Table 1.

DeepGuide ablation analysis.

Row	Training	Layer	Pearson, r
			Cas12a	Cas9
1	Random weights	Encoder⇨flatten7	0.070	0.003
2	Back prop-flatten7	Encoder⇨flatten7	0.455	0.312
3	Pretrained + back prop flatten7⇨fc8–10⇨mult11	Encoder⇨flatten7	0.532	0.353
4		Encoder⇨flatten7⇨fc8	0.534	0.310
5		Encoder⇨flatten7⇨fc8⇨fc9	0.517	0.291
6		Encoder⇨flatten7⇨fc8⇨fc9⇨fc10	0.514	0.305
7		Encoder⇨flatten7⇨fc8⇨fc9⇨fc10⇨mult11	0.514	0.388
8	Pretrained + back prop-all	Encoder⇨flatten7	0.641	0.409
9		Encoder⇨flatten7⇨fc8	0.658	0.424
10		Encoder⇨flatten7⇨fc8⇨fc9	0.664	0.414
11		Encoder⇨flatten7⇨fc8⇨fc9⇨fc10	0.664	0.414
12		Encoder⇨flatten7⇨fc8⇨fc9⇨fc10⇨mult11	0.664	0.501

Row 1 shows the performance of the encoder (followed by a flatten layer) using random weights (no pre-training or back-propagation); row 2 shows the performance of the encoder (followed by a flatten layer) using random weights and then performing back-propagation only on the flatten layer; rows 3–7 show the performance after pre-training the encoder and then running back-propagation only layers downstream of the encoder; rows 8–12 show the performance after pre-training and then running back-propagation on the whole network (including the encoder); correlation coefficients in bold corresponds to the best performance.

fc fully connected layer, flatten flatten layer, mult multiplication layer (see Supplementary Table 3 for the list of layers).