We place a very strong prior on this transition that effectively fixes it and encourages the Baum-Welch algorithm to learn motifs of higher information content.
We also compare the merits of several ways of representing binding sites, and demonstrate the ability of our method to learn motifs that are associated with CaM binding.
Maticzka and colleagues developed the graph kernel-based GraphProt, and applied it to learn motifs from CLIP-seq data [ 118], producing motifs that were highly predictive of binding: the certainty of predicted motifs for PTB correlated with measured RBP affinity.
At the first step of creating the binding model, we learn the motifs in the input motif collection that are cooperative to TFexp.
The task of learning a motif from ChIP-seq data is not addressed.
Our method needs to use motifs of TFs, representing binding specificities, to identify putative binding sites in target sequences (though it is theoretically possible to learn novel motifs under our framework, similar to [19]).
Learning RBP binding motifs can be accomplished by applying DNA motif finders, which only consider the RNA sequence, or by considering RNA secondary structure either explicitly or as a layer on top of sequence preference.
After learning all significant motifs, we combine them into a single model and estimate the model parameters.
In this study, we used machine learning with DNA motifs and relevant array-based ChIP data to computationally classify cell type-specific Drosophila cardiac enhancers.
RankMotif++ [ 18] learned PWM motif models by maximum likelihood estimation of a probabilistic model for binding preferences.
