AI for Bio · Model

scPRINT-2

Institut Pasteur's second-generation single-cell foundation model pretrained on 350M cells from 16 species

Single-cell & perturbationAcademic labFranceActive中文

Overview

NamescPRINT-2 · aliases: scPRINT-2, scprint2
Full nameThe preprint gives no expanded name
Ambiguous names (not auto-linked)scPRINT
FamilyNot a family entry. The predecessor scPRINT has no record and is mentioned as background
OrganizationInstitut Pasteur, CNRS
Organization typeAcademic lab
CountryFrance
Model classSingle-cell FM
DomainSingle-cell & perturbation
First release · paper revision
Basis: bioRxiv preprint v1 source
Other milestones
  • First public evidence: · bioRxiv preprint v1 source
  • Preprint v1: · bioRxiv preprint v1 (v3 on 2026-02-11) source
  • Code release: · First PyPI release source
Latest update† partly verified · new version
Basis: GitHub release 2.3.8
GitHub release 2.3.8; assets not inspected, so not to be read as new weights. PyPI is a separate version axis, latest 1.0.3 (2026-02-24)
Note: Whether this release contains new weights is unverified source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Code release within 12 months
OpennessCode: Open · GPL-3.0 (GitHub) / MIT (PyPI metadata)
Weights: Open
Code public and weights on Hugging Face; the GitHub licence is GPL-3.0 while PyPI metadata say MIT, an inconsistency
Benchmarks reported0 counted (unconfirmed 2)
Wet-lab validationNone recorded
Unverified fields† latest_update, open_status.code_license, benchmarks_reported[0], benchmarks_reported[1]
As of2026-11-17

What it does

A Transformer single-cell model pretrained with masked expression prediction, denoising, contrastive cell embedding and hierarchical classification along the cell ontology; the paper first builds an "additive" benchmark to find which design changes help. It targets zero-shot denoising, cell embedding, cell-type prediction and gene regulatory network inference. Compared with CellFM and scGPT/Geneformer it stands out for 16-species pretraining and network inference.

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. 0 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
09 scIB (+Open Problems)Unconfirmed (not counted)—Unconfirmed sourceAbstract does not name scIB †disputed · evidence
21 DREAM5 / BEELINE / CausalBenchUnconfirmed (not counted)—Unconfirmed sourceThe preprint mentions its own GRN benchmark with scores in the body, not under this family's protocol. †disputed · evidence

Other evaluations (outside the 44)

  • "Additive" design benchmark (Own task set) Self-reported: The authors' own benchmark used to compare design changes and finalise the model
  • Denoising and cell-type prediction (Own task set) Self-reported: Zero-shot; the abstract claims state of the art without numbers

Developers report: state of the art on denoising, cell embedding and cell-type prediction, improved gene network inference, and generalisation to unseen modalities and species.

Independent evidence: No independent evidence found. No third-party leaderboard or independent evaluation result found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesOver 350M cells from more than 22k datasets and 16 species, mainly CZ CELLxGENE and scBaseCount source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic lab: the Machine Learning for Integrative Genomics junior group at Institut Pasteur, part of CNRS UMR3738, focused on single-cell data, with CNRS / ENS collaborators.
  • Funding: The group holds an ERC Starting Grant and a completed ANR project; funding lines for this preprint are not public.
  • Products and partnerships: The predecessor scPRINT appeared in Nature Communications in 2025-04; data loading and management use LaminDB tools; no company or pharma partnership disclosed.
  • Related models: CellFM, scGPT / Geneformer, TranscriptFormer, xTrimoSCPerturb / scFoundation

Reading notes

  1. Building a benchmark before the model is more transparent, but the readable pages have no scores, so it cannot be ranked.
  2. The state-of-the-art claim is author-reported within the authors' own benchmark scope and needs third-party confirmation.
  3. Licences differ between GitHub and PyPI; check the repository LICENSE before commercial use.

Caveats

  1. The scIB and DREAM5 pairings are unconfirmed and not counted.
  2. Whether GitHub 2.3.8 contains new weights is unverified.
  3. A large cell count does not imply balanced species or tissue coverage.

Sources

  1. bioRxiv preprint (preprint)
  2. Code repository (repository)
  3. Package registry page (package registry)
  4. bioRxiv version record (preprint)
  5. Predecessor journal paper (journal)
  6. Research group page (official site)

Corrections

1 correction
FieldWasNowBasisFound
latest_update2026-02-24(PyPI 1.0.3)2026-09-26 (GitHub release 2.3.8; PyPI is a separate version axis)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.