A Multi-Branch 3D-Aware Deep Learning Framework for Generalizable Small-Molecule Drug Discovery at HTS-Scale Performance

Overview

Currently, small-molecule drug discovery for diseases such as Alzheimer’s (AD) remains challenging due to the inefficiency and costliness of traditional high-throughput screening (HTS), regularly costing numerous days and thousands to millions of dollars. This study aims to develop HTS3D: a high-throughput AI-powered compound screening platform to discover small-molecule inhibitors as therapies for diseases, including AD. HTS3D is the first four-branch virtual screening platform that accounts for 3D protein-ligand interactions to identify promising small-molecule inhibitors. A training dataset of 610 molecules with confirmed NLRP3 inhibitory activity, as well as a discovery dataset of 1,000 unknown compounds, was extracted from ChEMBL. Additionally, the trained NLRP3 discovery model was tested on a set of 15 promising candidates discovered through manual virtual screening. The generalizability of HTS3D was evaluated using a previously synthesized set of 719 molecules relating to immune receptor CD-28 inhibition. Ultimately, the HTS3D platform achieved satisfactory precision and recall for the identification of both NLRP3 (above 80%) and CD-28 (above 60%) inhibitors. Additionally, the NLRP3 model identified 110 molecules in the NLRP3 discovery dataset as promising for experimental validation. The platform therefore exhibits comparable performance metrics to modern AI screening platforms as well as manual virtual screening, and surpassed experimental high-throughput screening in terms of cost and efficiency, being around 1.7 times faster. HTS3D has immediate applications and impact within drug development by filtering for promising compounds and accounting for various protein interaction possibilities to vastly reduce the wet-lab screening burden for drug discovery research.