Deep neural networks for interpreting RNA binding protein target preferences

Abstract

Deep learning has become a powerful paradigm to analyze the binding sites of regulatory factors including RNA-binding proteins (RBPs), owing to its strength to learn complex features from possibly multiple sources of raw data. However, the interpretability of these models, which is crucial to improve our understanding of RBP binding preferences and functions, has not yet been investigated in significant detail. We have designed a multitask and multimodal deep neural network for characterizing in vivo RBP binding preferences. The model incorporates not only the sequence but also the region type of the binding sites as input, which helps the model to boost the prediction performance. To interpret the model, we quantified the contribution of the input features to the predictive score of each RBP. Learning across multiple RBPs at once, we are able to avoid experimental biases and to identify the RNA sequence motifs and transcript context patterns that are the most important for the predictions of each individual RBP. Our findings are consistent with known motifs and binding behaviors of RBPs and can provide new insights about the regulatory functions of RBPs.

Graph neural representational learning of RNA secondary structures for predicting RNA-protein interactions

“…Our work also demonstrates that the removal of border artefacts is crucial for an end-to-end learning system to learn non-trivial protein binding hypothesis. An interesting alternative to our de-biasing technique is the one introduced by Ghanbari and Ohler (2019), who formulate the RBP binding prediction problem as a multi-class classification, aiming to simultaneously predict the binding of all RBPs in a collection of CLIP-seq data sets, which combines multiple biased RBP dataset into one. If all datasets are equally affected by sequence biases introduced by the experimental protocol, then this bias is uninformative for the prediction task and should not significantly affect the training.…”

“…Integrated gradients (Sundararajan et al, 2017) are an effective approach to assign an "attribution score" Attr(i) to each position i of a given input sequence s, measuring the extent to which the nucleotide at that position contributes to the entire sequence's prediction score. Figure 2: Positive examples from the PAR-CLIP dataset can be biased at the beginning and at the end of the viewpoint regions, emitting an unusually high frequency of Guanine (Ghanbari and Ohler, 2019) and some correlated residuals, which can be revealed by aligning the viewpoint borders of all positive examples. Such pattern, however, does not seem to exist in the negative examples, e.g.…”

“…Finally, another contribution of our paper is to quantify the impact of and rectify a specific type of sequence bias present in certain CLIP-Seq datasets, originally identified by Ghanbari and Ohler (2019), which artificially inflated the reported prediction accuracy of several approaches published recently.…”

Interpreting Deep Neural Networks Beyond Attribution Methods: Quantifying Global Importance of Genomic Features

“…Models that exploit the latter may not necessarily generalize well, Currently, the main approach to interpret a convolutional neural network (CNN) is to visualize learned representations in the input space. In genomics, such methods include visualizing the convolutional filters (Alipanahi et al, 2015;Kelley et al, 2016;Quang & Xie, 2016;Angermueller et al, 2016;Cuperus et al, 2017;Chen et al, 2018;Ben-Bassat et al, 2018;Wang et al, 2018), attribution methods (Alipanahi et al, 2015;Zhou & Troyanskaya, 2015;Kelley et al, 2016;Shrikumar et al, 2017;Ghanbari & Ohler, 2019), and more recently in silico experiments (Koo et al, 2018;Avsec et al, 2019). These approaches can be grouped into local and global interpretability methods.…”

“…For example, gradients (from predictions to the inputs) have been employed to reveal known transcription factor (TF) binding sites when trained to predict read profiles from high-throughput sequencing datasets (Kelley et al, 2018). Integrated gradients were used to uncover motifs for RNAprotein interactions (Ghanbari & Ohler, 2019). Recently, DeepLift was used to uncover known and novel TF binding sites, including their syntax with respect to other binding sites .…”