Fig. 5

Maximum Pearson correlation of the docking scores of each protein in DOCKSTRING to those of any other protein. The maximum correlation of each protein suggests how challenging it would be for meta-generalization. We identified three proteins from the DOCKSTRING regression benchmark (orange bars) that represented a range of dissimilarity to other proteins: PARP1 (highly correlated, very similar), ESR2 (medium correlation and similarity) and PGR (low correlation and similarity). We selected these proteins as our meta-test set, and all proteins not in the regression benchmark as meta-training set. In this way we aimed to illustrate meta-generalization across a range of correlations