Fig. 5: Evaluation of conformation quality for AI generated models.
From: Assessing conformation validity and rationality of deep learning-generated 3D molecules
Molecule generative models including Lingo3DMolv23, Pocket2Mol5, PocketFlow2, TargetDiff6, and PMDM4 are involved in this evaluation. a–c Distributions of target counts based on the passing rates of high-energy atom detector (HEAD) for ligand-protein interaction evaluation, HEAD for ligand conformation evaluation, and torsional energy descriptor (TED) for torsion energy evaluation, with histograms presented in panels a–c, respectively. Mean values are indicated with red dashed lines. d Scatter plot of HEAD and TED passing rates on a per-target basis, with points colored according to the AI models. Passing HEAD indicates that both the pocket-ligand interaction and conformation validity tests were successfully passed. e Bubble chart providing a comprehensive evaluation of AI models, where the position of each bubble centroid corresponds to the per-target average of HEAD and TED passing rates. Bubble size encodes the fraction of generated molecules satisfying the drug-likeness thresholds (QED ≥ 0.3; SAS ≤ 5), with the corresponding percentage shown in parentheses. QED stands for Quantitative Estimate of Drug-likeness, and SAS stands for Synthetic Accessibility Score.
