Extended Data Fig. 8: SAE features reveal semantic, structural, and organizational details of eukaryotic genomes.

(a) Activations of a frameshift-associated feature in a 100 bp region following different mutation types. (b) F1, precision, and recall scores across mutation types for features shown in (a). (c) Activations for SAE features associated with exons, introns, and their boundaries in the human genome, shown for a 6000 bp region in chromosome 1. (d) F1, precision, and recall scores for each SAE feature shown in (c) to its corresponding genomic element. These scores were calculated at the level of individual bases across 1,000 genes randomly selected from the human genome. (e) Mean activations for each SAE feature shown in (c) on different annotation types across the human genome, and the corresponding AUROC values of the features to their corresponding annotation type. AUROCs were calculated at the level of individual bases across 1,000 genes randomly selected from the human genome. (f) Recall-FDR curve of Evo 2 SAE features compared with HOMER on human H1-2CORE motifs and promoter-enriched motifs.