CoChat

annATAC: automatic cell type annotation for scATAC-seq data based on language model

Open access

The DOI matches its Crossref record; 49 of the 49 references it cites were found in a registry or an index.

Abstract

BACKGROUND: Cell type annotation serves as the cornerstone for downstream analysis of single cell data. Nevertheless, scATAC-seq data is characterized by high sparsity and dimensionality, presenting significant challenges to its annotation process. RESULTS: We introduce a novel method based on language model, named annATAC, which is designed for the automatic annotation of cell types in scATAC-seq data. This method primarily consists of three stages. During the pre-training stage, by training on a vast amount of unlabeled data, the model can learn the interaction relationships between peaks, thus building a preliminary understanding of the data features. Subsequently, in the fine-tuning stage, a small quantity of labeled data is utilized to conduct secondary training on the model, which enables the model to identify cell types accurately. Finally, in the prediction stage, the trained model is applied to annotate scATAC-seq data. CONCLUSIONS: Compared with other automatic annotation methods across multiple datasets, annATAC demonstrates superiority on the annotation performance. Further experiments have validated that annATAC holds great potential in identifying marker peaks and marker motifs. It is expected that annATAC will provide more profound and precise analysis outcomes for scATAC-seq research. As a result, it will effectively promote the progress of relevant biomedical research.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1186/s12915-025-02244-5

Cites

49 of 49 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Overview of the annATAC. In step 1, collect millions of scATAC-seq data, covering 13 kinds of tumor and normal tissues in the Homo sapiens. Then further convert these data into the form of a cell-peak island for the input of the model. In step 2, perform data masking and ignore the positions where the expression values are zero. In step 3, conduct unsupervised pre-training. Build an Encoder by introducing linear transformation into the multi-head attention mechanism. Train the model by minimizing the loss between the model’s output and the ground truth before masking. In step 4, carry out supervised fine-tuning. Through secondary training on a small amount of labeled data, the model is fully converged. In step 5, conduct downstream biological analysis
Fig. 1 Overview of the annATAC. In step 1, collect millions of scATAC-seq data, covering 13 kinds of tumor and normal tissues in the Homo sapiens. Then further convert these data into the form of a cell-peak island for the input of the model. In step 2, perform data masking and ignore the positions where the expression values are zero. In step 3, conduct unsupervised pre-training. Build an Encoder by introducing linear transformation into the multi-head attention mechanism. Train the model by minimizing the loss between the model’s output and the ground truth before masking. In step 4, carry out supervised fine-tuning. Through secondary training on a small amount of labeled data, the model is fully converged. In step 5, conduct downstream biological analysisCC BY-NC-ND · PubMed Central
Fig. 2 Performance evaluation of annATAC in annotating cell types. A Evaluation of the annotation performance of ACC, Jaccard_weighted, and Jaccard_micro on the scATAC-seq data of six adult tissues. The abscissa indicates the number of cells in each tissue. B Overall distribution performance evaluation of ACC on the scATAC-seq data of eight adult tissues. C t-SNE visualization display on the Heart and Stomach Tissues, where the first column shows the annotation of the original class labels. D Evaluation of annATAC’s performance in detecting novel cell types. The red color in the Sankey diagram represents novel cell types, with the real cell labels on the left side and the labels predicted by annATAC on the right side
Fig. 2 Performance evaluation of annATAC in annotating cell types. A Evaluation of the annotation performance of ACC, Jaccard_weighted, and Jaccard_micro on the scATAC-seq data of six adult tissues. The abscissa indicates the number of cells in each tissue. B Overall distribution performance evaluation of ACC on the scATAC-seq data of eight adult tissues. C t-SNE visualization display on the Heart and Stomach Tissues, where the first column shows the annotation of the original class labels. D Evaluation of annATAC’s performance in detecting novel cell types. The red color in the Sankey diagram represents novel cell types, with the real cell labels on the left side and the labels predicted by annATAC on the right sideCC BY-NC-ND · PubMed Central
Fig. 3 Performance evaluation of cell subtype annotation. A Comparing annATAC with other comparative algorithms in the AD datasets with the smallest sample size (2588) and the largest sample size (9761). B Semi-violin plots display the performance of annATAC and other algorithms in terms of ACC and F1 score on seven AD datasets with different sample sizes. C The experimental performance of two AD and control groups is evaluated in terms of ACC and F1 score. D A comparison displays the annotation results of annATAC and those of RAINBOW and Cellcano on four AD and one control datasets. The bottom row indicates different tissues and the number of cells. E Performing t-SNE visualization on the AD datasets shows that different colors represent different cell types
Fig. 3 Performance evaluation of cell subtype annotation. A Comparing annATAC with other comparative algorithms in the AD datasets with the smallest sample size (2588) and the largest sample size (9761). B Semi-violin plots display the performance of annATAC and other algorithms in terms of ACC and F1 score on seven AD datasets with different sample sizes. C The experimental performance of two AD and control groups is evaluated in terms of ACC and F1 score. D A comparison displays the annotation results of annATAC and those of RAINBOW and Cellcano on four AD and one control datasets. The bottom row indicates different tissues and the number of cells. E Performing t-SNE visualization on the AD datasets shows that different colors represent different cell typesCC BY-NC-ND · PubMed Central
Fig. 4 Biological analysis of AD. A UMAP visualization of annATAC’s cell type annotation for patient 4 with AD. B Proportions of regulatory elements contained in different cell types. C Display of marker peaks on different cell types, with red auxiliary lines indicating the positions of marker peaks. D Display of marker motifs on different cell types
Fig. 4 Biological analysis of AD. A UMAP visualization of annATAC’s cell type annotation for patient 4 with AD. B Proportions of regulatory elements contained in different cell types. C Display of marker peaks on different cell types, with red auxiliary lines indicating the positions of marker peaks. D Display of marker motifs on different cell typesCC BY-NC-ND · PubMed Central
Fig. 5 A Analysis of cell type-specific transcription factor footprints in AD patients. B Pseudo-time trajectory analysis of marker motifs and marker peaks in AD patients
Fig. 5 A Analysis of cell type-specific transcription factor footprints in AD patients. B Pseudo-time trajectory analysis of marker motifs and marker peaks in AD patientsCC BY-NC-ND · PubMed Central
Fig. 6 Model stability testing. A Model robustness testing was conducted on five tissues: Breast Bammary tissue (cell: 8547), Transverse Colon (cell: 6499), Esophagus Muscularis (cell: 10,783), Lung (cell: 3313), and Stomach (cell: 1266). RAW indicates that the labels were not shuffled during the fine-tuning process, dis_5 indicates that 5% of the labels were randomly shuffled, dis_10 indicates that 10% of the labels were randomly shuffled, and so on. B Randomly select 40%, 60%, and 80% of the fine-tuning data for training, and test the annotation ACC on two tissues, Gastrocnemius Muscle and Lung. C Pre-training ablation experiments. The upper part shows ACC and loss with pre-training, while the lower part shows the results without pre-training
Fig. 6 Model stability testing. A Model robustness testing was conducted on five tissues: Breast Bammary tissue (cell: 8547), Transverse Colon (cell: 6499), Esophagus Muscularis (cell: 10,783), Lung (cell: 3313), and Stomach (cell: 1266). RAW indicates that the labels were not shuffled during the fine-tuning process, dis_5 indicates that 5% of the labels were randomly shuffled, dis_10 indicates that 10% of the labels were randomly shuffled, and so on. B Randomly select 40%, 60%, and 80% of the fine-tuning data for training, and test the annotation ACC on two tissues, Gastrocnemius Muscle and Lung. C Pre-training ablation experiments. The upper part shows ACC and loss with pre-training, while the lower part shows the results without pre-trainingCC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 49 references

