Fig. 3

Comparison of the alternative strategies for defining cis-regulatory evidence. a Histogram of test LLRs derived from training-validation-test cross-validation for up- and down-analysis using DiffMark strategy. b, c Empirical CDFs of test LLR for different strategies (shown in different colors) suggest that DiffMark strategy performs better than alternatives in down- (b) as well as up-analysis (c). d, e Frequency of training-validation-test partitions (out of 100) where DiffMark results in a greater (blue) or lower (red) test LLR than an alternative strategy. f, g Histograms of LLR values on the entire dataset for three schemes designed to test the explanatory power of cis-regulatory evidences based on TF ChIP-seq data from a CRC (HCT116) cell line. Results are for down- (f) and up-analysis (g) using the DiffMark strategy. In “shuffled” scheme (blue), the model was trained using permuted evidence. In “K562” (red), ChIP-seq data from the K562 cell line, representing binding profiles of 20 randomly selected TFs, were used to generate the DiffMark evidence. The “K562-distinct” scheme (green) is similar to “K562,” except that the 20 TF profiles were randomly selected from the 90 ChIP-seq profiles most dissimilar to the 20 HCT116 TFs. Colored dashed lines represent means of respective distributions. The LLR of the analysis performed using the 20 available ChIP-seq profiles from the CRC cell line (black dashed line) is significantly larger than the average of “shuffled” and “K562-distinct” schemes in both down- and up-analyses. h, i Statistical assessment of the contribution of each CRC TF in down- (h) and up-analysis (i), respectively. Each point represents one TF. The y-axis represents the average LLR of 100 models, each trained using the TF and 19 randomly selected K562-distinct TFs. The x-axis represents the frequency with which the TF is ranked as the most significant contributor among the 20 TFs in these models