CoChat
Sign Up
Cited by 193Single-cell and spatial transcriptomicsGene expression and cancer classificationBioinformatics and Genomic Networks

Transformer for one stop interpretable cell type annotation

Open access
Chen, Jiawei, Xu, Hao, Tao, Wanyu, Chen, Zhaoxiong, Zhao, Yuxuan, Han, Jing-Dong J.·2023·Nature Communications
DOI 10.1038/s41467-023-35923-4PMC PMC9840170
Open this paper in CoChat — free accountOpen PDF

Abstract

Consistent annotation transfer from reference dataset to query dataset is fundamental to the development and reproducibility of single-cell research. Compared with traditional annotation methods, deep learning based methods are faster and more automated. A series of useful single cell analysis tools based on autoencoder architecture have been developed but these struggle to strike a balance between depth and interpretability. Here, we present TOSICA, a multi-head self-attention deep learning model based on Transformer that enables interpretable cell type annotation using biologically understandable entities, such as pathways or regulons. We show that TOSICA achieves fast and accurate one-stop annotation and batch-insensitive integration while providing biologically interpretable insights for understanding cellular behavior during development and disease progressions. We demonstrate TOSICA’s advantages by applying it to scRNA-seq data of tumor-infiltrating immune cells, and CD14+ monocytes in COVID-19 to reveal rare cell types, heterogeneity and dynamic trajectories associated with disease progression and severity.

AI snapshot

Written from the paper’s full text.

TL;DR

TOSICA is a Transformer-based, multi-head self-attention model for interpretable cell type annotation that uses biologically meaningful tokens such as pathways or regulons. It transfers labels from reference to query scRNA-seq datasets with high accuracy, batch-insensitive integration, and attention-based interpretability across development, cancer, COVID-19, and SLE case studies.

Key findings

  • Across six benchmark datasets and 19 methods, TOSICA had the highest mean accuracy at 86.69%, ranked top 6 on every dataset, and reached 81.06% on the largest dataset mAtlas versus 79.57% for the second-best method ACTINN.
  • On datasets with uneven cell-type distributions between reference and query, TOSICA achieved 76.47% mean accuracy on five unbalanced cell types, compared with 63.23% for SingleR and 68.18% for SciBet; Kullback–Leibler divergence had the strongest negative correlation with accuracy (PCC = −0.9).
  • TOSICA discovered unseen or rare cell types: after deleting alpha cells from the hPancreas reference, 76% of query alpha cells were labeled “Unknown,” while other methods mislabeled them as PP, delta, or beta.
  • In integration benchmarks, scIB ranked TOSICA second of 11 applicable methods on pan-cancer myeloid cells and first of 13 methods on a COVID-19 dataset; on the T cell dataset, TOSICA’s runtime was shortest, while scGen took nearly five days.

Methods

  • TOSICA uses a multi-head self-attention Transformer with a Cell Embedding layer, Multi-head Self-attention layer, and Cell-Type Classifier. Gene expression is transformed into tokens using a mask matrix based on prior knowledge, such as GSEA Reactome or C3 gene sets, so tokens represent pathways or regulons; a CLS token is used for cell-type prediction.
  • The model is trained supervised on reference scRNA-seq data with cell-type labels, with 30% of the training set used as validation, cross-entropy loss, SGD, cosine learning rate decay, and convergence typically within 20 epochs. Attention scores between CLS and pathway tokens serve as interpretable cell embeddings for downstream UMAP, clustering, and trajectory analyses.
  • Benchmarks used six datasets—hArtery, hBone, hPancreas, mBrain, mPancreas, and mAtlas—compared against 18 other annotators, with accuracy defined as the fraction of correctly predicted cells. Integration was assessed with scIB metrics including batch ASW, NMI, ARI, graph connectivity, kBET, and isolated label F1.

Limitations

  • The paper states that no statistical method was used to predetermine sample size, experiments were not randomized, and investigators were not blinded to allocation during experiments and outcome assessment.
  • Evident from the text: benchmark accuracy depends on “ground truth” labels from original publications, and some comparison methods were not applicable to large or highly batched datasets, which may affect relative rankings.
  • Evident from the text: random masks sometimes converged to lower accuracy and required more epochs than knowledge-based masks, though TOSICA was generally robust to mask choice.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-023-35923-4

Cites

61 of 61 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 30 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    SingleR, SciBet, and ACTINN do not automatically identify alpha cells as a new cell type, instead incorrectly label them as PP, delta or beta.

