hydrogel/organogel task yielded macro ROC-AUC 0.783 and MCC 0.247 on unseen molecules, with organic non-gelling conditions constituting the dominant error mode. For broad morphology, an Extra Trees classifier achieved macro ROC-AUC 0.980, macro PR-AUC 0.925, and MCC 0.665 on 59 conditions from 14 unseen molecular identities, although spherical/particulate morphology remained comparatively difficult.Fine-grained morphology prediction was not forced when independent class support proved insufficient for stable molecule grouped cross-validation. These results establish that reliable learning of short-peptide self-assembly requires explicit representation of experimental environment, source-level curation, and validation at the level of independent molecular systems rather than condition rows.
ER -