Table 4. Top 10 highest average importance descriptors from each data domain, as measured in the tripartite models. Overall, chemical descriptors are most important – although not all important descriptors are chemical. Protein names and cell lines are more interpretable than individual chemical descriptors.

MOE chemical descriptors			Protein target descriptors		Cytotoxicity descriptors
Ave. imp.	Code	Description	Ave. imp.	Protein name	Ave. imp.	Cell line and dose
90.6	a_ICM	Atom information content (mean)	35.5	Heat shock protein HSP 90-beta	33.5	Hek293, 92 μM
86.2	density	Molecular mass density	31.1	Endothelin B receptor	30.7	N2A, 92 μM
47.9	KierA3	A Kappa Shape index	29.3	KiSS-1 receptor	26.9	H4lle, 92 μM
47.0	SlogP_VSA9	A subdivided surface area descriptor	29.1	Carbonic anhydrase 14	25.9	jurkat, 46 μM
44.1	GCUT_PEOE_0	An adjacency and distance matrix descriptor	29.0	Tyrosine-protein phosphatase non-receptor type 7	21.3	jurkat, 92 μM
40.5	GCUT_PEOE_2	An adjacency and distance matrix descriptor	28.2	Plasminogen activator inhibitor 1	21.3	MRC5, 92 μM
39.2	PEOE_VSA-2	A partial charge descriptor	28.1	Somatostatin receptor type 4	21.2	SKNSH, 92 μM
38.1	BCUT_SLOGP_3	An adjacency and distance matrix descriptor	23.8	E3 ubiquitin-protein ligase Mdm2	16.6	N2A, 46 μM
34.0	GCUT_SMR_0	An adjacency and distance matrix descriptor	22.0	Multidrug resistance protein 1	15.5	SHSY, 46 μM
30.8	PEOE_RPC+	A partial charge descriptor	20.9	C5a anaphylatoxin chemotactic receptor 1	14.0	H4lle, 46 μM