Fig. 6: Unified Screening Pipeline for AI-Generated 3D Molecules Powered by HEAD-TED.
From: Assessing conformation validity and rationality of deep learning-generated 3D molecules
HEAD strands for a high-energy atom detector. TED stands for torsional energy descriptor. Molecule generative models including Lingo3DMolv23, Pocket2Mol5, PocketFlow2, TargetDiff6, and PMDM4 are involved in this evaluation. ‌a Schematic overview of the sequential screening pipeline for evaluating AI-generated molecules. The pipeline integrates four filters: drug-likeness assessment with passing criteria defined as QED ≥ 0.3 and SAS ≤ 5, pocket-ligand interaction and ligand conformation validity test using HEAD, and conformational rationality test using TED. ‌ QED stands for Quantitative Estimate of Drug-likeness. SAS stands for Synthetic Accessibility Score. b Performance comparison of five AI-driven molecular generative models. The average passing rates for molecules generated on DUD-E targets are shown for each model at every screening step. DUD-E stands for the Directory of Useful Decoys-Enhanced dataset20. c Comparison of the final passing rates of different AI models across various target types. Bars represent the average final passing rate of each model for each target class. DUD-E dataset includes the following target types (with the number of targets in parentheses): GPCR (5), Kinase (26), Protease (15), Other Enzymes (36), Nuclear Receptor (11), Cytochrome P450 (2), Miscellaneous (5), and Ion Channel (2). All values are obtained from a single deterministic processing of the dataset; repeated runs yield identical results. Source data including target name, target type, passing rate at each stage of the screening pipeline can be found in supplementary materials.
