RBP-Tar – a searchable database for experimental RBP binding sites

Implementation

Pipeline for reproducible data download and annotation

We have developed a reproducible and easy-to-use pipeline for downloading and annotating RBP eCLIP data from the ENCODE (Luo et al. 2020) database.

Metadata for eCLIP experiments, as well as additional files containing genomic coordinates are downloaded from the ENCODE database with the following parameters:

Following the download, information about the chromosome, start position, end position and strand of reads are extracted for each RBP binding site. Binding sites are filtered by length, excluding ones shorter than 20 and longer than 100 nucleotides. As a last step, genomic sequences of binding sites are retrieved.

The described pipeline is implemented as a set of python scripts and is freely available at GitHub.

Using this pipeline, data for 167 RBPs on two cell lines (K562, HepG2) were downloaded. In total, 1.8 Gb of data, representing more than 25 million binding sites, were thus processed.

Web server

Here we present RBP-Tar, a web server that can access the above curated dataset of RBP binding sites (https://ncbr.muni.cz/RBP-Tar) and was built with Python (RRID:SCR_008394), the web development framework Flask, and a simple SQLite database (RRID:SCR_017672). The application’s source codes can be found on the project’s GitHub page, along with the requirements and instructions for the deployment if a user wants to run the application locally.

The web user interface allows searching and filtering based on the start and end position of the binding site, strand, chromosome, and protein name. It offers the download of the filtered data based on the search done by the user. Due to the size of the dataset, the view is limited to 10 000 results. However, the whole dataset can be conveniently downloaded as a gzipped CSV (579 MB) (here).

Operation

The RBP-Tar web server can take as input any of the following user-provided parameters: [Start min, Start max] denote the limits of the low genomic coordinate of the locus of interest. Similarly, [End min, End max] denote the limits of the high coordinate. [Strand] and [Chromosome] can be used to narrow down the search to only one strand and a specific chromosome. Using these combinations of parameters, a user can easily search for binding sites on their favourite gene, exon, or even a whole chromosome. The last parameter is [Protein name], which brings out a drop-down menu of all the RBPs in the database. If this parameter is not set, all RBPs are queried.

Results are shown as a table with the [Chromosome, Start, End, Strand] genomic location of the binding site, followed by the [Protein name] of the associated RBP and the [Sequence] of the binding site contains the genomic sequence of the binding site. Results can be seen online in a table format or downloaded as a CSV file with one button click. We expect most users to download the results and use them for further downstream analyses.

Use case 1: all known RBP binding sites on the gene of interest

The first and potentially most common use case would be the query of all known RBP binding sites on a gene of interest. For example, we can query our web server with the coordinates of the Fused in Sarcoma (FUS) gene (chr16: 31180139-31191605, +) and leave the protein field empty. After this search, all 5 045 known RBP binding sites on this gene are returned and can be easily downloaded in a CSV file for further analysis (Figure 1).

Figure 1. Extracting binding sites of all RNA-binding proteins on one locus (use case 1).

In fact, the gene we used encodes the RBP FUS, which plays important roles, among others, in neurodegeneration and cancer progression. We can use the filter [Protein Name] to identify the potential self-targeting of FUS on itself. Indeed, we can thus identify 72 potential FUS self-targeting binding sites identified via eCLIP (Grešová & Raček 2023) (Figure 2). Of course, the biological relevance of this type of finding is left to the users, as is the further validation of high-throughput derived RBP binding sites.

Figure 2. Fused in Sarcoma (FUS) self-targeting identified by filtering by gene location and protein name.

Use case 2: training/testing dataset for one RBP

A potential user may want to develop an RBP binding site machine learning tool that would be able to predict binding sites based on a sequence. It is important to make sure that training and testing sets for their machine learning method are not overlapping. Using our web server, they can download all binding sites for a specific RBP, for example, on chromosome 1, and use them as a training set (Figure 3). Then do the same for chromosome 2 and use it as an independent testing set, thus ensuring that training and testing set do not overlap.

Figure 3. Extracting all binding sites of a single protein on a single chromosome (use case 2).