NNGeoGVS: A Unified Equivariant Graph Neural Network Framework for Integrated Structure and Ligand-based Virtual Screening.

Abstract

Introduction: Virtual screening (VS) is a widely used computational strategy in early- stage drug discovery for identifying bioactive compounds from large chemical libraries. Although structure-based and ligand-based approaches have demonstrated success individually, their integration into a unified deep learning framework remains limited. This study presents NNGeoGVS (Neural Network Geometric Graph-based Virtual Screening), an Equivariant Graph Neural Network (EGNN)-based framework that integrates ligand molecular topology and protein active-site geometry for activity prediction.

Methods: NNGeoGVS represents ligands and protein binding sites as molecular graphs and learns structure-activity relationships through end-to-end message passing using PyTorch Geometric. The framework combines node-level atomic features, edge-level bond descriptors, and global geometric and pharmacophore descriptors. Active-site geometry is characterized using Ultra-Fast Shape Recognition (USR) and convex hull algorithms. These features are integrated through a multi-branch GNN architecture with late fusion, while focal loss optimization and graph-level augmentation address class imbalance.

Results: NNGeoGVS was evaluated on the DUD-E dataset (8 kinase targets) and the LITPCBA benchmark (15 targets). Compared with NNGeoAS and other GNN-based methods, it achieved average EF1% values of 35.92 and 17.03 and ROC-AUC scores of 0.98 and 0.77 on DUD-E and LIT-PCBA, respectively.

Discussion: The integration of ligand topology with protein active-site geometry improves enrichment and classification performance, demonstrating the effectiveness of combining complementary structural information within a unified EGNN framework.

Conclusion: We proposed NNGeoGVS, a robust framework for virtual screening that integrates structural information of both proteins and ligands and demonstrated robust predictive performance on benchmark datasets.