AI-Assisted Molecular Docking for Novel Drug Discovery

Abstract

Molecular docking--the computational prediction of small molecule binding poses and affinities at protein active sites--is a
cornerstone of structure-based drug discovery, yet classical docking methods suffer from limitations in scoring function
accuracy, conformational sampling completeness, and computational throughput that constrain their utility for large-scale
virtual screening of multi-billion-compound chemical libraries. Artificial intelligence has transformed molecular docking
through deep learning scoring functions, generative molecular design, and end-to-end docking architectures that directly
predict binding poses without explicit energy minimisation. This study develops and benchmarks an AI-assisted drug
discovery pipeline integrating four AI components: (i) SE(3)-equivariant GNN-based binding site prediction (SiteNet); (ii)
DiffDock-v2 blind docking with improved confidence calibration; (iii) a graph transformer-based scoring function
(GraphDTA-v2) fine-tuned on the PDBbind 2024 dataset; and (iv) a conditional variational graph autoencoder (CVGAE)
for de novo ligand generation optimised for predicted binding affinity. The pipeline was applied to three high-priority drug
targets: SARS-CoV-2 main protease (Mpro), drug-resistant EGFR (T790M/C797S double mutant), and IDH1 (R132H
oncogenic mutant). Against the CASF-2016 benchmark, GraphDTA-v2 achieved a Pearson correlation of 0.869 between
predicted and experimental pKd values, outperforming AutoDock Vina (r=0.614) and Glide SP (r=0.703). DiffDock-v2
achieved 51.3% top-1 RMSD<2A success rate on PoseBusters v2 benchmark. CVGAE-generated ligands for Mpro
demonstrated mean QED of 0.847 and predicted pKd of 8.41 (nM affinity range), with 12 of 500 generated compounds
predicted with higher affinity than nirmatrelvir. Four top-ranked generated compounds were synthesised and validated by
SPR, with Kd values of 47-284 nM confirming AI-predicted binding, representing successful end-to-end AI-assisted drug
discovery from target to validated hit.