    “Three other annotators with high average accuracy, SingleR, SciBet, and ACTINN (Fig. 2a), do not automatically identify ‘alpha’ cells as a new cell type, instead incorrectly label them as ‘PP’, ‘delta’ or ‘beta’ (Supplementary Fig. 9a–c).”
    Quote in the paperSubstantial claim
  • Findingown data

    On five unevenly distributed cell types, TOSICA (76.47%) beats SingleR (63.23%) and SciBet (68.18%) in mean accuracy.

    “On the five cell types, prefibrochondrocytes (preFC), prehypertrophic chondrocytes (preHTC), homeostatic chondrocytes (HomC), regulatory chondrocytes (RegC) and hypertrophic chondrocytes (HTC) that are more unevenly distributed in test or reference set, TOSICA (76.47%) beats the second (SingleR, 63.23%) or third (SciBet, 68.18%) highest mean accuracy methods (Fig. 2e).”
    Quote in the paperSubstantial claim
  • Findingown data

    CaSTLe recognizes most cells of all cell types as 'Unknown', including well-represented cell types.

    “CaSTLe even simply recognizes most of the cells of all cell types as ‘Unknown’, including the cell types that are well-represented in the training sets (Supplementary Fig. 9f).”
    Quote in the paperSubstantial claim
  • Findingown data

    The cell type distribution of hBone dataset is the most unbalanced between training and test set.

    “Not surprisingly, the cell type distribution of hBone dataset is the most unbalanced between training and test set (Fig. 2a).”
    Quote in the paperNot a substantial claim
  • Findingown data

    Other methods also predict MHC II as 'macrophage' or 'Unknown' like TOSICA.

    “Other methods also predict MHC II as ‘macrophage’ or ‘Unknown’ like TOSICA (Supplementary Fig. 9).”
    Quote in the paperNot a substantial claim
  • May combine two claims. 44
  • Classified as probably not a substantial claim. 10
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    Some cell types at low abundance in the reference of mAtlas are still identified well and clustered together by TOSICA, and separated from other cell types in the query set.

    “Some cell types are at low abundance in the reference of mAtlas, may be insufficient for training a good predictor, but TOSICA still identifies them well and clusters them together, also separates them from other cell types as much as possible in the query set (Fig. 2b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    By connecting attention to prior biological knowledge and without any batch information, TOSICA interpretably integrates and annotates single-cell data in a batch-insensitive manner while retaining biological variation.

    “By connecting attention to prior biological knowledge and without any batch information, TOSICA interpretably integrates and annotates single-cell data in a batch-insensitive manner while retaining biological variation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On mAtlas, TOSICA annotated query cells with accuracy of 81.06%, while ACTINN has 79.57%.

    “Notably, on the biggest dataset mAtlas, which also has the most cell types, TOSICA annotated the cells in query with a high accuracy of 81.06%, while the second best tool ACTINN has an accuracy of 79.57%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The accuracy of TOSICA on every dataset ranks at top 6, and its mean accuracy of 86.69% is the highest among all 19 methods.

    “The accuracy of TOSICA on every dataset ranks at top 6 (Supplementary Dataset 3), and its mean accuracy of 86.69% is the highest among all 19 methods (Fig. 2a).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    TOSICA achieves fast and accurate one-stop annotation.

    “We show that TOSICA achieves fast and accurate one-stop annotation and batch-insensitive integration while providing biologically interpretable insights for understanding cellular behavior during development and disease progressions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Random masks usually result in similar accuracy as knowledge-based masks, but on mPancreas accuracy converges lower with random mask.

    “Random masks usually can result in the similar accuracy as knowledge-based masks, but in the case of mPancreas dataset, the accuracy converges lower with the random mask (Supplementary Fig. 8c).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    TOSICA meets the needs for accurate cell type annotation across different datasets, exceeding existing methods in accuracy, and often does so with reduced time cost.

    “TOSICA not only met the needs for accurate cell type annotation across different datasets, exceeding existing methods in accuracy, but also often do so with reduced time cost.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When alpha cells are deleted from hPancreas reference, alpha cells in query are clustered together and 76% are labeled as 'Unknown' by TOSICA, while the rest are labeled as PP.

