ADMET-Informed Machine Learning Prioritization of Hydrogel-Associated Compounds for Nanocarrier Evaluation
Year: 2026
Type: preprint
Abstract: Experimental screening of large hydrogel-associated chemical collections is slow and resource-intensive. We developed an ADMET-informed machine-learning workflow for 2060 HydrogelDB-linked compounds, using the median of a separately reported admetSAR3.0 biodegradability-related output to define higher- and lower-score classes and the remaining 115 predicted ADMET properties as model inputs; preprocessing, feature selection, hyperparameter optimization, probability calibration, and threshold selection were confined to the training partitions of nested cross-validation. The calibrated linear SVM achieved a mean held-out MCC of 0.9116 ± 0.0222, an F2-score of 0.9662 ± 0.0112, and a ROC–AUC of 0.9904 ± 0.0048; performance remained high after proxy-feature removal and feature-family ablation, whereas structure-focused models showed a weaker but detectable signal. Strict eligibility criteria retained 19 candidates. Six-component ranking under changes in model outputs and weights identified D-Alanyl-D-Alanine, Prolyl-Glutamate, and Pro-Asp as the most stable leading candidates. N-Glycylalanine showed lower ranking stability, with Isoleucylglycine as the main alternative. All shortlisted compounds require experimental validation of biodegradability, hydrogel formation, and nanocarrier performance.
Source: ChemRxiv
Institution: Iran University of Science and Technology
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7215119686