Table 1. The average predictive performance comparison between DEEPScreen and various novel DL-based and conventional DTI predictors.

Dataset	Reference	Method/architecture	Performance (metric)
ChEMBL temporal-split dataset		DEEPScreen: DCNN with 2-D images	0.45 (MCC)
	Lenselink et al.18	Feed-forward DNN PCM (best model)	0.33 (MCC)
		Feed-forward DNN	0.30 (MCC)
		SVM	0.29 (MCC)
		LR	0.26 (MCC)
		RF	0.26 (MCC)
		Naïve Bayes	0.10 (MCC)
Maximum unbiased validation (MUV) dataset		DEEPScreen: DCNN with 2-D images	0.88 (AUROC)
	Kearnes et al.11	Graph convolution NNs (W2N2)	0.85 (AUROC)
	Ramsundar et al.49	Pyramidal multitask neural net (PMTNN)	0.84 (AUROC)
		Multitask neural net (MTNN)	0.80 (AUROC)
		Single-task neural net (STNN)	0.73 (AUROC)
		RF	0.77 (AUROC)
		LR	0.75 (AUROC)