GraphProt: modeling binding preferences of RNA-binding proteins

Daniel Maticzka,Fabrizio Costa,Sita J Lange,Rolf Backofen

doi:10.1186/gb-2014-15-1-r17

Abstract

We present GraphProt, a computational framework for learning sequence- and structure-binding preferences of RNA-binding proteins (RBPs) from high-throughput experimental data. We benchmark GraphProt, demonstrating that the modeled binding preferences conform to the literature, and showcase the biological relevance and two applications of GraphProt models. First, estimated binding affinities correlate with experimental measurements. Second, predicted Ago2 targets display higher levels of expression upon Ago2 knockdown, whereas control targets do not. Computational binding models, such as those provided by GraphProt, are essential for predicting RBP binding sites and affinities in all tissues. GraphProt is freely available at http://www.bioinf.uni-freiburg.de/Software/GraphProt.

Highlights

Recent studies have revealed that hundreds of RNAbinding proteins (RBPs) regulate a plethora of posttranscriptional processes in human cells [1,2,3]
We proved the capacity of GraphProt to predict RBP target sites reliably and even to detect sites missed by experimental high-throughput methods
GraphProt is an accurate method for elucidating binding preferences of RBPs and it is highly flexible in its range of application

Summary

Introduction

Recent studies have revealed that hundreds of RNAbinding proteins (RBPs) regulate a plethora of posttranscriptional processes in human cells [1,2,3]. Despite the great success of these methods, there are still some problems to overcome: (1) the data may contain many false positives due to inherent noise [7,8]; (2) a large number of binding sites remain unidentified (a high falsenegative rate), because CLIP-seq is sensitive to expression levels and is both time and tissue dependent [9] and (3) limited mappability [10] and mapping difficulties at splice sites lead to further false negatives, even on highly expressed mRNAs. To analyze the interaction network of the RBPome and to find all binding sites of a specific RBP, a CLIP-seq experiment is only the initial step. Peak detection leads to high-fidelity binding sites; it again increases the number of false negatives.

Methods

Results

Conclusion