Machine learning framework for predicting DPP-4 inhibitory activity: a preliminary study on candidate screening and structural determinants
Journal of Medicine and Life Science · 31.05.2026 · 23 (2) · 81-92
DOI: 10.22730/jmls.2026.04.15 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Medicine and Life Science, erschienen am 31.05.2026. Verfasst von Kwang Hyeon Kim und 1 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
This preliminary study developed a machine learning-based computer-aided drug design framework to predict the inhibitory activity of dipeptidyl peptidase-4 (DPP-4) inhibitors for type 2 diabetes, aiming to streamline the identification of high-potential antidiabetic candidates. The bioactivity data for human DPP-4 were retrieved from ChEMBL, curated, and normalized to the pIC50 scale. The chemical structures were digitized into 2048-bit Morgan fingerprints. We implemented a dual- model random forest architecture using an 80/20 train-test split to perform qualitative activity classification (threshold pIC50≥6.0) and quantitative potency regression. The model robustness was evaluated using the area under the receiver operating characteristic curve (ROC-AUC) and shapley additive explanations (SHAP) analyses to ensure predictive accuracy and structural interpretability. The random forest classifier achieved superior performance with an ROC-AUC of 0.91, overall accuracy of 91.17%, a precision of 0.9281, and an F1-score of 0.9393. The regression model achieved moderate predictive accuracy suitable for early-stage candidate prioritization. The SHAP analysis identified key structural determinants, notably fingerprint bit 699, as major contributors to the inhibitory activity. In practical screening, the pipeline identified Candidate A as a high-potential lead with a predicted pIC50 of 6.4 and an 80.0% activity probability. Furthermore, the model validated the drug-likeness of successful candidates using Lipinski’s rule of five, ensuring favorable physicochemical properties for oral bioavailability. This integrated quantitative structure-activity relationship framework significantly enhanced drug discovery efficiency by providing accurate qualitative and quantitative predictions. By combining 91% classification accuracy with structural insights, this methodology offers a robust data-driven tool for the accelerated validation of novel DPP-4 inhibitors.
Zitieren
APA
Kim, K. H., & Cho, J. (2026). Machine learning framework for predicting DPP-4 inhibitory activity: a preliminary study on candidate screening and structural determinants. Journal of Medicine and Life Science, 23(2), 81-92. https://doi.org/10.22730/jmls.2026.04.15
Vancouver
Kim KH, Cho J. Machine learning framework for predicting DPP-4 inhibitory activity: a preliminary study on candidate screening and structural determinants. Journal of Medicine and Life Science. 2026;23(2):81-92. doi:10.22730/jmls.2026.04.15
BibTeX
@article{Kim2026,
author = {Kim, Kwang Hyeon and Cho, Jae-Keun},
title = {Machine learning framework for predicting DPP-4 inhibitory activity: a preliminary study on candidate screening and structural determinants},
journal = {Journal of Medicine and Life Science},
year = {2026},
volume = {23},
number = {2},
pages = {81--92},
doi = {10.22730/jmls.2026.04.15}
}
EUVIMED