Table 3.

Matthews correlation coefficients (MCCs) of single-task DNN models on cross validation data.

Feature set	Size	Biliary hyperplasia		Fibrosis		Necrosis
		Mean	SD	Mean	SD	Mean	SD
Gene-level feature sets
Toxicity Module (L1000)	154	0.85	0.09	0.77	0.15	0.53	0.07
PTGS (core)	199	0.82	0.11	0.73	0.12	0.51	0.06
A200	200	0.83	0.09	0.66	0.12	0.51	0.07
A600	600	0.87	0.08	0.75	0.18	0.55	0.08
L1000	978	0.87	0.06	0.72	0.18	0.56	0.10
A1200	1,200	0.85	0.05	0.77	0.13	0.55	0.09
Toxicity Module Gene	1,312	0.89	0.07	0.78	0.13	0.55	0.09
PTGS (all)	1,331	0.85	0.08	0.77	0.17	0.54	0.09
Pathway-level feature sets
MSigDB (hallmark)	50	0.75	0.13	0.67	0.14	0.49	0.09
Toxicity Module	89	0.81	0.09	0.65	0.17	0.48	0.10
MSigDB (C2) L1000	1,220	0.83	0.09	0.67	0.21	0.53	0.08
MSigDB (C2)	1,329	0.82	0.08	0.68	0.16	0.51	0.07

Bold indicates the greatest mean MCC value for each injury phenotype.