RNAGenesis

Non-coding RNAs regulate gene expression, catalyze reactions, and recognize molecular partners — and they are increasingly the substrate of therapeutics, from antisense oligonucleotides and siRNAs to aptamers and CRISPR guide scaffolds. RNA foundation models built to serve that work have mostly been encoder-only language models: RNA-FM, RiNALMo, and AIDO.RNA all learn strong sequence representations, but none of them generate sequences. Design tools, conversely, are usually narrow structure-conditioned models that cannot be reused for prediction.

RNAGenesis was built to close that split. It is a roughly 1-billion-parameter model that places sequence representation, 3D structure prediction, and de novo functional design inside a single generative framework, coupling a BERT-style encoder to a latent diffusion decoder through a query-based compression module. The work comes from Princeton University together with BioMap, Zhejiang University, and Stanford University, first posted as a preprint in December 2024 and substantially expanded in a July 2025 revision that added a therapeutics benchmark and wet-lab validation.

What distinguishes RNAGenesis from other generative RNA models is that steering happens at inference time rather than through retraining. Because generation runs in a compact latent space, a reward signal — a differentiable classifier, or a non-differentiable quantity such as minimum free energy — can be applied during the denoising trajectory to bias sampling toward sequences with desired structural or functional properties.

#Key Features

• Unified encode-and-generate framework: One pretrained backbone serves representation tasks, inverse folding, tertiary structure prediction, and sequence generation, rather than requiring a separate model per capability.
• Query-based latent compression: A Query Transformer maps variable-length encoder outputs into fixed-length latent vectors, giving the diffusion model a compact and well-behaved space to operate in.
• Inference-time alignment: Gradient guidance nudges latents toward a differentiable reward at each denoising step, while beam search explores diverse high-reward denoising paths for rewards that offer no gradient.
• Pluggable structure module: Pairing the encoder with an Invariant Point Attention transformer extends the model to RNA 3D structure prediction and de novo structure design.
• Therapeutics benchmark: The authors introduce RNATx-Bench, over 100,000 experimentally validated datapoints spanning ASO, siRNA, shRNA, circRNA, aptamer, and human UTR variant assays.

#Technical Details

The encoder is a 32-layer transformer with hybrid N-gram tokenization and 1D convolutional layers, combining single-nucleotide resolution with multi-scale context. It was pretrained by masked language modeling at 30% masking on RNAcentral v24.0 — 42 million sequences reduced to 6.4 million cluster centers by MMseqs2 clustering at 90% sequence identity, covering lncRNA, rRNA, tRNA, miRNA, snoRNA, and ribozymes. A Query Transformer adapted from the ESM-2 architecture bridges the encoder to a causal decoder based on ProGen2-small, which autoregressively reconstructs sequences from the latent soft prompts; a diffusion model then learns to denoise those latents from Gaussian noise.

On the BEACON benchmark, RNAGenesis ranks first on 11 of 13 tasks. In inverse folding it reaches 57.1% sequence recovery at 16.9 Å RMSD, ahead of RhoDesign (52.9%, 17.1 Å) and gRNAde. For tertiary structure, the encoder is paired with the IPA transformer from RNA-FrameFlow and fine-tuned on RNASolo; on the ZTP riboswitch it reaches 3.12 Å RMSD and a TM-score of 0.552 versus 3.79 Å for RhoFold+, and it is the fastest method among those compared while placing second overall on RMSD.

#Applications

RNAGenesis targets RNA therapeutic discovery end to end: ranking ASO, siRNA, and shRNA candidates, scoring UTR variants, and generating novel functional RNAs. In wet-lab validation, designed aptamers against IGFBP3 bound with dissociation constants of 4.02 nM and 6.06 nM, compared with 11.6 nM for a SELEX-derived control. For genome editing, 6 of 24 designed sgRNA scaffolds outperformed wild type in a GFP reporter assay, with up to 2-fold higher knockout at the AAVS1 and B2M loci; the RGen-6 scaffold raised cytosine base editing efficiency from 5.5% to 14.8% and improved prime editing by up to 1.2-fold.

#Impact

RNAGenesis is among the first RNA foundation models to carry a single pretrained backbone through representation, structure, and generation, and to close the loop with experimental validation of both aptamers and editing scaffolds. Its inference-time alignment strategy is notable beyond RNA: steering a frozen latent diffusion model with an external reward avoids retraining for each new design objective. Practical caveats remain. The work is a preprint and has not completed peer review; the wet-lab evidence rests on a small number of designs at a handful of loci; and while the code is MIT licensed, the released weights carry a non-commercial license and gated access, which limits commercial use.