    “As expected, ‘alpha cells’ in the query set of hPancreas are clustered together (Fig. 3a) and 76% of them are labeled as ‘Unknown’ by TOSICA (Fig. 3b), while the rest are labeled as ‘pancreatic polypeptide cell’ (PP), which is also an endocrine cell (Fig. 3b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    When tested on many datasets, TOSICA provides the advantage to interpret the attention feature genes and pathways, and automatically filters out batch effect, potentially as a consequence of direct mapping of cell types to genes (or pathways when using a pathway mask).

    “When tested on many datasets, TOSICA provides the advantage to interpret the attention feature genes and pathways, and surprisingly also automatically filtering out batch effect, potentially as a consequence of direct mapping of cell types to genes (or pathways when using a pathway mask).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CELLBLAST and chetah label alpha cells with 99% and 62% as PP, with 0 and 37% as a new cell type, respectively.

    “On contrary, CELLBLAST and chetah, two annotators that actively identify new cell types, label ‘alpha’ cells with 99% and 62% as ‘PP’, with 0 and 37% as a new cell type, respectively (Supplementary Fig. 9d, e).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    With increase of dataset size, time cost of TOSICA on mAtlas is the fourth shortest and does not explode exponentially like most other methods.

    “Meanwhile, with the increase of the dataset size, time cost of TOSICA on mAtlas is the fourth shortest and does not explode exponentially like most of the other methods (Fig. 2c).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A rare 'MHC class II' cell type only appeared in query is annotated as 'macrophage' or 'Unknown' and clustered separately by TOSICA.

    “There is also another rare cell type only appeared in query, ‘MHC class II’ cell, and is annotated as ‘macrophage’ or ‘Unknown’ and clustered separately by TOSICA (Fig. 3a, b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TOSICA ranks fifth and sixth on hArtery and hPancreas, with accuracy of 93.75% and 95.76%, close to top-ranked methods (Seurat 96.37% for hArtery and SingleCellNet 97.53% for hPancreas).

    “Although TOSICA ranks fifth and sixth on two easy-to-classify datasets (hArtery and hPancreas), where all top six methods have above 90% accuracy, its accuracy of 93.75% and 95.76% is close to the top-ranked methods (Seurat 96.37% for hArtery and SingleCellNet 97.53% for hPancreas).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    While scGen and Seurat show excellent ability on datasets with fewer batches and cells, neither works on mouse atlas dataset, on which TOSICA ranks at the top in both batch effect removal and biological conservation.

    “Conspicuously, while scGen and Seurat show excellent ability on datasets with fewer batches and cells, neither of them works on mouse atlas dataset, on which TOSICA ranks at the top in both batch effect removal and biological conservation (Supplementary Fig. 11b).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    Seurat-based methods, including Seurat v3 CCA and Seurat v3 RPCA, are unable to integrate datasets from more than 85 batches, and are thus not applicable for comparison.

    “Note that Seurat-based methods, including Seurat v3 CCA and Seurat v3 RPCA, are unable to integrate datasets from more than 85 batches, these methods are thus not applicable for comparison.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On 2 of 5 test datasets with more cells, batch ASW of TOSICA ranks in top 2 and is only slightly lower (0.02-0.06 or 2.1-5.6%) than top 1 method's batch ASW.

    “On 2 of the 5 test datasets, which have more cells, the batch ASW of TOSICA ranks in the top 2 and is only slightly lower (0.02–0.06 or 2.1–5.6%) than the top 1 method’s batch ASW (Supplementary Fig. 11b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    MM and MP are distinguished by MIR-6382 and MIR-29B-3P regulons, with attention score of MIR-29B-3P higher in MP.

    “We find that MM and MP are distinguished by MIR-6382 and MIR-29B-3P regulons (Fig. 3d), with attention score of MIR-29B-3P higher in MP.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Principal component analysis of the original expression matrix shows that MP shares similar PCs with PP compared with other MM, with transition ordering visible on PC1.

    “Principal component analysis of the original expression matrix also shows that MP shares similar PCs with PP (reference Prlf. Acinar, TOSICA Prlf. Acinar) compared with other MM, where the transition ordering is visible on PC1 (Supplementary Fig. 11a).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TOSICA is robust against the choice of masks in its batch effect removal ability, except batch effects removal ability is unexpectedly slightly stronger when using random masks.

