Fig. 4: Evaluations on protein-ligand affinity prediction task.

a We utilize RMSD intervals of 1 Å to calculate the average prediction values of all conformations for various methods. PoseRank is determined by sorting the energy from 1 to 10. To facilitate display, we multiplied the predicted values of PoseScore by 10. b The box plot presents the affinity prediction performance across four real-world test sets, categorized by the maximum similarity of ligands found in the training set. The ChEMBL-Kinase dataset (n = 2539), and the Mpro covalent test (n = 142), are derived from ChEMBL and patent databases. The LSD1 project (n = 55) and the Mpro project (n = 22) originate from our internal real-world drug design projects. c (left) The most potent compound, Cpd27 in the LSD1 project achieves an affinity of 0.7 nM. c (middle, right) The correlation plot between the Interformer predicted affinity value and experimental pIC50 in LSD1 project and prediction of Interformer without negative samples for training. d (left) The most potent compound, TAD 6-ref, in the Mpro project achieves an affinity of 16 nM. d (middle, right) The correlation plot in the Mpro project and prediction of Interformer without negative samples for training. e (left) The docking pose of Cpd27 aligns with the PDB ID: 6W4K, it holds crucial hydrogen bond interaction with LYS-661 and creating two additional hydrogen bond interactions with GLN-358 and ALA-539. e (right) The docking pose of TAD 6-ref aligns with the PDB ID: 7RFS and is designated to form a macrocyclic to stabilize the ligand. For the box plots in (b), the lower limit represents lower quartile, the center line represents the median and the upper limit represents the upper quartile. The whiskers do not include outliers. For the regression plot in 4c and 4d, the blue line indicates linear regression fit. The light-blue region indicates the corresponding 95% confidence interval computed via bootstrapping mean.