DatasetAdditionalInfo.MD

Additional Information on BDB and KIBA

The following figure illustrates the binding score distributions and quartiles in the BDB and KIBA datasets. A strong peak at $pK_d = 5$ is observed for BDB, because $K_d$ of weak interactions are often recorded as $K_d \geq 10000$ $(pK_d \leq 5)$.

We create train/validation/test splits out of BDB and KIBA with warm and cold biomolecules (see the manuscript for more details). We report the average number of proteins, ligands, and interactions in the training and test sets in the following table, alongside standard deviations in the parentheses.

Dataset	Fold	#Proteins	#Ligands	#Interactions
BDB	Train	403.4 $\pm$ 2.8	740.8 $\pm$ 19.46	17988.2 $\pm$ 646.45
BDB	Validation	355.0 $\pm$ 5.62	170.0 $\pm$ 11.05	1494.2 $\pm$ 56.17
BDB	Warm	354.4 $\pm$ 3.44	179.6 $\pm$ 5.28	1494.4 $\pm$ 56.32
BDB	Cold Ligand	376.0 $\pm$ 4.38	84.8 $\pm$ 5.53	2448.8 $\pm$ 373.48
BDB	Cold Protein	43.6 $\pm$ 2.15	264.8 $\pm$ 90.17	2360.0 $\pm$ 216.02
BDB	Cold Both	41.4 $\pm$ 3.07	30.8 $\pm$ 11.92	274.6 $\pm$ 36.19
KIBA	Train	200.6 $\pm$ 1.36	1834.6 $\pm$ 6.41	77264.4 $\pm$ 814.94
KIBA	Validation	193.0 $\pm$ 1.67	1467.2 $\pm$ 23.75	6650.2 $\pm$ 69.53
KIBA	Warm	192.0 $\pm$ 3.16	1476.2 $\pm$ 17.7	6650.6 $\pm$ 69.1
KIBA	Cold Ligand	193.0 $\pm$ 2.45	140.0 $\pm$ 5.59	6810.0 $\pm$ 570.52
KIBA	Cold Protein	14.6 $\pm$ 0.8	1296.0 $\pm$ 179.09	6259.6 $\pm$ 1024.25
KIBA	Cold Both	14.0 $\pm$ 1.1	100.2 $\pm$ 14.55	468.6 $\pm$ 37.89