    “Also, TOSICA is robust against the choice of masks in its of batch effect removal ability, except batch effects removal ability is unexpectedly slightly stronger when using random masks, and it is expected that some knowledge-based masks are better than others for a specific dataset, for example for hBone (Supplementary Fig. 11b).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Among all cell types in the COVID-19 atlas, DC_LAMP3, Epi and Mast are unknown cell types for reference but TOSICA can still identify them de novo as an isolated cluster on UMAP with little batch effect.

    “Among all cell types, DC_LAMP3, Epi and Mast are unknown cell types for reference but TOSICA can still identify them de novo as an isolated cluster on UMAP (Fig. 5a and Supplementary Fig. 15a) with little batch effect (Supplementary Fig. 15b).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    INNATE_IMMUNE_SYSTEM is downregulated and INTERFERON_SIGNALING is upregulated during aging in Mono_CD14.

    “INNATE_IMMUNE_SYSTEM is downregulated and INTERFERON_SIGNALING is upregulated during aging in Mono_CD14.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The runtime of TOSICA is the shortest (minutes), while scGen takes nearly five days to finish.

    “In addition, the runtime of TOSICA is the shortest (minutes), while it takes scGen nearly five days to finish (Supplementary Fig. 13c).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Despite no batch information included in training or test sets from different batches, studies or subjects, TOSICA can consistently predict cell types with great accuracy and generate batch insensitive embedding.

    “In contrast, despite no batch information is included in either the training set or test set when they both comes from different batches, different studies or subjects (Supplementary Dataset 1), TOSICA can consistently predict cell types with great accuracy (Fig. 2a) and generate batch insensitive embedding, perhaps due to direct mapping of cell types to genes (or pathways when using a pathway”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TOSICA's TF regulon attentions in C3 and C4 show that AP2_Q6 and FOXO4_01 have low activities and AP4_01, MIR3617_5P, NFKB_Q6, and ATF3_Q6 have upregulated activities during COVID-19 disease progression.

    “TOSICA’s TF regulon attentions in C3 and C4 show that AP2_Q6 and FOXO4_01 have low activities and AP4_01, MIR3617_5P, NFKB_Q6, and ATF3_Q6 have upregulated activities during COVID-19 disease progression (Fig. 5g).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). The quote only partly states this claim.

    Findingnone given

    TOSCIA converges within 20 epochs.

    “Typically, TOSCIA converges within 20 epochs.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Transformer is successful in interpretability benefited by self-attention mechanism, which calculates the relationship (attention) between tokens of object representation.

    “Transformer is successful in interpretability benefited by self-attention mechanism, which calculates the relationship (referred to as “attention”) between tokens of object representation12.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [12] Vaswani, A. et al. Attention is all you need. In Advances in Neural Information Processing Systems Vol. 32 (NeurIPS 2017); https://papers.nips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    Through supervised training, TOSICA learns the projection function from gene expression to cell type, transferring high-dimensional and sparse expression space to low-dimensional and dense feature space.

    “Through supervised training, our model learns the projection function from gene expression to cell type, meanwhile transfers high-dimensional and sparse expression space to low-dimensional and dense feature space.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    TOSICA is not limited by expert knowledge and robust to mask choice.

    “So TOSICA is not limited by expert knowledge and robust to mask choice, and one can choose mask depending on biological context or research interests, but expert knowledge helps to converge to the best model faster.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The Transformer framework does not involve dimensionality reduction, thus keeping all attention layers traceable to the original input features, making the models interpretable.

    “However, the Transformer framework does not involve dimensionality reduction10,12, thus keeping all attention layer traceable to the original input features13, thus making the models interpretable.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [10] Ranftl, R., Bochkovskiy, A. & Koltun, V. J. a. e.-p. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 12179–12188 (ICCV, 2021). source

      Not checked in the free run.

    • cites [12] Vaswani, A. et al. Attention is all you need. In Advances in Neural Information Processing Systems Vol. 32 (NeurIPS 2017); https://papers.nips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.

      Not checked in the free run.

    • cites [13] Abnar, S. & Zuidema, W. J. a. e.-p. Quantifying attention flow in transformers. Preprint at https://arxiv.org/abs/2005.00928 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The change of dimensions and non-linear aggregation of features throughout the autoencoder's deep processing stages leads to untraceable and uninterpretable latent space and loss of information and feature resolution.

    “For example, the change of dimensions and non-linear aggregation of features throughout the autoencoder’s deep processing stages leads to untraceable and uninterpretable latent space and loss of information and feature resolution9,10.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9] Gong, C., Wang, D., Li, M., Chandra, V. & Liu, Q. J. a. e.-p. Vision transformers with patch diversification. Preprint at https://arxiv.org/abs/2104.12753 (2021).

