Table 2.

Comparison of fastISM with standard ISM for large models

Architecture	Layers	Input size	Output size	fast ISM	Standard ISM
Akita	11 conv+max pool, 8 dilated conv block, 7 Conv2D blocks	1 048 576	99 681×5	6463	8700
					(1.3)
Enformer	14 conv, 7 attention pool, 11 transformer + feed-forward blocks, 2 conv	196 608	896×5313	1642*	3792*
					(2.3)
	Same as above with compressed output		896×20	10 086	31 980
					(3.2)
	Same as above except with 4 transformer		896×20	4878	26 694
					(5.5)

Note: All times in seconds per 100 input sequences for ISM of central 1000 bp windows, except the Enformer model with 896×5313 outputs for which central 100 bp windows were scored (*). Time relative to fastISM in parentheses.