Loading…
RNA foundation model pretrained on 223 eCLIP experiments to predict base-resolution RBP binding, with frozen embeddings that transfer downstream.
No providers recorded yet. Browse all providers
A masked-language objective teaches an RNA model which nucleotides tend to follow one another. It never tells the model that a U-rich stretch is where TIA1 binds, or that TIA1 bound downstream of a 5' splice site pushes an exon toward inclusion. Parnet replaces that proxy objective with a measurement: it is pretrained to predict, base by base, where each of 150 RNA-binding proteins (RBPs) crosslinks to a transcript. The combinatorial "RBP code" governing how an RNA is spliced, stabilized, localized and translated becomes the training signal itself, anchoring the representation in observed protein–RNA contacts rather than sequence statistics.
Parnet — named PanRBPNet in its codebase — comes from Annalisa Marsico's group at Helmholtz Munich, with collaborators in London, Munich, Frankfurt and Ljubljana. It is the multi-task successor to RBPNet, the same lab's single-experiment binding-profile model. Where RBPNet fits one network per CLIP-seq dataset, Parnet trains a single 21-million-parameter network end-to-end across 223 ENCODE eCLIP experiments at once, letting sparsely sampled RBPs borrow structure from related ones.
The result behaves like a foundation model at a fraction of the usual size. Frozen Parnet embeddings, the encoder never updated, support RNA biotype classification, lncRNA chromatin half-life, translational efficiency, splice-site recognition, intron retention and zero-shot variant scoring — matching or exceeding far larger self-supervised models such as RiNALMo and Nucleotide Transformer. It extends into post-transcriptional regulation the strategy Borzoi applies to chromatin-level assays, and contrasts with Orthrus, which grounds RNA representations in cross-species conservation rather than measured binding.
Parnet one-hot encodes a 600-nucleotide window, passes it through a 1D convolution (128 filters, kernel size 12) and nine residual blocks of dilated 1D convolutions (128 filters, kernel size 6) with batch normalization, ReLU and 25% dropout, then branches into 223 heads, each producing a 600-position probability distribution. That is 21M parameters against RBPNet's 0.5M, on a context window twice as wide. Training used ENCODE eCLIP for 150 RBPs in K562 and HepG2, with GENCODE v48 transcripts tiled and filtered by read depth to 700,114 windows, split by chromosome.
On held-out transcripts Parnet reaches Pearson correlations of 0.17–0.70 across RBPs (mean 0.48), improving on retrained single-task RBPNet by 35% in Pearson and 19% in Spearman correlation, with the noisiest datasets gaining most: bottom-quartile RBPs improve 40% in Spearman versus 3.6% for the top quartile. Frozen embeddings predict lncRNA chromatin half-life at Pearson 0.54, translational efficiency at 0.82 in muscle and 0.73 in PC3 cells (against RiNALMo's 0.78 and 0.67), and intron retention at 0.90 AUROC. Motif agreement with in vitro RNA Bind-n-Seq and RNAcompete data peaks at λ = 10 and degrades by λ = 80, as the target track over-shrinks toward the control.
One sequence model answers questions that normally need a separate tool each: which RBPs occupy a transcript, whether an intron is retained, how efficiently an mRNA is translated, whether a lncRNA stays chromatin-tethered, and what a non-coding variant does to splicing. Because Parnet predicts binding rather than classifying outcomes, it generalizes beyond its training context — trained on TDP-43 eCLIP in K562, it recovers TDP-43 binding at the STMN2 cryptic exon central to ALS, validated against iCLIP in SH-SY5Y cells. For variant curation it attributes the damage to identifiable spliceosomal factors rather than only ranking it.
Parnet argues that the RBP interactome is a compact, functionally sufficient basis for RNA representation learning: measuring what binds an RNA teaches a model more about how it is regulated than reconstructing its sequence does, with two orders of magnitude fewer parameters than typical RNA language models. Its embeddings also resist the global-composition shortcuts that inflate other scores: removing the top two principal components barely dents Parnet's accuracy while costing RiNALMo and Nucleotide Transformer more than 10% on some tasks. The authors state the scope limits: pretraining covers roughly 10% of the human RBP repertoire from two cell lines, and the model reads sequence alone. Code is released under Apache-2.0, and checkpoints are committed to the repository, though the paper's availability statement does not mention weights and none of the released files matches the parameter count of the published network. The work remains a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.