      Not checked in the free run.

    • cites [10] Ranftl, R., Bochkovskiy, A. & Koltun, V. J. a. e.-p. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 12179–12188 (ICCV, 2021). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    An important step in scRNA-seq analysis is to identify cell populations or types by clustering.

    “An important step in scRNA-seq analysis is to identify cell populations or types by clustering1.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Sandberg, R. Entering the era of single-cell transcriptomics in biology and medicine. Nat. Methods 11, 22–24 (2014). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    With the increase of non-linear aggregation layers to achieve more powerful learning capability, the model gets deeper and the contribution from input gets harder to trace, leading to loss of interpretability.

    “In addition, with the increase of non-linear aggregation layers to achieve more powerful learning capability, the model gets deeper meanwhile the contribution from input gets harder to trace, which leads to the loss of interpretability11.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [11] Li, X. et al. Interpretable deep learning: interpretation, interpretability, trustworthiness, and beyond. Knowl Inf Syst 64, 3197–3234 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    TOSICA has an acceptable time cost on large datasets, performs better than any other methods on tough tasks, making it a universal cell type annotator.

    “Altogether, TOSICA has an acceptable time cost on large datasets, while performs better than any other methods on tough tasks, making it a universal cell type annotator.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Across different datasets, the same cell may have different marker gene sets, thus annotated differently.

    “Thus, across different datasets, the same cell may have different marker gene sets, thus annotated differently.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Specificity of marker genes is determined by comparing with the other cell types in the same dataset.

  • May combine two claims.

    Premisecitation

    The homolog of CEBP has been reported to inhibit proliferation of mouse chondrocytes in vitro, and Nf1 ablation in Fgfr1Col2cKO mice reverses their hypertrophic zone phenotype.

    “Indeed, the homolog of CEBP has been reported to inhibit proliferation of mouse chondrocytes in vitro45, and Nf1 ablation in Fgfr1Col2cKO mice reverses their hypertrophic zone phenotype46.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [45] Okuma, T. et al. Regulation of mouse chondrocyte differentiation by CCAAT/enhancer-binding proteins. Biomed. Res. 36, 21–29 (2015). source

      Not checked in the free run.

    • cites [46] Karolak, M. R., Yang, X. & Elefteriou, F. FGFR1 signaling in hypertrophic chondrocytes is attenuated by the Ras-GAP neurofibromin during endochondral bone formation. Hum. Mol. Genet. 24, 2552–2564 (2015). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Due to the good interpretability of attention score, it can well reconstruct the trajectory and reveal the key pathways in the biological process.

    “Due to the good interpretability of attention score, it can well reconstruct the trajectory and reveal the key pathways in the biological process.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The interpretability does not stop at the high-level structures; important low-level entities, genes, that significantly contribute to these high-level annotations are also available from the networks within TOSICA, and can be revealed simultaneously thus generating a comprehensive hierarchical annotation structure.

    “Furthermore, as shown by the discovery of MIR-29B-3P regulon (Fig. 3d) and its important target gene Sparc in the development of acinar cell (Supplementary Fig. 10a), the interpretability does not stop at the high-level structures, the important low-level entities, genes, that significantly contribute to these high-level annotations are also available from the networks within TOSICA (Supplementary”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    By embedding genes to higher level of biological processes tokens, TOSICA directly learns the biological processes and signaling pathways giving rise to the cell types, thus separating cell types, including new cell types with accurate high-resolution annotation and allowing direct trajectory regulation discovery, while immune to batch effect.

    “Instead, by embedding genes to higher level of biological processes tokens, TOSICA directly learns the biological processes and signaling pathways giving rise to the cell types, thus separating cell types, including new cell types (Fig. 2d) with accurate high-resolution annotation (Fig. 3d) and allowing direct trajectory regulation discovery (Fig. 3g), while immune to batch effect (Supplementary”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    This high-level attention framework not only allows interpretability but is essential for the high accuracy of TOSICA.

    “This high-level attention framework not only allows interpretability but is essential for the high accuracy of TOSICA (Fig. 2a).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Previous observations show low expression of TLR signaling genes and low innate immune activity of cDC3_LAMP3 and pro-inflammatory properties of cDC2_FCN1 and cDC2_IL1B in blood.

