Extended Data Fig. 2: Influence of mutation rate and library size on SQUID attribution maps.
From: Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models
a,b, attribution variation (left) and average attribution maps (right) found for 50 SQUID attribution maps computed for the TF AP-1 as in Fig. 2 but using in silico MAVE libraries having (a) variable mutation rate (r) and fixed number of mutagenized sequences N = 100, 000, or (b) fixed mutation rate r = 10% and variable number of mutagenized sequences (N). All SQUID attribution maps were computed using additive models with GE nonlinearities, followed by cropping these maps using a flank size of nf = 50 nt.