Research works printed but not linked: 1

  1. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–30.

Cited by

See all 3

Metadata

Authors
Cui, L., Wang, F., Li, H., Liu, Q., Zhou, M., & Wang, G.
Affiliations
  • Northeast Forestry University (CN)
  • Wenzhou Medical University (CN)
  • University of Electronic Science and Technology of China (CN)
  • Henan University (CN)
  • Harbin Institute of Technology (CN)
Corresponding author
Wang, Guohua
Venue
BMC Biology
Published
2025-05-28
Volume
23
Issue
1
Pages
145-145
ISSN
1741-7007 (ISSN-L)
OpenAlex
W4410796331
Cited by
3
References
49
Citations by year
2026: 3
FWCI
0.90
Type
journal article
Language
en
Open access
Licence
Locations
  • BMC Biology (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Natural Science Foundation of China (32400546, 62302342, 62450112)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Chromatin Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
Keywords
cell type annotation, single-cell ATAC-seq, language models, unsupervised pre-training, fine-tuning, single-cell chromatin accessibility
MeSH terms
Humans, Software, Molecular Sequence Annotation, Single-Cell Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: annATAC: automatic cell type annotation for scATAC-seq data based on language model. https://app.cochat.ai/papers/10.1186/s12915-025-02244-5
The paper
Cui, L., Wang, F., Li, H., Liu, Q., Zhou, M., & Wang, G. (2025). annATAC: automatic cell type annotation for scATAC-seq data based on language model. BMC Biology. https://doi.org/10.1186/s12915-025-02244-5

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.