Table 1.

Comparison of fastISM with standard ISM, Gradient × Input, Integrated Gradients and DeepSHAPfor three different models with 1000 and 2000 bp length inputs

Architecture	Layers	Input size	Outputs	fastISM	Standard ISM	Gradient × Input	Integrated Gradients	DeepSHAP
Basset	3 conv+3 max pool, 3 fully connected	1000	Single scalar	2.70	27.36	0.04	2.34	1.75
					(10.1)	(<1)	(0.8)	(0.7)
		2000		6.49	100.44	0.08	4.61	3.03
					(15.4)	(<1)	(0.7)	(0.4)
Factorized Basset	9 conv+3 max pool, 3 fully connected	1000	Single scalar	5.47	68.97	0.09	4.82	2.64
					(12.6)	(<1)	(0.9)	(0.5)
		2000		18.04	262.24	0.17	9.47	4.63
					(14.5)	(<1)	(0.5)	(0.25)
BPNet	2 conv, 9 dilated conv, skip connections	1000	Profile (length 1000 vector) + scalar	28.97	46.09	41.49*	4399*	1743*
					(1.6)	(1.4)	(151)	(60)
		2000	Profile (length 2000 vector) + scalar	81.52	173.96	126.41*	12 440*	6427*
					(2.1)	(1.5)	(152)	(78)

Note: All times in seconds per 100 input sequences. Integrated Gradients is computed with 50 steps and a single all-zeros reference. DeepSHAP is computed with 10 references. Time relative to fastISM in parentheses. For BPNet models, which output a profile vector as well as a count scalar, Gradient × Input, Integrated Gradients and DeepSHAP were computed in a loop with respect to each output of the profile and the count scalar (*).