BEYOND THE BOTTLENECK: SCALING GPU-ACCELERATED MOLECULAR DOCKING AND MACHINE LEARNING INTEGRATION FOR RAPID NEUROPHARMACOLOGICAL LEAD DISCOVERY

The discovery of novel neurotherapeutics, particularly for complex membrane-bound targets such as G protein-coupled receptors (GPCRs), requires efficient screening of extensive chemical libraries. Conventional CPU-based virtual ligand screening (VLS) workflows often impose limitations on throughput and pose accuracy. This study investigates GPU-accelerated molecular docking combined with machine learning refinement to overcome these bottlenecks in neuropharmacological lead discovery.
We benchmarked the ICM-PRO docking platform by comparing traditional CPU-based VLS with GPU-optimized implementations, including RIDGE and GINGER-GPU tools. Performance and accuracy were evaluated against widely used tools such as GOLD, AutoDock Vina, and DiffDock using the Astex Diverse Set and the PoseBusters benchmark. To enhance pose selection, we developed an automated machine learning (AutoML) framework in PyCaret that integrates classical force-field descriptors with AI-based scoring functions.
GPU acceleration achieved up to a 3000-fold increase in docking throughput using RIDGE, while GINGER-GPU reduced conformer generation time by over 100-fold. In pose accuracy assessments, ICM-VLS and RIDGE outperformed deep learning-based approaches for PoseBusters dataset, achieving success rates of 80.5% and 67.5%, respectively (RMSD ≤ 2 Å and PoseBusters Valid). The AutoML refinement layer reduced false-positive rates by identifying native-like binding poses from high-speed screening outputs.
The integration of GPU-accelerated docking with ML-driven pose refinement enables rapid yet accurate large-scale virtual screening. This scalable framework offers a powerful approach for drug discovery, facilitating the identification of subtle ligand–receptor interactions critical for next-generation neurotherapeutics.