Extended Data Figure 1. Additional comparisons of K-Lasso LIME to SQUID.

Shown are the results of analyses, performed as in Fig. 2b for $n=50$ genomic sequences, comparing the performance of SQUID to the performance of the K-Lasso implementation of LIME for four different values of $K$. P-values were computed using a one-sided Mann-Whitney U test; ***, p < 0.001. We note that the attribution variation values obtained for SQUID in these tests varied systematically with the choice of $K$. The reason is as follows. The K-Lasso LIME algorithm produces sparse attribution maps that have only $K$ nonzero parameters. Consequently, the variation observed in K-Lasso LIME attribution maps systematically decreases as $K$ decreases. This gives K-Lasso LIME an unfair advantage in the attribution variation test described in Main Text and in Methods. To fairly compare K-Lasso LIME to SQUID in this figure, we therefore modified this test. In the analysis of each in silico MAVE, the attribution map elements inferred by SQUID were first set to zero at the same positions where all K-Lasso LIME attribution map elements were exactly zero. Attribution variation values were then calculated as described in Main Text and in Methods.