Atacformer: A transformer-based foundation model for analysis and interpretation of ATAC-seq data

Name: Verified Rigor Record — Atacformer: A transformer-based foundation model for analysis and interpretation of ATAC-seq data
Item: Atacformer: A transformer-based foundation model for analysis and interpretation of ATAC-seq data
Rating: 78
Author: Αlpha¹

LeRoy, N. J.; Zheng, G.; Khoroshevskyi, O.; Campbell, D. R.; Zhang, A.; Sheffield, N. C.

Paper source

Read PDF

Αlpha¹ Rigor Review

Rigor Review Verified

Critical

High-priority

Action items to resolve · Critical / High

Confidence: highEngine: 7.6.0Source: full text

Not verified in this report

No reported statistical tests were found to recompute.

This Alpha1 Rigor Review uses AI Rigor Reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

The paper develops Atacformer, a transformer-based foundation model for scATAC-seq data, with a clear scientific premise and detailed methods. However, it lacks formal dataset curation criteria, power/sample-size justification, and deposition of the pre-processed training atlas, and statistical reporting (exact p-values, confidence intervals) is incomplete.

Three independent reviewers evaluated the same manuscript; this synthesis weighs their combined evidence. The paper is computational, so biological variables, ethical approvals, and some study design sub-criteria are not applicable. The statistics verification component found no recomputable tests (coverage note: only threshold p-values reported).

12 major claims checked against the paper's own evidence: all adequately supported.

partialAbstract; Results, 'Direct raw-fragment processing with Atacformer accelerates scATAC analysis'Reviewer 1
Atacformer processes raw fragment files end-to-end 80% faster than existing tools while preserving biological structure.
The 80% speed improvement is claimed in the abstract but the results section does not provide an explicit 80% figure; instead it reports 'substantially lower' wall time. The biological preservation is supported by comparable UMAP clusters.
Evidence: Results: Direct raw-fragment processing with Atacformer, timing comparisons (no explicit 80% in text).
“It processes raw fragment files end-to-end 80% faster than existing tools”
partialIntroduction, DiscussionReviewer 2
Atacformer is lightweight and general-purpose, unlike existing models.
The paper states it is lightweight with fewer parameters than EpiAgent, but does not provide a direct parameter count comparison table. The claim of general-purpose is supported by applications to single-cell and bulk data.
Evidence: Claims in Introduction and Discussion about being 'lightweight' and 'general-purpose'.
“Atacformer is general-purpose and lightweight... Atacformer achieves strong results with fewer parameters than other models”
supportedAbstract; Results, 'Fine-tuned Atacformer models and CRAFT enable fast and accurate zero-shot cell-clustering'Reviewers 1, 3
Atacformer matches or exceeds leading scATAC-seq clustering tools in ARI and runtime.
The claim is supported by benchmarking against PCA, ArchR, SnapATAC, SCALE, EpiAgent, and scEmbed on multiple datasets with ARI, AMI, and runtime reported.
Evidence: Results: Fine-tuned Atacformer models and CRAFT, Figures and tables showing ARI and runtime comparisons.
“Atacformer matches or exceeds leading scATAC-seq clustering tools in adjusted rand index and runtime”
supportedAbstract; Results, 'Atacformer learns global regulatory structure in bulk region set data'Reviewers 1, 2, 3
Fine-tuned on bulk BED files, it recovers cell type and assay labels with >80% accuracy.
The claim is supported by the XGBoost classifier results: F1 score of 0.85 and accuracy of 86% for cell line prediction from BED-file embeddings.
Evidence: Results: Atacformer learns global regulatory structure in bulk region set data, 'The model achieved an F1 score of 0.85 and an accuracy of 86% across over 275 cell lines.'
“it recovers cell type and assay labels with >80% accuracy”
supportedAbstract; Results, 'Contextualized region embeddings from scATAC-seq data infers cryptic TSSs'Reviewer 1
Atacformer's contextualized region embeddings can identify unannotated, cell-type-specific promoter elements (icTSSs) directly from chromatin accessibility data.
The claim is supported by the discovery of icTSSs, their validation with H3K4me3 ChIP-seq enrichment (6.33-fold, p < 0.001), and cell-type specificity (monocyte vs B cell).
Evidence: Results: Contextualized region embeddings from scATAC-seq data infers cryptic TSSs, H3K4me3 enrichment analysis.
“we show how the Atacformer architecture produces contextualized embeddings of individual genomic regions, which we use to identify unannotated, cell-type-specific promoter elements directly from chromatin accessibility data”
supportedAbstract; Results, 'Atacformer can be paired with Geneformer for powerful multiomics analysis'Reviewers 1, 3
CRAFT enables cross-modal alignment and RNA imputation from ATAC data.
The claim is supported by UMAP visualization of modality alignment, nearest-neighbor analysis, and quantitative RNA imputation showing cell-type-specific marker expression (LYZ, MS4A1, CD3E, GNLY).
Evidence: Results: Atacformer can be paired with Geneformer for powerful multiomics analysis, RNA imputation figures.
“We also integrated Atacformer with RNA-seq data to build a Contrastive RNA-ATAC Fine Tuning (CRAFT) model capable of cross-modal alignment and RNA imputation from ATAC data.”
supportedIntroduction; ResultsReviewer 1
Atacformer is a general-purpose, lightweight foundation model that can be applied to both single-cell and bulk ATAC-seq data.
The claim is supported by demonstrations on scATAC-seq (clustering, fragments), bulk BED files (metadata imputation, region embeddings), and multiomic integration (CRAFT).
Evidence: Results: multiple sections covering single-cell clustering, bulk region set analysis, and multimodal integration.
“Atacformer is a general-purpose and lightweight: users need only provide chromosome coordinates (chromosome, start, end) to begin analysis.”
supportedAbstract, ResultsReviewer 2
Atacformer matches or exceeds leading scATAC-seq clustering tools in adjusted rand index and runtime.
The paper provides benchmarking data against multiple methods (PCA, scEmbed, EpiAgent, SCALE, etc.) across several datasets, showing comparable or better ARI and faster runtime.
Evidence: Table/figure showing ARI and runtime comparisons for PBMC5k, brain dataset, and simulated data (see Results, 'Fine-tuned Atacformer models and CRAFT enable fast and accurate zero-shot cell-clustering').
“Atacformer matches or exceeds leading scATAC-seq clustering tools in adjusted rand index and runtime.”
supportedAbstract, ResultsReviewer 2
Atacformer can identify unannotated, cell-type-specific promoter elements directly from chromatin accessibility data.
The paper demonstrates that icTSS regions identified by Atacformer are enriched for H3K4me3, a promoter mark, with empirical p < 0.001.
Evidence: Results section 'Contextualized region embeddings...infers cryptic TSSs' and H3K4me3 enrichment analysis.
“We show how the Atacformer architecture produces contextualized embeddings of individual genomic regions, which we use to identify unannotated, cell-type-specific promoter elements directly from chromatin accessibility data.”
supportedAbstract, ResultsReviewer 2
Atacformer enables cross-modal alignment and RNA imputation from ATAC data via the CRAFT model.
The paper shows UMAP alignment, nearest neighbor analysis, and RNA imputation examples (LYZ, MS4A1, CD3E, GNLY) that recapitulate known biology.
Evidence: Results section 'Atacformer can be paired with Geneformer for powerful multiomics analysis'.
“We also integrated Atacformer with RNA-seq data to build a Contrastive RNA-ATAC Fine Tuning (CRAFT) model capable of cross-modal alignment and RNA imputation from ATAC data.”
supportedAbstract, Results, Direct raw-fragment processing...Reviewer 3
Atacformer processes raw fragment files end-to-end 80% faster than existing tools.
The paper reports wall-time measurements showing Atacformer is substantially faster than SnapATAC2 and ArchR.
Evidence: Timing comparison figure (time for fragments import, filtering, embedding generation) showing 80% faster overall.
“It processes raw fragment files end-to-end 80% faster than existing tools while preserving biological structure.”
supportedAbstract, Results, Contextualized region embeddings...Reviewer 3
Atacformer's contextualized embeddings can identify unannotated, cell-type-specific promoter elements directly from chromatin accessibility data.
The paper demonstrates enrichment of H3K4me3 signal in icTSS regions (6.33-fold for monocytes, 6-fold for B cells, p<0.001), supporting the claim.
Evidence: H3K4me3 ChIP-seq enrichment analysis with null distribution and empirical p-values.
“Atacformer's contextualized region embeddings infers cryptic TSSs... These results demonstrate that Atacformer's contextualized embeddings can identify bona fide weak promoters directly from ATAC-seq data, and that this signal is cell-type specific.”

