PyaiVS unifies AI workflows to accelerate ligand discovery and yields ABCG2 inhibitors

Abstract

Developing optimized AI models for virtual screening requires coordinated selection of algorithms, molecular representations, and data splitting strategies, yet lacks integrated tools. We present PyaiVS, a Python package that integrates nine machine learning algorithms, five molecular representations, and three data splitting strategies. This study demonstrates that constructing efficient AI-driven virtual screening models for small molecules requires coordinated optimization of algorithm architectures (e.g., prioritizing deep learning models such as GCN, GAT, and Attentive FP), molecular representations (ECFP4/MACCS fingerprints for small datasets and molecular graph-based representations for large-scale data), and data splitting strategies (clustering-based splitting achieving 68.5 % optimal AUC-ROC performance). To demonstrate utility, we combined PyaiVS with pharmacophore modeling and docking to screen 4,188,623 compounds for ABCG2 inhibitors. Experimental validation identified four compounds (C1/C6/C7/C9) binding ABCG2 with sub-100 μM kd values (5.31–51.35 μM) that potentiate topotecan cytotoxicity. PyaiVS streamlines virtual screening by unifying critical components into an accessible platform, freely available at https://github.com/danqingmk/OpenVS_PyaiVS.