    “This is in agreement with previous observations showing low expression of Toll-like receptor (TLR) signaling genes and low innate immune activity of cDC3_LAMP348 and the “pro-inflammatory” properties of cDC2_FCN1 and cDC2_IL1B in blood51.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [48] Cheng, S. et al. A pan-cancer single-cell transcriptional atlas of tumor infiltrating myeloid cells. Cell 184, 792–809 e723 (2021). source

      Not checked in the free run.

    • cites [51] Dutertre, C. A. et al. Single-cell analysis of human mononuclear phagocytes reveals subset-defining markers and identifies circulating inflammatory dendritic cells. Immunity 51, 573–589 e578 (2019). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    TOSICA's gene set attentions automatically distinguish cells originally labeled as Mat. Acinar and Prlf. Acinar, and further identified an intermediate state between the two, which is closer to Prlf. Acinar and incorrectly labeled as Mat. Acinar in the annotation database.

    “Thus, TOSICA’s gene set attentions automatically distinguish cells originally labeled as Mat. Acinar and Prlf. Acinar, and further identified an intermediate state between the two, which is closer to Prlf. Acinar and incorrectly labeled as Mat. Acinar in the annotation database.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    With high biological resolution and batch insensitivity of attention, TOSICA reveals many novel dynamic and functional status of single cells, with their key contributors hierarchically annotated at both pathway and gene levels.

    “With high biological resolution and batch insensitivity of attention (Supplementary Fig. 12a), TOSICA reveals many novel dynamic and functional status of single cells, with their key contributors hierarchically annotated at both pathway and gene levels to guide further experimental explorations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    These results indicate that TOSICA is insensitive to batch effect and good at biological conservation, and excels on large datasets with many batches, especially considering that we never provide batch information to it.

    “These results indicate that TOSICA is insensitive to batch effect and good at biological conservation, and excels on large datasets with many batches, especially considering that we never provide batch information to it.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    These monocyte subtypes have their own biological signatures and potentially different functions.

    “Furthermore, benefiting from its high resolution, TOSICA identifies several subtypes of monocytes that have not been discovered in the original publication48 (Fig. 4h), all having their own biological signatures and potentially different functions (Supplementary Fig. 12e).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Various systems level comparisons with existing methods and case-by-case close examination of different datasets and tasks demonstrate the accuracy, robustness, flexibility, and generalizability of TOSICA as an indispensable new tool for advancing single-cell studies.

    “The various systems level comparisons with existing methods and case-by-case close examination of different datasets and tasks demonstrate the accuracy, robustness, flexibility, and generalizability of TOSICA as an indispensable new tool for advancing the single-cell studies.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    TOSICA is a Transformer-based cell type annotation and integration tool that offers accurate, transferrable, high-resolution, batch insensitive, biologically interpretable cell type annotations under many scenarios, including new cell type discovery, dynamic trajectory analysis, cross platform, and population dataset integration.

    “In this study, we develop and establish TOSICA, a Transformer-based cell type annotation and integration tool that offers accurate, transferrable, high-resolution, batch insensitive, biologically interpretable cell type annotations under many scenarios, including but not limited to new cell type discovery, dynamic trajectory analysis, cross platform, and population dataset integration.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Only data with poor labels were excluded from the analyses.

    “Only data with poor labels were excluded from the analyses.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The experiments were not randomized.

    “The experiments were not randomized.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    No statistical method was used to predetermine sample size.

    “No statistical method was used to predetermine sample size.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The investigators were not blinded to allocation during experiments and outcome assessment.

    “The investigators were not blinded to allocation during experiments and outcome assessment.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The intelligible and interpretable high-resolution annotation transfer between completely independent studies on different biological processes is demonstrated, allowing interdisciplinary data integration of single-cell studies.

    “The intelligible and interpretable high-resolution annotation transfer between completely independent studies on different biological processes is demonstrated, thus allowing interdisciplinary data integration of single-cell studies.”
    Quote in the paperSubstantial claim
51 of 51

  • The paper relies on this without support.

    Consistent annotation transfer from reference dataset to query dataset is fundamental to the development and reproducibility of single-cell research.

    “Consistent annotation transfer from reference dataset to query dataset is fundamental to the development and reproducibility of single-cell research.”

  • The paper relies on this without support.

    Deep learning based methods are faster and more automated than traditional annotation methods.

    “Compared with traditional annotation methods, deep learning based methods are faster and more automated.”

