AI for Bio · Model

C2S-Scale

Single-cell model that turns ranked expression into "cell sentences" and fine-tunes general LLMs from 410M to 27B parameters

Single-cell & perturbationAcademic labUnited StatesActive中文

Overview

NameC2S-Scale · aliases: C2S-Scale, Cell2Sentence
Full nameCell2Sentence-Scale
Ambiguous names (not auto-linked)None
FamilyPart of the Cell2Sentence family
OrganizationYale University, Google Research, Google DeepMind
Organization typeAcademic lab
CountryUnited States
Model classSingle-cell language model
DomainSingle-cell & perturbation (also: Text, LLMs & agents)
First release · preprint
Basis: Preprint v1
Note: The site recorded 2025-04, month level only source
Other milestones
  • First public evidence: · Preprint v1 source
  • Preprint v1: · Preprint v1 source
  • Weights release: † · Largest-size weights released on the model hub alongside an official blog post source
Latest update · paper revision
Basis: Preprint v4
The preprint was revised to v4 (v2 on 2025-10-11, v3 on 2025-10-26); the earlier event is the 2025-10 release of the largest size. No journal version was found source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Preprint revisions and a weights release within 12 months
OpennessCode: Open
Weights: Open · CC-BY-4.0
Code is in a public repository and weights on the model hub with a Creative Commons licence on the model card; the code licence was not re-checked this round
Benchmarks reported1 counted (modified protocol 1)
Wet-lab validation1 recorded campaign
Unverified fields† release_milestones.weights_release, wetlab_validation[0], open_status.code_license
As of2026-11-16

What it does

Ranks each cell's most expressed genes into a text sequence and uses a general language model for cell-type annotation, conditional generation and question answering, extending to spatial neighbourhoods and perturbation response; it keeps the general LLM architecture rather than a single-cell-specific one, the opposite of TranscriptFormer. The developers report that some context-dependent reasoning appears only at the largest size.

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
24 Virtual Cell Challenge and perturbation benchmarksDeveloper report, modified protocol—Self-reportedUses its own datasets and metrics, not the VCC protocol. · evidence

Other evaluations (outside the 44)

  • Cell-type annotation comparison (Comparison against baselines only) Self-reported: Self-reported parity with two single-cell foundation models on annotation; a comparison rather than a public-protocol evaluation
  • Self-chosen metric for integration (Own task set) Self-reported: The preprint does not use the common integration metrics and substitutes another alignment measure, which cannot be listed beside the site's integration evaluation

Developers report: Developers report that handling single-cell data with a general language model improves with scale and can produce biological hypotheses testable in the lab.

Independent evidence: No independent evidence found. The preprint is not peer reviewed and no third-party evaluation of its perturbation or annotation results was found

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesOver 50 million human and mouse cells from public cell atlases and other sources plus biological text and metadata, about 1 billion tokens and 150 million training samples source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationYesOne drug hypothesis was tested in the lab, as validation rather than training source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Experimental test of a single drug hypothesis (2025-10 and later preprint versions)
Targets 1† · Designs tested 1† · Hits 1† · Success 1/1 hypothesis tested† · Best affinity —

Denominator basis: One hypothesis taken forward from a computational screen of over 4,000 drugs; batch and replicate counts are not public, so no hit rate can be computed
The model predicted that a kinase inhibitor would raise antigen presentation only under low-interferon conditions, and an increase was seen in a previously unseen human cell model; the result exists only in the 2025-10 and later preprint versions, so citations must point at the current version
Executor: Run by the collaborating academic lab; not an independent reproduction

Self-reported · † · †partly verified · source

Background

  • Organization: A collaboration between a US university lab and two research units of a large technology company, built on that company's open model family.
  • Funding: The corporate partner supplied models and compute; no grant numbers are listed on the paper page or the blog.
  • Products and partnerships: The model family was released in 2025-04 and the largest size in 2025-10 with an official blog post; the same company's genomics work is AlphaGenome.
  • Related models: TranscriptFormer, AlphaGenome, scGPT / Geneformer

Reading notes

  1. Comparisons are self-reported, with the main text giving directional conclusions and per-item values left to the supplement.
  2. The drug-hypothesis test is a single case, not systematic wet-lab backing.
  3. It reports no public integration protocol, so it cannot be compared directly with single-cell models that do.
  4. The wet-lab result appears only from the 2025-10 version onward, so citing the earlier version misses it

Caveats

  1. The site linked the first preprint version, which does not contain the wet-lab result; the link must point at the current version.
  2. Replicate counts and executor attribution are unverified; no replicate figure is given.
  3. The preprint is not peer reviewed, so all comparison values are the authors' own.

Sources

  1. Preprint full text (preprint)
  2. Preprint version record (registry)
  3. Model hub weights page (model hub)
  4. Official blog (wet-lab result) (blog)
  5. Code repository (repository)

Corrections

3 corrections
FieldWasNowBasisFound
first_release2025-042025-04-17 (preprint v1 posted)source2026-11-16
latest_update2025-10-15 (weights release, recorded as new weights)2026-01-01 (preprint v4, kind paper revision); the weights release is recorded as an earlier eventsource2026-11-16
wetlab_validationCited the first preprint versionCites the 2025-10 and later versions; the first version does not contain this resultsource2026-11-16

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