SARNAclust: Semi-automatic detection of RNA protein binding motifs from immunoprecipitation data

Iván Dotú, Scott I. Adamson, Benjamin Coleman, Cyril Fournier, Emma Ricart-Altimiras, Eduardo Eyras, Jeffrey H. ChuangPLOS Computational Biology 2018 · DOI

TL;DR

This paper introduces SARNAclust, a novel computational method that can automatically identify RNA motifs that bind to proteins, even when these motifs depend on both sequence and structure. By applying SARNAclust to real-world data, the researchers not only discovered known motifs but also identified novel ones, which they then experimentally validated, paving the way for a deeper understanding of RNA-protein interactions.

What Problem Does This Paper Address?

RNA-protein interactions are fundamental to gene regulation, influencing processes like splicing, translation, and localization. Understanding these interactions is crucial for deciphering cellular mechanisms and disease pathologies. A key aspect of these interactions is the specific RNA sequence and structure that a protein recognizes – its "binding motif." However, identifying these motifs has been a significant challenge. Existing methods often struggle with motifs that depend on RNA structure, have limited success in distinguishing multiple motifs that a single protein might bind, and can be confounded by experimental noise inherent in techniques like CLIP-seq. This paper aims to address these limitations by providing a more robust and comprehensive approach to RNA motif discovery.

Methods

The core of this work is the SARNAclust algorithm. It's a semi-automatic, unsupervised clustering method designed to identify RNA motifs from CLIP-seq data. Here's a breakdown of its key features:

  • Graph-based Representation: SARNAclust represents RNA sequences and their predicted secondary structures as graphs. This allows it to capture both sequence and structural information simultaneously.
  • Flexible Graph Transformations: The method offers various ways to transform these RNA graphs (e.g., "complete graph," "bulge graph"), allowing users to tailor the analysis to the expected nature of the binding motif (e.g., sequence-dependent, structure-dependent, or a combination).
  • Graph Kernel for Similarity: It uses a graph kernel (specifically, the EdEN kernel, similar to those used in GraphClust and GraphProt) to compute similarity between these graph representations. This is crucial for grouping similar RNA sequences/structures.
  • Clustering: SARNAclust then employs clustering algorithms (like DBSCAN) to group the similar RNA sequences/structures, effectively identifying potential motifs.
  • Experimental Validation Pipeline: Beyond computational prediction, the authors developed a complementary experimental validation protocol using RNA Bind-N-Seq (RBNS). This involves designing RNA sequences based on predicted motifs and testing their binding affinity to the target protein.

Datasets

  • Synthetic Data: The authors generated synthetic datasets consisting of 5 distinct hypothetical RNA motifs mixed with random sequences to benchmark SARNAclust's ability to distinguish and recover motifs.
  • ENCODE eCLIP Data: They utilized publicly available eCLIP datasets for various RNA-binding proteins (RBPs) from the ENCODE project. This real-world data was used to test SARNAclust's performance on actual experimental results.
  • Known Motifs: The study also leveraged existing databases of known RNA-binding motifs (e.g., RBPDB, ATTRACT) for comparison and validation.

Key Insights and Ideas

  • Structure Matters: The study strongly emphasizes that RNA structure is often as critical as sequence for protein binding. SARNAclust's ability to integrate both is a major advantage over sequence-only methods.
  • Multiple Motifs are Common: The paper highlights that a single RBP can bind to multiple distinct RNA motifs, a phenomenon that many previous methods struggled to deconvolve. SARNAclust's clustering approach is well-suited for this.
  • Unsupervised Learning for CLIP-seq: Applying an unsupervised clustering approach to CLIP-seq data is particularly valuable because the experimental data can be noisy and the exact binding specificities are not always well-defined beforehand.
  • Experimental Validation is Key: The integration of an RBNS-based validation pipeline demonstrates the practical utility of SARNAclust, bridging the gap between computational prediction and biological confirmation.

Key Results

  • High Performance on Synthetic Data: SARNAclust achieved very high V-measure scores (over 0.95) on synthetic data, significantly outperforming existing methods like GraphClust and RNAcontext in distinguishing multiple motifs.
  • Identification of Known and Novel Motifs: When applied to ENCODE eCLIP data, SARNAclust successfully identified known motifs for proteins like SLBP and HNRNPC. Crucially, it also predicted novel motifs for several other RBPs, including AGGF1, AKAP8L, and ILF3.
  • Experimental Validation Success: The RBNS experiments confirmed the binding of predicted motifs for SLBP and ILF3. Notably, a novel, partially double-stranded UUUUUGAGA motif for ILF3 was experimentally validated, showing a strong preference for paired regions.
  • Comparison with Existing Tools: SARNAclust demonstrated superior performance compared to GraphClust and RNAcontext, especially in its ability to handle multiple motifs and incorporate structural information effectively.

Limitations and Future Work

  • Parameter Optimization: The performance of SARNAclust can be sensitive to parameter choices (e.g., graph transformation options, clustering thresholds). While the authors explored various options, finding the optimal settings for new datasets might require further investigation.
  • Multi-protein Complex Interactions: The current RBNS assay primarily tests direct binding. The paper acknowledges that indirect binding interactions mediated through multiprotein complexes might be relevant for some motifs and are not captured by this method.
  • Generalizability to All RBPs: While SARNAclust shows promise, the authors note that different RBPs have diverse binding affinities and modalities, and further studies will be needed to confirm its generality across all RBP types.
  • Future Directions: The authors envision SARNAclust as a valuable tool for discovering new motifs, probing combinatorial interactions of RBPs, and elucidating their functional importance. They anticipate its utility will grow as more eCLIP data for double-stranded binding RBPs become available.

Summary updated

BibTeX
@article{dotu2018sarnaclust,
  author = {Iván Dotú and Scott I. Adamson and Benjamin Coleman and Cyril Fournier and Emma Ricart-Altimiras and Eduardo Eyras and Jeffrey H. Chuang},
  title = {{SARNAclust: Semi-automatic detection of RNA protein binding motifs from immunoprecipitation data}},
  journal = {PLOS Computational Biology},
  year = {2018},
  doi = {10.1371/journal.pcbi.1006078},
}
Figure