AI for Bio · Model

BulkFormer

Foundation model for human bulk transcriptomes at about 150 million parameters

Genomics & RNAAcademic labChinaActive中文

Overview

NameBulkFormer · aliases: BulkFormer, BulkFormer-Atlas, BulkFormer-Latest
Full nameThe name is a compound word with no expanded form in the paper
Ambiguous names (not auto-linked)None
FamilyThe paper version spans five sizes, with two larger variants added in November 2026 under different hosting and licence statements
  • BulkFormer (paper checkpoints) · †
  • BulkFormer-Atlas ·
  • BulkFormer-Latest ·
OrganizationPeking University (biomedical informatics), Jilin University, Wuhan Sports University
Organization typeAcademic lab
CountryChina
Model classTranscriptome foundation model (masked reconstruction)
DomainGenomics & RNA (also: Single-cell & perturbation)
First release† unverified · paper revision
Basis: Preprint v1
Note: The date comes from the existing record and was not rechecked this round source
Other milestones
  • First public evidence: † · Preprint v1 source
  • Preprint v1: † · Preprint v1 source
  • Journal online: † · Journal publication (registry records that month) source
  • Code release: · Repository tag releasing two new variants source
  • Weights release: · Final checkpoint committed source
Latest update · new weights
Basis: Final checkpoint committed
The repository tagged two new model variants on 4 November 2026 and committed the final checkpoint of one on 6 November 2026; the journal paper appeared in July 2026.
Note: The site recorded 2026-11-04; the later commit is 11-06 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Journal publication and new weights within 12 months
OpennessCode: Open · MIT
Weights: Open · MIT
Paper-version code and weights under MIT; the two November 2026 variants are hosted on a general file-sharing service with no separate licence statement
Benchmarks reported0 counted · none of the 44
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.paper_v1, release_milestones.journal_online, open_status.weights_license
As of2026-11-16

What it does

Works on human bulk RNA sequencing: it takes expression over about twenty thousand protein-coding genes, pretrains by masking and reconstructing values with a gene co-expression graph as an attention prior, and attaches simple heads to gene or sample embeddings for imputation, disease annotation, prognosis, drug response, compound perturbation and gene essentiality. The other omics foundation models here, such as CellFM, scFoundation and scGPT/Geneformer, pretrain on single-cell data; this is the only bulk one, and those models are its comparisons.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Six downstream tasks (preprint)

Other evaluations (outside the 44)

  • Self-built downstream task set (Own task set) Self-reported: The preprint abstract lists six downstream tasks and the published abstract five, against single-cell foundation models and simple baselines
  • Third-party pan-cancer mutation probing (Public benchmark not among the 44) Third-party: A corporate lab's conference abstract probes frozen embeddings linearly and places this model behind two single-cell foundation models
  • A snapshot of a public bulk leaderboard (Public benchmark not among the 44) Self-reported: The repository states that the new variant combined with the organiser's baseline topped one day's snapshot while ranking lower alone, and warns that rankings move

Developers report: a roughly 150-million-parameter bulk transcriptome foundation model pretrained on over half a million public samples that beats single-cell foundation models and simple baselines across its chosen downstream tasks.

Independent evidence: A corporate lab's conference abstract ranks it behind two single-cell foundation models on pan-cancer mutation probing, and a research institute's preprint uses it as a comparison reference.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesHuman bulk transcriptomes from two public expression archives covering 20,010 protein-coding genes; the preprint gives 522,769 samples and the published abstract 581,503 source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoDownstream tasks use public cohorts and public drug-response and essentiality datasets

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: A biomedical informatics lab in a Chinese university's school of basic medical sciences together with a national key laboratory, with co-authors from two other universities.
  • Funding: The preprint acknowledges one national natural science foundation grant and the published version lists an additional grant number.
  • Products and partnerships: Code and weights are maintained in a public repository and model hub, with the repository also recording a submission to a public bulk leaderboard; no company or pharmaceutical partnership is disclosed.
  • Related models: CellFM, xTrimoSCPerturb / scFoundation, scGPT / Geneformer

Reading notes

  1. The sweep is over tasks the authors chose, comparing models pretrained for single-cell data through one shared interface.
  2. A third-party comparison places it behind two single-cell models, so the advantage varies by task rather than covering bulk work generally.
  3. Preprint and published version differ on sample count and task count, so citations must state the version.
  4. Its value is filling the gap left by single-cell-only coverage; by foundation-model standards it is small.

Caveats

  1. It has no pairing among the 44 benchmarks here; all evaluations are self-built.
  2. The first-release date comes from the existing record and was not rechecked.
  3. The journal date is verifiable only to the month; the day given in a Chinese report was not reproduced.
  4. The November 2026 weights sit on a general file-sharing service with no separate licence statement.
  5. The leaderboard snapshot is self-reported, and the developer itself warns the ranking moves.

Sources

  1. Journal paper 2026 (journal)
  2. Preprint 2025 (preprint)
  3. Code repository (repository)
  4. Third-party conference abstract 2026 (third-party paper)

Corrections

1 correction
FieldWasNowBasisFound
latest_update2026-11-04 (new release tag)2026-11-06 (final checkpoint commit)source2026-11-16

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