scalable. 1000s sequences. 100s motifs highly parallel. extendable
Actually, looking at metamotifs and motif classification in this talk.
Different TF families have different modes of binding.
Can build a metamotif that encodes the distribution across a number of motifs in a family. Like a PWM, but with confidence intervals. Developed a nested sampler to learn these from a set of motifs (eg for bHLH TFs) describes repeating patterns in a set of motifs. Can use these as a prior for motif finding which improves detection of real motif in a test set. (I think…)