AI for Bio · Model

RIBOSPAN

1.61B-parameter bidirectional RNA foundation model with a native 10,240-nt context and single-nucleotide tokens, able to take a full transcript at once

Genomics & RNAAcademic labChinaActive中文

Overview

NameRIBOSPAN · aliases: RIBOSPAN, RIBOSPAN-FM
Full nameThe name is a model name; the paper gives no expanded form
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single model whose four checkpoints are configuration axes inside one record
OrganizationShanghai Innovation Institute, CAS Center for Excellence in Molecular Cell Science, Shanghai Jiao Tong University
Organization typeAcademic lab
CountryChina
Model classRNA foundation model (representation)
DomainGenomics & RNA (also: Molecule generation)
First release · preprint
Basis: Preprint v1 source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Code release: † · Public code repository source
  • Weights release: † · Model hub collection updated: short-context checkpoints public, one long-context checkpoint gated and another marked as coming source
Latest update† partly verified · new weights
Basis: Model hub collection update (weights axis)
After the update one long-context checkpoint is still marked as coming; the preprint also has a v2 revision dated 2026-08-28, whose date was not independently re-checked this round source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Preprint and weights released within 12 months
OpennessCode: Open
Weights: Non-commercial
Code is public; weights carry the developer's own non-commercial research licence, with one checkpoint gated and another still marked as coming
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, release_milestones.weights_release, open_status.weights_license, open_status.code_license
As of2026-11-16

What it does

Addresses full-length mRNA exceeding existing RNA models' context using dense bidirectional attention, single-nucleotide tokens and attention-isolated sequence packing, pretraining at a native 10,240-nt context so a whole transcript fits at once; the same backbone also carries a discrete-diffusion mRNA generation line whose evaluation is not covered on this site. Comparators are several encoder models of the same class, with a large version of Evo 2 as a decoder-side reference.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
39 RNAGymDeveloper report, official protocol—Self-reportedThe developer self-reports zero-shot results on this set's fitness task under its own aggregation. · evidence

Other evaluations (outside the 44)

  • Whole-transcript linear-probing benchmark (Public benchmark not among the 44) Self-reported: Self-reported results on six task classes of an mRNA benchmark not among this site's 44, against encoder models of the same class
  • Own long-context reconstruction and representation tests (Own task set) Self-reported: Long-context reconstruction and RNA-class separation under frozen representations, defined by the authors

Developers report: Developers report the strongest representations among encoder models of its class and the ability to process a whole transcript at once in a native long context.

Independent evidence: No independent evidence found. No third-party reproduction was found; all evaluation results are self-reported

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYes67.6 million deduplicated RNA sequences and 85.7 billion nucleotide tokens from specified versions of three public sequence and genome-annotation resources source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationNoDownstream evaluation uses existing experimental datasets; no new experiments source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The lead organisation is a new AI research and doctoral-training institute founded in Shanghai in 2024-09, not a company; collaborators include a national molecular cell science centre and a university.
  • Funding: No separate funding disclosure was found; backers and scale are not public.
  • Products and partnerships: The lead team was previously noted for work on reasoning in large models, and this model is its entry into RNA; no pharmaceutical partnership is disclosed.
  • Related models: Evo 2, RhoFold+/DRfold2/trRosettaRNA

Reading notes

  1. Its lead is confined to encoder models of its own class; the decoder-side reference scores higher on the same evaluation.
  2. Short-context checkpoints do better on some tasks, with the long-context advantage concentrated where a whole transcript is needed.
  3. All results are self-reported with checkpoint choice and protocol set by the authors, and no third party has reproduced them.
  4. A non-commercial licence and gated checkpoints limit external use and checking.

Caveats

  1. The decoder comparator's values come from the evaluation's own report, not a re-run in the same environment.
  2. One long-context checkpoint was still unreleased at the check date, so its reported results cannot be reproduced.
  3. The other whole-transcript benchmark is not among this site's 44, so its values cannot be listed beside entries here.

Sources

  1. Preprint (preprint) · 2026-08-24
  2. Preprint full text PDF (preprint)
  3. Model hub collection (model hub)
  4. Code repository (repository)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.