Interpretation of the AGO-binding model learned from CLIP-seq data.

<p>(<b>A</b>) Positional distribution of AGO binding sites (blue/green) and unbound sites (grey) within 3’UTRs in CD4+ T cell (top) and HEK293 (bottom), showing enrichment of bound sites near the start of the 3’UTR (left) and in the region upstream of internal 3’ cleavage sites of multi-UTR transcripts (right). There is also enrichment of AGO-bound sites ~200nt downstream of internal 3’ cleavage sites, suggesting that the resolution of the PolyA-seq peaks can be limited and/or that clusters of nearby 3’ cleavage sites confound the analysis. All distances were between the position aligned against nucleotide 2 of the miRNA and the start/end of the corresponding 3’UTR. (<b>B</b>) RBPs with motifs that match the most discriminative <i>k</i>-mers in the common sequence model. Positions with the highest differential POIM for 6-mers upstream and downstream of the miRNA seeds were chosen, and then a signed rank test was used to assess the enrichment of POIM <i>k</i>-mers in RNAcompete array probes. False discovery rates (FDRs) were estimated using the empirical <i>p</i>-value distribution from 1,000 SVMs trained on random permutations of the +/- labels. Motif logos summarized from the original RNAcompete assays are shown for the top 5 RBPs as ranked by FDR. (The same RBP symbol may appear multiple times since in some cases several constructs of the same protein were assayed by RNAcompete.) (<b>C</b>) An example of co-binding of Pumilio and Argonaute at miRNA target sites. Two miR-17/20/106 seed matches within the 3’UTR of UBNX2A are shown, one with AGO2 binding and one without, along with the coverage profiles of AGO2 and PUM2 CLIP in HEK293. For each site, the prediction scores from the SVM sequence model are decomposed into positional scores and displayed. Sequence features near the target site including the Argonaute motif, Pumilio motif, m1A and m9U are also highlighted.