2026-09-17
Parnet (21M) predicts base-resolution RBP binding from 223 eCLIP experiments. Frozen embeddings match or beat larger RNA LMs on translation, intron retention, and splice variants.
Most RNA language models pretrain by filling in masked bases on unlabeled sequence. That objective is enough for structure prediction and ncRNA family labels. For splicing, translation, and localization, those models usually need heavy labeled fine-tuning, because pretraining never saw which bases a protein actually reads.
RBPNet already predicted a single RNA-binding protein (RBP) CLIP footprint from sequence. It learned one experiment at a time, and those weights were not treated as a reusable foundation. This Helmholtz Munich / Francis Crick Institute preprint asks whether jointly modeling binding maps for 150 RBPs is a compact enough "RBP code" to serve as RNA embeddings.
Parnet turns RBPNet into a multi-task CNN. GENCODE transcripts are tiled into 600 nt windows with 150 nt overlap, one-hot encoded, then passed through a 1D convolution (128 filters, kernel 12) and nine residual blocks with exponentially dilated convolutions. The bottleneck splits into 223 output heads, one per ENCODE eCLIP experiment (K562 and HepG2, 150 unique RBPs). Capacity rises from RBPNet's 0.5M parameters to 21M. Training used 700,114 tiles; chromosomes 3, 8, and 15 are held out.
Each head does not emit a single binding curve. eCLIP has strong experimental background, so every head has three tracks: a control track for the size-matched input (SMI) crosslink distribution, a target track for protein-specific binding, and a total track that mixes them with coefficient π. An extra loss term λπ pushes noise onto the control track. λ was swept from 0 to 80; agreement with in vitro motifs from RNA Bind-n-Seq and RNAcompete peaked at λ=10, which is used downstream.
Most downstream tasks freeze the encoder and train a linear layer or a two-layer MLP on mean-pooled embeddings. Splice-site classification also has an end-to-end fine-tune control. Variant scoring is zero-shot: cosine distance between reference and alternate embeddings, or Jensen-Shannon divergence across the 223 binding profiles.
On held-out transcripts, Pearson r for binding profiles ranges from 0.17 to 0.70 (mean 0.48), a 35% Pearson and 19% Spearman lift over retrained single-task RBPNet. AUPRC on ENCODE NarrowPeaks rises 21% on average; on PureCLIP crosslink sites it rises 40%. Gains concentrate on weak data: the bottom quartile of RBPNet RBPs gain 40% Spearman, the top quartile only 3.6%. The paper flags the confound: 21M versus 0.5M parameters plus a longer context window, so multi-task heads are not the whole story.
Frozen embeddings plus a light classifier match or beat Nucleotide Transformer and RiNALMo on RNA-family and genomic-region classification. Dropping the first two principal components (length- and GC-correlated) cuts those two models by more than 10% in some settings; Parnet barely moves. lncRNA chromatin dissociation half-life reaches Pearson 0.33 with linear regression and 0.54 with a three-layer net, above both language models. On four Ribo-seq translation-efficiency sets, muscle is 0.82 versus RiNALMo's 0.78, PC3 is 0.73 versus 0.67, HEK293 is tied. Intron retention hits AUROC 0.90 and AUPRC 0.91 (figure top-k 0.906) against RiNALMo 0.878 and SpliceAI-400 0.863.
Ternary splice-site classification (donor / acceptor / neither) has AUROC and AUPRC above 90% for every method. Frozen Parnet is already indistinguishable from its fine-tuned version; frozen RiNALMo drops by as much as 15%. Zero-shot MutSpliceDB splice-altering variants versus nearby gnomAD controls: embedding cosine distance AUROC 0.974 and AUPRC 0.891, versus OpenSpliceAI 0.799 / 0.801. Profile-level perturbation still beats phyloP at AUPRC 0.68. On aggregated SpliceBench, Parnet ranks second (AUROC 0.782, AUPRC 0.661) behind SpliceAI (0.794 / 0.709).
Cell-type transfer is shown on STMN2: the gene is not expressed in the K562 eCLIP used for training, yet the model places TDP-43 on the cryptic exon, matching SH-SY5Y iCLIP.
| Task | Parnet | Baseline |
| Binding Pearson | mean 0.48 (+35%) | single-task RBPNet |
| Intron-retention AUPRC | 0.91 | RiNALMo 0.878 / SpliceAI 0.863 |
| MutSpliceDB AUPRC | 0.891 | OpenSpliceAI 0.801 |
| SpliceBench AUPRC | 0.661 | SpliceAI 0.709 |
| TE (muscle) | 0.82 | RiNALMo 0.78 |
This is the Borzoi / BigRNA idea moved into post-transcriptional regulation: the pretraining signal is measured RBP binding, not masked reconstruction. With 21M parameters and two cell lines, frozen embeddings already match or beat much larger RNA and DNA language models on a spread of RNA function tasks. The RBP interactome is a dense inductive bias.
Two practical uses stand out. Variant interpretation yields a score and a shortlist of disrupted RBP tracks: an RB1 donor mutation maps to PRPF8/AQR, an NF1 acceptor mutation to U2AF1/U2AF2, the latter nearly invisible to SpliceAI. For experimental hypotheses, Integrated Gradients plus TF-MoDISco on intron retention recover hnRNPH1, ELAVL, and RBM15B. Training code is public.
This is not a wholesale replacement for SpliceAI. Parnet is second on SpliceBench. Most translation-efficiency baselines see only the 5' UTR, while Parnet reads the full transcript. The backbone is a scaled CNN, not a new architecture. The actual claim is about pretraining: functional tracks can be cheaper and more interpretable than unlabeled sequence at scale.
Two hard limits are stated. Training covers about 10% of the human RBP catalog (150 of roughly 1,500), and only K562 and HepG2; proteins without CLIP data are blind spots in attribution. The model takes sequence only, no secondary structure and no m6A. Associations with modification-linked RBPs such as RBM15 are indirect footprints.
The controls are imperfect. Gains over RBPNet are entangled with capacity and window length. Optimus 5', UTR-LM, and FramePool use 5' UTRs; Parnet and RiNALMo use full transcripts. eCLIP can mix "the protein is nearby" with "the protein binds", a caveat the authors write in the variant section. This is a bioRxiv preprint. Figure and text disagree once: the body quotes MutSpliceDB AUPRC 0.801 for Parnet embeddings, the figure shows 0.891 for Parnet and 0.801 for OpenSpliceAI. The numbers above follow the figure.