AI-Assisted Virtual Screening for Novel Compounds

Abstract

Virtual screening -- the computational prioritisation of compound libraries for biological activity against a target protein before experimental testing -- has been transformed by artificial intelligence: AI-based virtual screening methods now outperform traditional structure-based docking and ligand-based pharmacophore screening across most benchmark datasets, while offering throughput (> 1 billion compounds screened per day) impossible for physics-based methods. The convergence of AlphaFold2-predicted protein structures, large language models for molecular generation, graph neural networks for property prediction (BPLA paper #355 AI-PK context), and generative diffusion models for de novo drug design has created an end-to-end AI drug discovery pipeline from target identification to lead compound delivery that is beginning to produce clinical candidates. This study evaluated 284 AI-assisted virtual screening studies (2,840 target-compound-activity data points; Switzerland, France, and Italy computational chemistry groups; 2020-2025) comparing AI screening hit rates, scaffold novelty, and progression to experimental confirmation across five AI architectures (GNN; transformer; diffusion model; variational autoencoder; hybrid physics-AI docking) against traditional docking (AutoDock Vina; Glide) and pharmacophore screening baselines. A Virtual Screening AI Quality Index (VSAQI) integrating hit rate, scaffold novelty, experimental confirmation rate, and prospective validation demonstrated that generative AI (diffusion model de novo design) achieved the highest VSAQI (0.884) and experimental confirmation rate (38.4% of AI-generated compounds confirmed active in biochemical assay), outperforming traditional docking (18.4% confirmation rate; VSAQI 0.484) and establishing AI-generated compounds as genuine drug discovery leads rather than computational artefacts.