A Hybrid Machine Learning Pipeline for Reliable Prediction of Potential HIV-1 Inhibitors

The discovery of potent antiviral inhibitors remains a major challenge in combating viral infections. In this study, we present a hybrid computational pipeline that integrates machine learning for accurate prediction of small-molecule HIV-1 inhibitors. Five classification algorithms were trained on 7552 known inhibitors from ChEMBL using five classes of molecular fingerprints. Among these, Random Forest (RFC) models consistently outperformed the others, achieving accuracy values of 0.9526 to 0.9932, while K-Nearest Neighbors (KNN) and Multilayer Perceptron (MLP) models, although slightly less accurate, still demonstrated robust performance, with accuracies ranging from 0.9170 to 0.9482 and 0.9071 to 0.9179 for selected descriptors, respectively. Based on model predictions, 4511 natural compounds from the COCONUT database were identified as potential inhibitors. After 3D shape similarity filtering (Tanimoto Combo > 1 and Shape Tanimoto > 0.8), eight top-ranked compounds were prioritized for further assessment of their physicochemical, ADMET, and drug-likeness properties. Two natural compounds, CNP0194477 and CNP0393067, were identified as the most promising candidates, showing low cardiotoxicity (hERG risk: 0.096 and 0.112), favorable hepatotoxicity and genotoxicity profiles, and good predicted oral absorption. This integrated workflow provides a robust and efficient computational strategy for the identification of natural compounds with antiviral potential, facilitating the selection of promising HIV-1 inhibitors for further experimental validation.