Discovering sequence and structure landscapes in RNA interaction motifs

Abstract

RNA molecules are able to bind proteins, DNA and other small or long RNAs using information at primary, secondary or tertiary structure level. Recent techniques that use cross-linking and immunoprecipitation of RNAs can detect these interactions and, if followed by high-throughput sequencing, molecules can be analysed to find recurrent elements shared by interactors, such as sequence and/or structure motifs. Many tools are able to find sequence motifs from lists of target RNAs, while others focus on structure using different approaches to find specific interaction elements. In this work, we make a systematic analysis of RBP-RNA and RNA-RNA datasets to better characterize the interaction landscape with information about multi-motifs on the same RNAs. To achieve this goal, we updated our BEAM algorithm to combine both sequence and structure information to create pairs of patterns that model motifs of interaction. This algorithm was applied to several RNA binding proteins and ncRNAs interactors, confirming already known motifs and discovering new ones. This landscape analysis on interaction variability reflects the diversity of target recognition and underlines that often both primary and secondary structure are involved in molecular recognition.

Figure 1.

Absolute frequency of co-localized pairs of motifs in the CLIP (left) and eCLIP (right) datasets. The majority of sequence-structural pairs are centered around 19% with extreme cases of ∼40% (orange line). More cases of co-localization on RNAs are present in the sequence-sequence pairs while the structure-structure pairs are relatively low, with a mean of 10%. eCLIP results show a generalized higher co-localization because of the lower experimental noise produced by the assay.

Figure 2.

RBM5 interaction landscape. Top: sequence and structure motifs shared by 45% of the RNA molecules. Bottom: location of sequence (left) and structure (right) motifs; the residue in position 0 corresponds to the start of the binding site reported by the experiment. Lower right: score distributions of input sequences and background sequences with respect to the motif model.

Figure 3.

TIAL1 interaction landscape. Top: sequence and structure motifs shared by 49% of the sequences. Bottom: location of sequence (left) and structure (center) motifs; the residue in position 0 corresponds to the start of the binding site reported by the experiment. Lower right: score distributions of input sequences and background sequences with respect to the motif model.

Figure 4.

Structural trend and sequence motifs for miRNA targets. Red lines correspond to PSSS mean score; the weblogo shows the sequence motif identified (top) in that specific regions compared to background (bottom). Here, position 0 corresponds to miRNA binding site as described in methods.

Figure 5.

Overview of motifs results for miRNA targets. Each spot is a sequence-structure motif pair found in one of the 383 miRNA set of targets. Ten representative pairs are shown in red circles. The size of the spot is proportional to the size of the dataset and the color represents the co-coverage, the position along x- and y- axes are respectively the reverse log₁₀P-value for structure and sequence motifs.

Figure 6.

TINCR interaction landscape. Top: sequence and structure motifs shared by 16% of the sequences. Bottom: location of sequence (left) and structure (center) motifs; the residue in position #0 corresponds to the start of the binding site reported by the experiment. Lower right: score distributions of input sequences and background sequences with respect to the motif model.

Figure 7.

XIST interaction landscape. Top: sequence and structure motifs shared by 42% of the sequences. Bottom: location of sequence (left) and structure (center) motifs; the residue in position 0 corresponds to the start of the binding site reported by the experiment. Lower right: score distributions of input sequences and background sequences with respect to the motif model.