Figure A: For each task (i.e. cell type), we examine how the motifs discovered on the peaks by MEME, HOMER, and DiChIPMunk are recovered by TF-MoDISco. After merging the motifs discovered by TF-MoDISco on the count and profile heads, we compute the maximum similarity of each motif discovered by MEME/HOMER/DiChIPMunk with a TF-MoDISco motif, and the maximum similarity to a JASPAR motif. We find that many motifs found by MEME/HOMER/DiChIPMunk are recovered by TF-MoDISco.

Figure B: For each task, we plot all motifs discovered by MEME/HOMER/DiChIPMunk and order them by the ratio of TF-MoDISco similarity to JASPAR similarity. Motifs with a high ratio are estimated to have been recovered by TF-MoDISco, but are not in JASPAR. These motifs tend to be low-complexity. Motifs with a low ratio are estimated to correspond to a known motif, yet was not discovered by TF-MoDISco. These motifs tend to spurious and bear no relevance to the biological system at hand.

Construct plots of benchmark motif rank and TF-MoDISco similarity