For visualization of a representative subset of enriched terms in Fig. 7b and Fig. 7c, we select representative terms for display (after diagonalizing the enrichment matrix by re-ordering the rows). We do this using a weighted bag-of-words approach to select highly-enriched terms that contain many words that are overrepresented in gene-set labels showing similar enrichment patterns. Briefly, sliding along the row names (gene-set terms) of the, diagonalized, enrichment matrices, we collect word counts and multiply these by integer-rounded -log10(q-values) obtained from GREAT. We do this in sliding windows of size 33 for Fig. 7b (resulting in 35 terms) and size 16 for Fig. 7c (resulting in 15 terms). For each word in a window, these values are expressed relative to the same words across all row names, registering to what extend they are over-represented. Each gene-set term in the window is then assigned a score based on the mean over-representation of all words it consists of. Lastly, gene-sets are co-ranked based on this mean over-representation and their GREAT significance. The best-ranked gene set label is selected as the representative label for that window. All terms are shown in Fig. S11d and available for download on the supplementary website.
