RBPSignal

RBPSignal predicts the binding signal of RNA-binding proteins along an RNA sequence and recovers the sequence motifs that drive it. RNA-binding proteins govern splicing, localisation, stability, and translation, and mapping where they engage a transcript is central to understanding post-transcriptional regulation. Experimental crosslinking assays such as eCLIP deliver that map, but only for the proteins, cell lines, and conditions actually assayed — leaving most transcripts and most contexts uncovered.

The model was developed by Xiaohan Sally Ding, Zidong Alex Shu, Xiaoyong Pan, Hong-Bin Shen, and colleagues at the Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, and posted as a preprint in May 2025. It differs from most binding-site predictors in what it predicts. Rather than classifying a window as bound or unbound, RBPSignal regresses the continuous binding intensity that eCLIP reports, so its output tracks the strength of the interaction rather than its mere presence.

That framing pairs naturally with the model's second contribution: interpretability. Integrated gradients attribute each prediction back to individual nucleotides, letting the authors read out motif syntax — not just which k-mers a protein prefers, but how they are arranged and combined. The recovered motifs are compared against an established motif database, and the resulting binding profiles are used to examine protein-protein interaction structure among RNA-binding proteins. RBPSignal is a distinct approach from the cataloged MuSIC, which attacks the same problem by transferring cross-species conservation into label smoothing.

#Key Features

• Continuous signal regression: The network outputs a scalar binding intensity for each 101-nucleotide window rather than a binary label, preserving the quantitative structure of the underlying eCLIP measurement.
• Integrated-gradients interpretability: Per-nucleotide attributions expose the sequence patterns behind each prediction, supporting analysis of motif syntax and comparison against known motif catalogues.
• Per-protein checkpoints from a shared architecture: One architecture is trained separately for each RNA-binding protein and cell line, keyed to the corresponding ENCODE eCLIP experiment, so predictions are specific to a named protein rather than pooled across proteins.
• Chromosome-held-out validation: Performance is assessed on held-out chromosome data, a stricter split than random windows because it prevents near-duplicate sequences from leaking between training and test.
• Free hosted prediction: A public web server accepts user RNA sequences and returns binding-signal predictions without local installation or model training.

#Technical Details

The released implementation is a compact convolutional-recurrent network operating on encoded 101-nucleotide RNA windows. A residual convolutional block applies two 1D convolutions with kernel size 7, expanding to 16 and then 32 channels with batch normalisation and ReLU, and adds a 1×1 projected shortcut from the input. The resulting feature sequence feeds a single-layer recurrent unit with 64 hidden units, whose outputs are averaged across positions and passed through a three-layer fully connected head of widths 128, 64, and 1, with dropout at 0.4. The whole model is small by contemporary standards — the released example checkpoint is under 200 KB — which is what makes per-protein training across a large eCLIP compendium tractable.

Training data are ENCODE eCLIP peak files in BED format, one per protein and cell line. The repository ships one example dataset (CPSF6 in K562) and one example checkpoint, with the remaining data available from the corresponding author and the remaining checkpoints served through the web server. The preprint reports improved predictive accuracy over prior approaches and validates the discovered motifs against an existing motif database; it has not yet completed peer review.

#Applications

The practical use is extending a finite experimental resource. A researcher studying a transcript that was not covered by an eCLIP experiment — a lowly expressed isoform, a lncRNA, a variant allele — can score it against a trained protein of interest and get a positional binding profile. The interpretability layer supports a second use: dissecting which sequence elements a protein actually reads, which informs the design of reporter constructs, minimal binding elements, and mutations intended to disrupt a specific interaction. Because the web server requires no installation, both are accessible to wet-lab groups without a computational pipeline.

#Impact

RBPSignal contributes a quantitative, interpretable alternative to binary binding-site classification, and its use of integrated gradients to characterise motif syntax rather than isolated k-mers is a step toward mechanistic rather than merely predictive models of RNA-protein recognition. Its limitations follow from its design: models are protein- and cell-line-specific rather than generalising to unassayed proteins, predictions inherit whatever biases the underlying eCLIP experiments carry, and the model reads sequence alone without RNA structure or cellular context. The public repository contains the model definition, one example dataset, and one example checkpoint, but carries no license file.