RBPsuite: RNA

RNA-binding proteins (RBPs) are involved in many biological processes, their binding sites on RNAs can give insights into mechanisms behind diseases involving RBPs [1]. Thus, how to identify the RBP binding sites on RNAs is very crucial for follow-up analysis, like the impact of mutations on binding sites. With high-throughput sequencing developing, there is an explosion in the amount of experimentally verified RBP binding sites, e.g. eCLIP [2] in ENCODE [3]. However, these CLIP-seq data still cannot provide the full view of the RBP binding landscape, it is because CLIP-seq relies on gene expression which can be highly variable between experiments. But these big data can serve as training data for machine learning models to predict missing RBP binding sites that may not be detected in some experiments. For example, GraphProt encodes a RNA sequence and structure in a graph [4], which is fed into a support vector machine to classify RBP bound sites from unbound sites. GraphProt can detect the binding sequence and structure preference of RBPs and further predict the RBP binding sites on any input RNAs. Considering that RBPs have difference binding preferences, the machine leaning-based methods train RBP-specific models; each model is trained per RBP.

Recently, deep learning-based methods have achieved remarkable results on predicting RBP sites [5, 6]. For example, DeepBind is the first method to train a convolutional neural network (CNN) [7] to predicting RBP binding preference [6]. Inspired by DeepBind, iDeep integrates multiple sources of features to predict RBP binding sites using a multi-modal deep learning, which consists of a CNN and multiple deep belief networks [8]. RBPs bind to RNAs by recognizing both the sequence and structure context. Thus, iDeepS trains a hybrid network with two CNNs and a long-short temporary memory (LSTM) network [9] to infer binding sequences and structure preferences of RBPs [10]. In iDeepS, two CNNs handle the sequence input and structure inputs, respectively and the LSTM learns the dependency between sequences and structures to improve prediction performance. Different from iDeepS, pysster encodes the sequence and structure in a one-hot encoded matrix based on an extended alphabet, which combines the sequence and structure alphabet [11]. DeepCLIP applies a similar network architecture consisting of a hybrid CNN and LSTM to predict RBP binding sites on RNAs [12] and the network architecture is similar to iDeepS. iDeepE trains a local CNN and a global CNN to predict RBP binding sites from sequences alone [13]. The binding mechanism of RBP binding circular RNAs (circRNAs) is different from that of linear RNAs, and thus the trained models on RBP binding linear RNAs cannot generalize well to circRNAs, CRIP is specially developed for predicting RBP binding sites on circRNAs by using a codon-based encoding schema and hybrid deep models [14].

There exist several online webservers for RNA-protein interaction prediction based on traditional machine learning models, e.g. omiXcore [15] and SMARTIV [16, 17]. omiXcore is an RBP-general method, which trains a non-linear algorithm on pooled RNA-protein interactions and accepts the proteins and large RNAs with a size between 500 and 20,000 as inputs. Considering that different RBPs have different binding specificities, the RBP-specific method in general is superior to the RBP-general method, as demonstrated in [13]. SMARTIV accepts a set of RNA sequences in BED format file as the input, and applies Hidden Markov Model (HMM) to find the enriched combined sequence and structure motifs from in vivo binding data. In addition, SMARTIV cannot predict RBP binding sites for a single RNA sequence. The backend predictor of the above webservers are non-deep learning-based methods, which are proved to be inferior to deep learning-based methods for predicting RBP binding sites [18]. Moreover, no online webserver is currently available for predicting RBP binding sites on circRNAs.

However, to date, there is no online webserver available for predicting RBP binding sites on both linear and circular RNAs using deep learning. Most published approaches for predicting RBP binding sites only provide source code with different input data format, like GraphProt, our developed iDeepS and CRIP, their dependency is difficult to configure due to frequent update of deep learning framework, like TensorFlow. In addition, for deep learning-based approaches, the training of models is very time-intensive and computationally intensive. Thus, it is imperative to develop an easy-to-use webserver to integrate the state-of-the-art prediction methods for predicting RBP binding sites on RNAs and cover as many RBPs as possible. RBPsuite holds a broad application potential, it can be used to expand our knowledge about RBP binding RNAs, e.g. identifying interactions between RNA regions of SARS-COV-2 and human proteins. In addition, RBPsuite may be used to investigate the effect of mutations on RNA-protein binding sites, we can use RBPsuite to predict binding scores for an RNA sequence and a mutated RNA sequence, then check whether the mutation will greatly decrease the binding score to determine the effect of this mutation.

We implement an online webserver RBPsuite for predicting RBP binding sites on full-length linear and circular RNAs from sequences alone. For the linear RNAs, the server predicts the RBP binding scores using our updated iDeepS, which is retrained on binding RNA targets of 154 RBPs derived from ENCODE. For circRNAs, RBPsuite predicts the RBP binding scores using our developed CRIP. RBPsuite first breaks a full-length input sequence into multiple segments of 101 nucleotides without overlap, then outputs the scores between the segments and the chosen RBP. RBPsuite further detects the verified motifs on the predicted binding segments and visualizes the score distribution within the input sequence.

We downloaded peaks of 154 RBPs of K526 and HepG2 through eCLIP-seq from ENCODE corresponding to human genome hg19 version. These narrow peaks were produced by the eCLIP-seq Processing Pipeline v2.0 of ENCODE [19]. To prepare the positive and negative RBP binding training data sets, several steps were processed. 1) We merge the peaks files of one RBP. It should be noted that some studies [20] used the intersection of the bed files to obtain a set of most probably peaks. 2) We select regions overlapped with reference gene by intersectBed of bedtools [21]. 3) The gene overlapped regions are extended to 101 nts in upstream and downstream centering at the read peaks, and we got the positive regions of RBPs. 4) Negative RBP binding regions were produced by implementing shuffleBed of bedtools, these negative sites are those regions without any peak located from the same gene of each peak. 5) The fasta files of positive and negative regions were retrieved by fastaFromBed of bedtools. To save the training time, for each RBP, we only keep 60,000 positive sites and 60,000 negative sites if the extracted positive and negative samples are more than 60,000, respectively. Otherwise we use all the extracted samples for this RBP.

For circRNAs, we use the trained models of 37 RBPs on the benchmark dataset of CRIP [14]. For each RBP, the number of training circRNAs (bound and non-bound) is different, they range from 992 to 40,000. Each circRNA is also a sequence segment of a size 101. More details are given in Table 1. All the collected benchmark datasets for linear and circular RNAs are freely available at http://www.csbio.sjtu.edu.cn/bioinf/RBPsuite/.

In addition, we downloaded verified motifs of RBPs from CISBP-RNA [22]. In total, we obtain verified motifs for 43 RBPs, which are further scanned against the sequence segments using FIMO in MEME suite [23] with p-value  0.5 by RBPuite, the p-value threshold 0.01 is used and other parameters are defaulted values.

iDeepS and CRIP in RBPsuite are implemented under the TensorFlow framework in Python. Given a full-length RNA sequence, it will break the sequence into multiple segments of 101 nts (used by iONMF [27] and our previous iDeep) without overlap, if the input sequence or the remaining sequence is shorter than 101 nt, we pad it to a length of 101 using ‘N’ as another 101 nt-long segment. Then these generated segments are fed into the iDeepS and CRIP to give the binding scores between individual segments and a specified RBP.

The frontend of RBPsuite webserver uses JQuery framework of JavaScript and Ajax technology to implement asynchronous loading. The backend uses PHP to call shell and python scripts. For the visualization, RBPsuite directly uses Matplotlib to display the results.