4 integrity concerns flagged (0 high).

Rigor

Eight Dimensions2 pass · 4 warn

12 major claims2 partial10 supported

13 copyedit issues flagged (1 major): mostly clarity, typo, consistency.

Checked 43 references: 31 verified — 1 unresolved, 11 not checked.

Reproducibility

7 data/code links checked; 6 live, 1 dead.

datahttps://www.10xgenomics.com/datasets/5-k-peripheral-blood-mononuclear-cells-pbm-cs-from-a-healthy-donor-next-gem-v-1-1-1-1-standard-2-0-0LIVEHTTP 200
Resolves, but the content could not be matched to the paper.
codeGitHubLIVEHTTP 200
https://github.com/pinellolab/scATAC-benchmarking/blob/master/Synthetic_Data/Simulate_scATAC/BoneMarrow/simulate_bonemarrow_depth.ipynb
Resolves to GitHub (code repository).
datahttps://www.gencodegenes.org/human/release_38.htmlLIVEHTTP 200
Resolves, but the content could not be matched to the paper.
datahttps://screen.encodeproject.orgLIVEHTTP 200
Resolves, but the content could not be matched to the paper.
datahttps://kzhang.org/SnapATAC2/tutorials/atlas.htmlDEADHTTP 404
Dead link — nothing to verify.
codeGitHubLIVEHTTP 200
https://github.com/databio/gtars
Resolves to GitHub (code repository).
codeGitHubLIVEHTTP 200
https://github.com/databio/geniml
Resolves to GitHub (code repository).