  • The paper relies on this without support.

    Traditional annotation methods often consist of preprocessing, dimensionality reduction, clustering, differential analysis, and manual annotation based on prior knowledge.

    “These traditional methods often consist of preprocessing, dimensionality reduction, clustering, differential analysis, and manual annotation based on prior knowledge.”

  • The paper relies on this without support.

    When subtypes are annotated manually based on a small set of marker genes, the same subtype can sometimes be given a new name in another research due to a slight difference.

    “When subtypes are annotated manually based on a small set of marker genes, the same subtype can sometimes be given a new name in another research due to a slight difference.”

  • The paper relies on this without support.

    Transferring cell type annotation from a reference dataset to newly generated query datasets with consistency and reproducibility is increasingly important and necessary.

    “Thus, transferring cell type annotation from a reference dataset to newly generated query datasets with consistency and reproducibility is increasingly important and necessary.”

  • The paper relies on this without support.

    Specificity of marker genes is determined by comparing with the other cell types in the same dataset.

    “However, specificity of marker genes is determined by comparing with the other cell types in the same dataset.”

  • The paper relies on this without support.

    All previous cell type annotators are gene-based, thus reveal little on the biological insight behind the cell type marker genes, many more subsequent analyses are needed to infer the potential enriched pathways and regulators behind the marker genes.

    “All previous cell type annotators are gene-based, thus reveal little on the biological insight behind the cell type marker genes, many more subsequent analyses are needed to infer the potential enriched pathways and regulators behind the marker genes.”

  • The paper relies on this without support.

    One of the most common demands in single-cell analysis is the transfer of identified cell population from an original reference to newly generated data, which may come from different batch and biological state.

    “One of the most common demands in single-cell analysis is the transfer of identified cell population from an original reference to newly generated data, which may come from different batch and biological state (e.g., disease).”

  • The paper relies on this without support.

    Query and reference datasets are generated in different laboratories with different experimental protocols and thus contain batch effects.

    “Generally, query and reference datasets are generated in different laboratories with different experimental protocols and thus contain batch effects.”

  • The paper relies on this without support.

    Batch information is necessary for conventional data-integration method to try to overcome these batch effects.

    “Batch information is necessary for conventional data-integration method to try to overcome these batch effects, which are difficult to completely remove and mixed up with biological differences.”
10 of 10

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The pancreas dataset contained human pancreas cells from five different scRNA-seq studies, with train and test sets from different datasets.

    “The pancreas dataset contained human pancreas cells from five different scRNA-seq studies, which were reprocessed by Chen at al. [20].”

    From A Deep Dive into Single-Cell RNA Sequencing Foundation Models (2023) · cites this paper as [20]

Self-check

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

Each contradiction between two passages listed here was confirmed by an independent check.

  • COVID-19 monocyte subtype trends are reversed

    Major

    The figure caption says subtype 3 increases and subtype 4 decreases from healthy to severe COVID-19, but the Results text says C3 decreases and C4 increases over the same progression.

  • Eight integration metrics claimed, but only seven listed

    Minor

    The myeloid analysis claims eight evaluation metrics, but names three for batch-effect removal and four for biological-variation retention. ASW appears in both groups, so the list contains only seven entries.

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Chen, J., Xu, H., Tao, W., Chen, Z., Zhao, Y., & Han, J.-D. J.
Affiliations
Peking University (CN)
Corresponding author
Han, Jing-Dong J.
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2023-01-14
Volume
14
Issue
1
Pages
223-223
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-023-35923-4
PMC
PMC9840170
OpenAlex
W4316096017
Cited by
189
References
59
Citations by year
2026: 36 · 2025: 69 · 2024: 70 · 2023: 12 · 2022: 1
FWCI
16.66
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Natural Science Foundation of China (92049302, 32088101)
  • Ministry of Science and Technology of the People’s Republic of China (2020YFA0804000)
  • Science and Technology Commission of Shanghai Municipality (2017SHZDZX01)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene expression and cancer classification
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
Keywords
cell type annotation, single-cell transcriptomics, transformer models, self-attention, interpretability, annotation transfer, multi-head attention, regulon, tumor-infiltrating immune cells, CD14+ monocytes, COVID-19 pandemic, rare cell types, cellular heterogeneity, disease progression
MeSH terms
Exome Sequencing, COVID-19, Humans, Reproducibility of Results, Sequence Analysis, RNA, Disease Progression, 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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.

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