RNA foundation models

To design RNA,
first understand it.

RiboLoop is a family of foundation models that has learned the language of RNA from more than 47 million natural sequences. It is the understanding layer for the next generation of RNA medicines.

See the benchmarks Our roadmap

Illustration of a chain of nucleotides folding into a functional RNA molecule

#1: at predicting which mutations break an RNA, across 31 laboratory experiments
48×: smaller than the strongest competing model, and more accurate
8/8: biology task families where a RiboLoop model ranks first
2.8×: faster than the next-fastest RNA model we tested

The idea

Every RNA medicine starts as a sequence someone had to design.

mRNA vaccines, gene-silencing drugs and RNA-targeting small molecules all depend on one thing: knowing how a string of A, C, G and U will fold, function and fail inside a cell.

The space of possibilities is enormous. A 50-letter RNA already has more possible sequences than there are stars in the observable universe. Labs can test thousands. Design needs a model that already understands.

Illustration of a folded RNA being scanned, its base pairs lighting up

It learns RNA the way people learn a language: by filling in the blanks.

We hide letters in millions of natural RNAs and train RiboLoop to restore them. To succeed, it has to learn which letters pair, what shapes they form, and which changes nature never tolerates. Nobody labels any of it.

restored by RiboLoop

One compact block, used again and again.

Instead of stacking hundreds of different layers, RiboLoop re-reads each sequence with a single compact block. It reaches the understanding of models many times its size while staying small, fast and inexpensive to run.

Benchmarks

How RiboLoop compares

Public benchmarks, run against the leading RNA foundation models. RiboLoop models are shown in coral.

Accuracy vs. model size

Mutation-effect score against parameters (log scale). Top-left is best: accurate and small.

Accuracy vs. speed

Mutation-effect score against RNAs read per second (log scale). Top-right is best: accurate and fast.

Methodology and caveats

Mutation effect. RNAGym non-coding RNA set: 31 deep-mutational-scanning experiments (26 ribozyme, 3 tRNA, 2 aptamer), 856,628 measured variants. Zero-shot, wild-type-marginal scoring. Score is the absolute Spearman correlation averaged within each RNA type, then across types, ×100. All models scored by us with the same protocol. Evo 2 is excluded because its published score uses a different scorer.
Structure. bpRNA test set (TS0), median per-RNA base-pair F1 ×100. ERNIE-RNA, RiNALMo, RNA-FM, RNA-MSM, RNAErnie and RNABERT use their published scores; all other models were trained and scored by us.
Biology tasks. RNA-Scope, 8 task families (42 endpoints). Each cell is the model's mean rank across the family's endpoints (1 = best); the best of three standard prediction heads is used per endpoint. Baselines are published multi-seed means.
Speed. Full forward pass over 1,022-nucleotide RNAs, batch 16, bfloat16, one NVIDIA RTX PRO 6000; median of five windows.
RiboLoop settings. RiboLoop can re-read a sequence a chosen number of times. Scores use the best observed setting (RiboLoop-10M: 12 passes; RiboLoop-100M: 8 passes on mutation effect and speed, 12 on structure and tasks). The point marked "fast" is RiboLoop-10M at 8 passes. RiboLoop results are single training runs.

Evaluation data as of October 2026.

A real molecule

It reads the fold from the letters alone.

Crystal structure of the twister ribozyme with the base pairs RiboLoop inferred drawn as glowing rods — stem pairs found pseudoknot pairs found

17/19

base pairs of the twister ribozyme recovered from its sequence alone, including all 7 pseudoknot pairs, the hardest kind to predict. RiboLoop never saw this structure.

Shown on the experimentally solved crystal structure (PDB 4OJI). Model: RiboLoop-33M, zero-shot.

It spots a fatal typo, and suggests the fix.

The natural ribozyme

A self-cutting RNA. In the lab it is highly active.

0.82

measured activity
One letter changed

RiboLoop ranks this change the 2nd most damaging of all 162 possible single-letter changes. Its predicted activity falls from 0.97 to 0.17.

0.06

measured activity
The suggested fix

After the change, RiboLoop expects the partner letter to flip from G to C (probability 0.00001 → 0.71), restoring the pair. That double change brings activity back in the lab.

0.62

measured activity

Laboratory measurements: Kobori & Yokobayashi, Angew. Chem. Int. Ed. 2016. Predicted activity is a calibration of the zero-shot score fitted on the other 143 measured single changes.

Applications

What understanding unlocks

With a small amount of task data, the same model leads on problems that matter for RNA medicine.

Illustration of a ribosome reading an mRNA strand

mRNA medicines and vaccines

Reads the signals that decide where an mRNA goes in the cell, how it is processed, and whether it codes for protein.

#1 mRNA localisation · coding potential · polyadenylation

Illustration of an intron being spliced out of an RNA

Splicing-correcting therapies

Predicts how cells cut and join RNA, the mechanism behind approved drugs for spinal muscular atrophy.

#1 splice sites · splicing events

Illustration of a small-molecule drug bound in a folded RNA pocket

RNA-targeting small molecules

Maps how an RNA folds, revealing the pockets where small-molecule drugs can bind.

#1 secondary structure · 3D contacts

Illustration of a protein cradling an RNA hairpin

RNA–protein interactions

Predicts which proteins latch onto an RNA, a key control point for RNA stability and many diseases.

#1 protein binding across 22 RNA-binding proteins

Illustration of an RNA switch lighting up when it binds its target

Ribozymes, aptamers and biosensors

Scores every possible mutation before anyone makes it, so labs only build the designs most likely to work.

#1 zero-shot mutation effect (RNAGym)

Illustration of many RNA molecules sorting themselves into clusters

RNA diagnostics

Recognises what an unknown RNA is and where in the cell it belongs, from sequence alone.

#1 RNA type · miRNA localisation

Roadmap

From understanding, to design, to medicine

01 Now Built

Understand RNA

A family of RiboLoop foundation models, from 10 million to 160 million parameters, trained on more than 47 million natural RNAs. Leading results on mutation effect, structure and all 8 biology task families.
02 Next In development

Design and redesign RNA

Generate new RNAs with a chosen function from scratch, and improve existing ones: more stable, better expressed, tighter binding. The model's understanding becomes the engine that scores and steers every design.
03 Long term Vision

Design mRNA cancer vaccines

From the mutations in a patient's tumour to a personalised mRNA vaccine: choose the right targets, then design an mRNA that is stable, highly expressed and safe.