Protein-conditional RNA design model that generates binding RNA sequences for any target protein, with no post-generation optimization step.
No providers recorded yet. Browse all providers
Only about 15% of human proteins present a pocket a small molecule can occupy, which leaves most of the proteome nominally undruggable. RNA offers a different mode of attack: it can wrap a protein surface, carries no genotoxicity risk, and is comparatively cheap to synthesize. The obstacle has always been design. The standard route is SELEX — build a large random library, then run iterative in vitro selection cycles — while computational alternatives generate RNA unconditionally and then optimize the output against a binding-affinity predictor trained for one specific target.
RNAtranslator removes that second stage. It treats protein-conditional RNA design as a translation problem: the protein sequence is the source language, a binding RNA sequence is the target language, and an encoder-decoder transformer learns the mapping from millions of observed protein-RNA interactions. At inference the model takes a protein sequence and samples RNA sequences directly, with no affinity predictor in the loop and no fine-tuning on the target's known binders. That matters because those per-target resources mostly do not exist: DeepBind models cover 732 proteins and DeepCLIP models 221, against a proteome of tens of thousands.
The model comes from the Çiçek lab at Bilkent University, with a Carnegie Mellon co-author, and was posted as a preprint in March 2025.
RNAtranslator implements the T5 encoder-decoder architecture with 6 encoder-decoder layers and 12 attention heads, totaling 41.4M parameters. Protein and RNA are tokenized separately with byte-pair encoding over a 1,000-token vocabulary, and sequences are padded or truncated to 1,024 tokens. Training runs in two stages: pretraining on approximately 26 million protein-RNA interactions from RNAInter, which mixes experimental and computationally predicted pairs, followed by fine-tuning on 12 million high-confidence interactions sampled from CLIPdb, augmented with the roughly 30,000 non-redundant complexes in PRI30K derived from the PDB. Because CLIPdb interaction counts per protein span four orders of magnitude — 85 for EZH2 against 1,079,145 for ELAVL1 — fine-tuning uses cluster-weighted oversampling to balance protein representation without collapsing RNA diversity. Pretraining ran for 350,000 iterations on two TITAN RTX GPUs with fully sharded data parallelism; fine-tuning added 75,000 iterations on three L40S GPUs. Generation uses top-k sampling with k = 30 and temperature 1.5.
Against RBM5 and ELAVL1, designs scored higher predicted binding affinity under DeepCLIP than those from GenerRNA and RNAGEN, and HDOCKlite docking of the top candidate gave a score distribution for RBM5 whose mean falls below that of experimentally validated natural binders. OpenMM molecular dynamics showed more protein-RNA contacts and lower binding energies than natural or random natural RNAs for both targets. Novelty results are mixed relative to baselines: GenerRNA overfits after target fine-tuning and produces far fewer novel sequences, while RNAGEN — which conditions only through a third-party affinity predictor — reaches comparable novelty for RBM5 and higher for ELAVL1. Designs for PRPF8, absent from training, and PRP4K, which has no known RNA interactions in the source databases, retained high predicted affinity; design structures were predicted with RhoFold+ before docking.
The immediate use is generating starting libraries for RNA aptamer campaigns, replacing a random SELEX pool with a target-biased one and shortening the selection cycle. Because conditioning needs only a protein sequence, the model reaches targets previously out of scope for computational design — understudied proteins, disordered regions without a ligand pocket, and engineered proteins. Synthetic biology groups can use the same interface to design RNA elements binding a chosen regulatory protein.
RNAtranslator reframes conditional biomolecule design as sequence-to-sequence translation, a formulation that generalizes past RNA and is the paper's most portable contribution. Its practical significance is decoupling design from per-target supervision, which confined earlier protein-conditional methods to a few hundred well-characterized proteins. The evidence base is entirely computational — predictors, docking, and simulation, with no wet-lab validation of any design — and the work remains an unreviewed preprint. The code carries no LICENSE file, and the terms declared in the repository and model card — CC BY-NC-SA 2.0, academic use only — restrict commercial application.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.