Fig. 2
From: End-to-end protein–ligand complex structure generation with diffusion-based generative models
Benchmark results on the PDBbind test set. A Median modeling accuracies for each complex in the test set compared between our models DPL (blue) and DPL+S (orange). B Fraction of complexes where a structure with TM-score above the threshold, varying from 0.5 to 1.0, was sampled. C Fraction of complexes where a structure with L-rms below the threshold, varying from 0 Å to 5 Å, was sampled. The results obtained with the molecular docking methods (GNINA and AutoDock Vina) are also shown. D Performance as a function of the number of generative samples. Two thresholds 2 Å and 5 Å are employed for L-rms. E Relation between performance and the number of ligand rotatable bonds. F Relation between performance and the number of related training data