/We next sought to measure how well interaction patterns transfer between protein-protein and protein-ligand instances. Of all the contact types, 87 had sufficient data for the asymptotic assumptions of the test to be valid (at least 30 samples in protein-ligand data and at least 600 in protein self-interaction data). Of these 87 contact types, 57 had a p-value under 0.05 (not transferable) while the remaining contact types had fairly uniform p-values (Supplemental Figure S5, left panel) that were not correlated with sample sizes in the protein-ligand data (Supplemental Figure S5, right panel). These results indicate that interaction geometries often, but not always, transfer between the two domains. We highlight the full distributions for four contact types that were particularly well sampled, two of which were rejected and two of which were not rejected (Figure 4). In particular, distributions with similar mode structures, but dissimilar density in each mode were rejected by our test as being out-of-distribution (Figure 4, top). This is a potential limitation of our definition of 'transferability' depending on the application of interest.