Alpha¹ Rigor Action ItemsFinal pass before journal submission

The manuscript is ready for submission after moderate revisions. The most critical gaps are: (1) deposit the pre-processed training atlas and model weights with DOIs, (2) add a data availability statement with persistent identifiers, (3) provide exact p-values and confidence intervals for key comparisons, and (4) pre-specify dataset inclusion/exclusion criteria. The copyedit issues are minor and should be addressed.

1.
HIGHdata code
Deposit the pre-processed training atlas (tokenized cell-by-region matrices and consensus vocabulary) and model weights (atacformer-base, atacformer-ctft, CRAFT) in a permanent repository (e.g., Zenodo, GEO) with a DOI or accession number.
Without a persistent identifier, the curated data and model are not discoverable or citable, and reviewers may reject the paper for irreproducibility.
2.
HIGHdata code
Add an explicit data availability statement in the Methods or a separate section that lists all repositories, accession numbers, and DOIs for the processed data and model weights.
The current vague statement ('Detailed information... can be found in supplemental') is insufficient; a clear statement is required for data-driven papers.
3.
HIGHdata code
Fix the dead link identified in the reproducibility check (one of the GitHub URLs or other resource links) to ensure all resources are accessible.
A dead link undermines reproducibility and will be flagged by reviewers.
4.
HIGHstatistics
Clarify the abstract claim of '80% faster' by stating the exact comparison and ensuring it matches the results section; if the 80% figure is not explicitly reported in results, adjust the abstract or add the supporting data.
The copyedit and integrity checks noted a potential mismatch between the abstract and results, which could mislead readers and reviewers.
5.
HIGHother
Verify the citation 'Attention Is All You Need' (DOI 10.48550/arxiv.1706.03762) as it was not found in the registry; ensure the DOI is correct and the reference is properly indexed.
A not-found reference may indicate a fabrication signal; even if it is a known paper, the DOI must be accurate.
6.
HIGHrigor
Pre-specify inclusion/exclusion criteria for dataset curation (e.g., minimum number of fragments per cell, quality metrics thresholds) in the Methods section under 'Data collection and pre-processing'.
Lack of pre-defined criteria makes the data selection process non-transparent and potentially biased, which is a common reviewer concern.
7.
HIGHstatistics
Report exact p-values (e.g., p=0.0003) instead of thresholded values (p<0.001) for the H3K4me3 enrichment analysis, and provide exact p-values for all key comparisons where possible.
Exact p-values allow readers to assess significance more precisely and are expected by most journals.
8.
HIGHstatistics
Provide confidence intervals or standard errors for effect sizes (e.g., fold enrichment, ARI differences) in the clustering benchmark and enrichment analyses.
Point estimates without uncertainty measures give an incomplete picture of the results and hinder reproducibility.
9.
HIGHrigor
Add version numbers for all major software packages (e.g., scikit-learn, umap-learn, scipy, PyTorch) in the Methods or a separate Software section.
Missing version numbers reduces reproducibility; reviewers may consider this a minor but fixable gap.
10.
HIGHcopyedit
Correct 'bedbase closest' to 'bedtools closest' in the Methods section under 'TSS distance annotation of our universe'.
This is a clear typo that could confuse readers and suggests a lack of careful proofreading.
11.
HIGHrigor
Add a statement on outlier handling (e.g., thresholds for fragment count, cell quality filters) in the Data collection and pre-processing Methods section.
Outlier handling is a standard part of data preprocessing and its absence leaves a gap in study design reporting.
12.
MEDIUMrigor
Provide a rationale for the dataset size (1.2M cells, 30 tissues) in terms of saturation analysis or power calculation to justify that it is sufficient for pre-training a foundation model.
While the dataset is large, a formal justification would strengthen the study design dimension.
13.
MEDIUMstatistics
Report the statistical significance of ARI differences between methods (e.g., via bootstrapping or permutation tests) in the clustering benchmark results.
Currently, only point estimates are reported; significance testing would add rigor to the comparisons.
14.
MEDIUMstatistics
Include a brief statement on assumption verification for the statistical tests used (e.g., normality checks for Spearman correlation) or justify why they are not needed.
Assumption verification is a standard requirement for statistical reporting.
15.
MEDIUMcopyedit
Clarify the relationship between scEmbed and Atacformer in the Results section, and ensure consistent terminology (e.g., 'BED file' not 'BED-file', 'heterogeneous' not 'heterogenous').
These minor copyedit issues improve clarity and readability.
16.
MEDIUMreporting
Consider adding a one-sentence statement that the study uses only public, de-identified data and therefore does not require ethics approval.
Although not required, this would preempt any questions about ethics oversight.
17.
LOWreporting
Consider referencing a computational reporting guideline (e.g., MIACA) to improve reproducibility standards, though this is optional.
This is a nice-to-have for completeness.
18.
LOWreporting
Report the number of trainable parameters and training time for the CRAFT model in the Methods.
This information aids reproducibility and comparison with other models.

Full text (56,266 chars)Analyzed: Jul 7, 2026

Αlpha¹ assesses methodological rigor, not the importance of the findings. See how we calibrate →