RPI-Zero benchmark: RNA-protein interaction data from CLIP peaks with unseen-RNA, unseen-protein and unseen-pair splits

This record contains the data of RPI-Zero, a benchmark for predicting RNA-protein interactions on RNAs and proteins that are absent from the training data. The code that builds the benchmark, trains the model and evaluates it is available at https://github.com/ciceklab/RPI-Zero.

The benchmark is built from 11.9 million CLIP peaks of 209 human RNA-binding proteins, collected from CLIPdb, ENCODE eCLIP, POSTAR2 and POSTAR3. For each gene with a peak, one spliced transcript is chosen from GENCODE release 50 (hg38), which gives 29,989 transcripts. The peaks that lie fully inside one exon are labelled on the transcripts at nucleotide resolution. A transcript too long for the model input is cut into overlapping parts on a fixed grid. An example is a part paired with a protein, and it is positive if the part holds at least one whole peak bound by the protein.

Transcripts are clustered with MMseqs2 at 80% sequence identity and proteins at 30%, and whole clusters are assigned to the training, validation and test sets. The proteins are split into 161 training, 24 validation and 24 test proteins. The validation and test sets each have three slices, for unseen RNAs, unseen proteins and unseen pairs. There are two kinds of negative examples. The first are parts of a transcript the protein binds that hold no peak of the protein. The second are parts of transcripts the protein does not bind; these are matched to the protein's positive examples in RNA length and in the number of other proteins that bind the transcript. The benchmark has 2,074,753 examples, of which 971,450 are positive.

Files:

- rpi_zero_benchmark.tar.gz: the final benchmark, ready for training and evaluation. It holds one parquet file per slice with the nucleotide labels as intervals, the transcript sequences and the protein sequences.
- rpi_zero_splits.tar.gz: the inputs of the benchmark build. These are the CLIP peaks split into seven files, the protein split and the transcript clusters. From these files, the build code reproduces the final benchmark byte for byte.
- clip_peaks_raw.tar.gz: the CLIP peak collection the pipeline starts from, with the peak coordinates, the peak sequences and the protein sequences.
- README.md: the content and format of each file and the column descriptions.

Each archive extracts into a data/ directory. Extracting it in the root of the code repository places the files where the scripts expect them. The hg38 genome and the GENCODE annotation are not included; they can be downloaded from UCSC and GENCODE.