Loading…
RNA-binding protein binding profiles predicted base by base across 800 bp windows, with an m6A signal channel that exposes methylation-RBP crosstalk.
No providers recorded yet. Browse all providers
An eCLIP experiment reports where an RNA-binding protein (RBP) sat, in one cell line, on transcripts abundant enough to sequence. Most computational stand-ins answer a coarser question than the assay does — given a window of sequence, is this protein bound or not — discarding the shape of the binding profile and the positional detail that makes a peak interpretable. TransRBP predicts the binding signal at every base across an 800 bp window, and takes a second input alongside the RNA sequence: local N6-methyladenosine (m6A) signal from MeRIP-seq. That second channel is the point. m6A and RBP binding are two post-transcriptional regulatory layers acting on the same molecule, and modeling them together makes the network itself interrogable for how one shifts the other.
TransRBP was developed by Jianche Liu, Xinlu Zhu, Yang Yin and colleagues in Xushen Xiong's group at Zhejiang University, spanning the Second Affiliated Hospital of the university's School of Medicine, Liangzhu Laboratory, and the university's biomedical systems and informatics centre. The authors first screened 139 ENCODE eCLIP datasets for proteins whose binding sites overlap m6A peaks, then trained one model per protein for the 32 that ranked highest — a set including established m6A readers, writers and erasers together with splicing factors and translation regulators not previously tied to methylation. Within the catalog it sits alongside RBPSignal, which also trains per-protein regressors on ENCODE eCLIP, and m6A-FORM, which comes at the same biology from the methylation side.
The architecture is a convolution-attention hybrid. A single 1D convolution with 256 filters of width 5 and an exponential activation reads all five input channels together and extracts local motifs, two residual convolutional blocks propagate them, a 4-layer Transformer encoder with 8 heads and hidden width 256 mixes information across all 800 positions, and a decoder of four dilated residual blocks (dilations 2, 4, 8, 16) rebuilds a single-channel output track. Training used K562 eCLIP from ENCODE paired with size-matched input controls, K562 MeRIP-seq from GEO accession GSE137752, and hg38; 20,000 augmented windows per protein were split by chromosome, with an MSE loss and Adam at learning rate 1e-3. Across the 32 proteins TransRBP reaches a median Pearson correlation of 0.59 and median Spearman of 0.53 against held-out eCLIP profiles, beating RBPnet — the prior sequence-to-signal model for this task — on 29 of 32 proteins, which the authors report as a 28% accuracy improvement.
The immediate use is imputing RBP binding where eCLIP has not been run: lowly expressed transcripts, non-coding RNAs, or alleles no experiment sampled. The interpretation tools extend that into mechanism. Running the variant scorer over more than 18,000 ClinVar variants across 24 brain, heart, lung and muscle diseases yielded 1,806 variant-RBP combinations with strong predicted binding disruption, concentrated in 3' UTRs and near splice junctions, and including synonymous variants whose effect is invisible to protein-coding annotation. Six protein-disease pairs showed variant impact falling off with distance from the nearest m6A site — among them UPF1 in Alzheimer's disease and DDX3X in cardiomyopathy and muscular dystrophy — the signature of an m6A cis-acting regulatory variant, and a concrete hypothesis to test at the bench.
TransRBP demonstrates that a modification track and a sequence track can be trained together and then pulled apart by ablation, turning an accuracy gain into a statement about regulatory crosstalk: the binding motifs of these 32 proteins are enriched for splicing consensus, implicating m6A in splicing regulation through RBPs not previously counted as m6A machinery. The practical boundary is coverage — thirty-two checkpoints are published, all trained on K562 data, and any other protein means training a new model from ENCODE eCLIP BAM files. The model carries no explicit representation of RNA secondary or tertiary structure, which the authors name as the clearest direction for improvement. The work remains a preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.