BulkFormer: A large-scale foundation model for bulk transcriptomes
Clinical Snapshot
PICO Framework
| P — Population | Human bulk RNA-seq transcriptomic profiles (581,503 samples spanning diverse tissues, disease states, and experimental conditions) |
| I — Intervention | BulkFormer — a ~150 million parameter transformer-based foundation model pretrained on bulk RNA-seq data, incorporating a hybrid encoder combining a graph neural network (GNN) for gene-gene relationships and a Performer module for global expression dependencies, covering 20,010 protein-coding genes |
| C — Comparator | Existing single-cell RNA-seq foundation models (e.g., scGPT, Geneformer, and related architectures) pretrained on sparse single-cell data (~3,000 genes per cell) |
| O — Outcomes | Performance across five downstream transcriptomic tasks including tissue/cell-type classification, disease state prediction, gene expression imputation, drug response prediction, and biological embedding quality; training computational cost |
Bottom Line
BulkFormer represents a methodologically significant advance in transcriptomic foundation modelling by addressing the fundamental mismatch between single-cell pretraining data and bulk RNA-seq inference requirements. Its pretraining on over 580,000 human bulk RNA-seq profiles — covering nearly the complete human protein-coding genome — and its hybrid GNN-Performer architecture are conceptually well-motivated and architecturally novel. The reported superiority over single-cell foundation models across five downstream tasks, combined with lower computational cost, is a compelling claim. However, the abstract provides no quantitative performance metrics, confidence intervals, or statistical testing, making independent assessment of the magnitude and reliability of these advantages impossible at this stage. Critical methodological details — including data curation, batch correction, train-test separation, and comparator fine-tuning conditions — are absent from the available summary. For Australian clinicians and clinical scientists, BulkFormer is a research-stage tool with no current regulatory approval or prospective clinical validation. It warrants attention from academic genomics centres but should not be applied to clinical decision-making without independent validation in prospective cohorts. Full manuscript review, including supplementary benchmarking data and transparent peer review documentation, is essential before drawing firm conclusions about its clinical utility.
Key Findings
P Value: Not reported
Effect Size: Not reported in abstract — no quantitative performance metrics provided
Primary Outcome: BulkFormer outperformed existing single-cell RNA-seq foundation models across five downstream bulk transcriptomic tasks (specific tasks not enumerated in abstract but likely include tissue classification, disease prediction, gene expression imputation, drug response prediction, and embedding quality assessment)
Nnt Or Sensitivity: Not applicable in standard clinical sense; no sensitivity/specificity, AUROC, or equivalent diagnostic accuracy metrics reported in the abstract. Training cost described as 'substantially lower' than comparators but not quantified.
Confidence Interval: Not reported
Clinical Application
Feasibility for routine clinical application is currently limited. The model requires substantial computational infrastructure (~150M parameters, large-scale pretraining). Integration into clinical bioinformatics pipelines would require prospective validation, regulatory consideration, and standardised input data pipelines. Research feasibility in well-resourced academic medical centres is reasonable given public availability of the model (assumed, not confirmed in abstract). BulkFormer has no current direct PBS or TGA regulatory pathway — it is a research-grade computational tool, not a therapeutic or in vitro diagnostic device. However, it is relevant to Australian clinical genomics initiatives including the Australian Genomics Health Alliance, ZERO Childhood Cancer Program, and the Australian Cancer Atlas. The RACGP does not currently have guidelines addressing transcriptomic foundation models. Australian researchers at institutions such as the Garvan Institute, WEHI, and QIMR Berghofer may find BulkFormer relevant for bulk RNA-seq research pipelines. For clinical translation, TGA regulation as a Software as a Medical Device (SaMD) under the ARTG would be required. The model's training data likely underrepresents Aboriginal and Torres Strait Islander genomic diversity, a critical equity consideration for Australian implementation. Researchers and clinical scientists performing bulk RNA-seq analysis in human tissue samples across oncology, immunology, pharmacogenomics, and systems medicine. Potential future application to clinical transcriptomic profiling pipelines for disease classification, biomarker discovery, and treatment response prediction — but currently at a research/pre-clinical stage only.
Abstract
Foundation models are transforming transcriptome analysis, yet most RNA sequencing (RNA-seq) foundation models are pretrained on sparse single-cell RNA-seq data, which typically captures only ∼3,000 genes per cell. This creates a need for models tailored to bulk transcriptomes, a modality that profiles ∼16,000 genes per sample and supports clinical and tissue-level analyses. Here, we present BulkFormer, a foundation model for bulk transcriptome analysis. BulkFormer contains ∼150 million parameters, covers 20,010 protein-coding genes, and is pretrained on 581,503 human bulk RNA-seq profiles. Its hybrid encoder combines a graph neural network to model explicit gene-gene relationships with a Performer module to capture global expression dependencies. Across five downstream tasks, BulkFormer outperformed existing single-cell foundation models while requiring substantially lower training cost. These results identify pretraining data modality as a key determinant of foundation model performance and establish BulkFormer as a framework for bulk transcriptome modeling. A record of this paper's transparent peer review process is included in the supplemental information.
References
- 1.Kang, B., Fan, R., Yi, M., Cui, C., & Cui, Q. (2026). BulkFormer: A large-scale foundation model for bulk transcriptomes. Cell Systems. https://doi.org/10.1016/j.cels.2026.101657
Related Research
Journal of chemical information and modeling
Functionally Guided Graph Learning for Robust Cross-Patient Cell-Type Annotation in Single-Cell RNA Sequencing
29 July 2026
Journal of chemical information and modeling
Unveiling Large-Scale Kinase-Centric Protein-Protein Interactions through a Knowledge-Informed Workflow
28 July 2026
Molecular genetics and genomics : MGG
The use of artificial intelligence in advancing molecular biology in Africa: a narrative review
23 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service