Improving QSAR model predictions using ensembled heterogenous features

Quantitative structure–activity relationship (QSAR) models are widely used computational tools in drug discovery for predicting molecular activities and prioritizing compounds for experimental testing. However, many existing QSAR methods still suffer from limited applicability domains, overfitting, and instability across different datasets. To address these challenges, we present GT-QSAR, an automated machine learning workflow for structure-centric QSAR modeling that integrates heterogeneous molecular representations through ensemble learning. GT-QSAR supports both receptor- and ligand-based modeling strategies and combines diverse feature groups, including molecular descriptors, pharmacophore fingerprints, graph-based embeddings, grid-based shape and electrostatic representations, and protein–ligand interaction features. Systematic benchmarking using the Papyrus dataset demonstrates that GT-QSAR provides improved predictive accuracy, model stability, and generalizability compared with control tools. In addition, virtual screening experiments using the LIT-PCBA benchmark show that GT-QSAR substantially improves enrichment compared with conventional similarity-based and docking-based screening methods and previously reported target-specific machine learning models. These results demonstrate that GT-QSAR provides a robust and generalizable framework for QSAR modeling, enabling improved prioritization of active compounds in realistic drug discovery campaigns. The data used for QSAR modeling as part of this work is made publicly available. GT-QSAR is available online at https://gtqsar.standigm.com.