Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

scMamba: A Pre-Trained Model for Single-Nucleus RNA Sequencing Analysis in Neurodegenerative Disorders

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read scMamba, a pre-trained bidirectional Mamba model, outperforms benchmark methods on cell type annotation, doublet detection, imputation, and differential expression analysis for brain snRNA-seq.

desk verdict Solid empirical extension of the authors' scHyena; the gains are real but modest, and the abstract oversells them. read the letter →

arxiv 2502.19429 v1 pith:22CGABMD submitted 2025-02-12 q-bio.GN cs.LG

Gyutaek Oh , Baekgyu Choi , Seyoung Jin , Inkyung Jung , Jong Chul Ye This is my paper · ORCID
classification q-bio.GNcs.LG
keywords single-nucleusRNAsequencingfoundationmodelMambastatespacecelltypeannotationdoubletdetectionexpressionimputationdifferentialneurodegenerativedisease
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

scMamba is a pre-trained model for single-nucleus RNA sequencing that aims to make brain snRNA-seq analysis work without the usual compromises: no selection of highly variable genes and no discretization or dimension reduction. The paper argues that a bidirectional Mamba architecture with learned gene embeddings, pre-trained by masked expression modeling on roughly 1.6 million human brain nuclei, learns generalizable cell and gene representations. Across five Alzheimer's and Parkinson's snRNA-seq datasets, the authors report that scMamba matches or beats existing tools on cell type, subtype, and subcluster classification, doublet detection, expression imputation, and the reproducibility of differential gene expression. A sympathetic reader would care because postmortem brain data are noisy and heterogeneous, and a model that uses the full transcriptome could make cross-dataset integration and disease discovery more reliable.

What carries the argument

The load-bearing object is the bidirectional Mamba block built on selective state space models (S6), where the transition parameters B, C, and the step size depend on the input, letting the model selectively keep or discard gene information. Expression levels are encoded through a linear adapter layer into continuous embeddings, and gene embeddings are added in place of positional encodings; three [CLS] tokens are inserted at the beginning, middle, and end for classification. Pre-training uses masked expression modeling on nonzero entries, and imputation fine-tuning uses different masking probabilities for zero (0.04) and nonzero (0.4) values. This combination is what lets the model see the full transcriptome at linear cost and carry gene identity through the sequence.

What would settle it

Take a snRNA-seq dataset in which cell identities and doublets are independently validated by orthogonal methods such as cell hashing, species-mixed nuclei, or in situ RNA detection; if scMamba's F1, imputation MSE, and DEG overlap advantages over scBERT and scHyena vanish or reverse on those true labels, the central claim that full-transcriptome bidirectional Mamba pretraining improves snRNA-seq analysis would be falsified.

Watch

Extended reading notes

Core claim

The central discovery is that a state-space-sequence model, Mamba, can serve as a foundation model for snRNA-seq when adapted with two design choices: a linear adapter that keeps expression values continuous, and per-gene embeddings that give each position its gene identity. The authors claim that pre-training with masked expression modeling on roughly 1.6 million nuclei yields cell embeddings that cluster by cell type and gene embeddings that group marker genes, and that fine-tuning this model gives the highest or near-highest F1 scores in subcluster classification, the best recall, F1, AUROC, and AUPRC in doublet detection, low-MSE imputation that also reduces batch effects, and higher DEG overlap under subsampling. The claim is that these gains come from processing all 19,306 genes without highly variable gene selection, using bidirectional Mamba's linear-time, input-dependent mixing.

Load-bearing premise

The cell type labels and in vivo doublet labels used for fine-tuning and evaluation come from computational pipelines, namely Leiden clusters assigned by marker expression after Harmony integration and Scrublet doublet scores thresholded at 0.1, so if those proxy labels are systematically wrong, the reported accuracies measure agreement with the labeling pipeline rather than biological identity.

Editorial extensions

If this is right

  • scMamba can be used as a drop-in pretrained encoder for brain snRNA-seq datasets, improving annotation of fine-grained subtypes without requiring users to select highly variable genes.
  • Because the architecture scales linearly with sequence length, whole-transcriptome analysis becomes practical at far larger cell counts than transformer-based single-cell foundation models.
  • The asymmetric masking strategy for imputation provides a concrete recipe for separating true zeros from dropout zeros in other single-cell modalities.
  • The reported improvement in DEG overlap under subsampling implies that imputation with scMamba should make multi-cohort disease studies more reproducible when sample sizes are limited.
  • Fine-tuning the same pretrained weights for classification, doublet detection, and imputation suggests that one pretrained model can replace several task-specific pipelines in snRNA-seq workflows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The gain from gene embeddings likely generalizes beyond brain tissue, since the same positional-encoding replacement could transfer to any gene-expression modality where gene identity, not order, carries meaning; this is an untested extension of the paper's design logic.
  • The optimal zero-versus-nonzero masking ratio (0.04 versus 0.4) is a data-dependent choice; varying sequencing depth or dropout rate across datasets would reveal whether the model is learning a biological dropout prior or a technical artifact.
  • The choice of three [CLS] tokens is motivated only empirically, so ablating their number and positions would clarify whether this is an essential mechanism or an incidental benefit of fine-tuning.
  • Because baseline models were pretrained on the same data, the comparison isolates architecture; an unexplored question is whether the ~1.6 million-cell scale, rather than the Mamba block, drives the performance gap, testable by pretraining scMamba on smaller subsets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript introduces scMamba, a bidirectional Mamba-based model for single-nucleus RNA-seq (snRNA-seq) analysis, pre-trained by masked expression modeling on roughly 1.6 million human brain nuclei. The architecture combines a linear adapter for continuous expression values with learned gene embeddings and bidirectional Mamba blocks, and is fine-tuned for cell type annotation, doublet detection, imputation, and differential expression reproducibility. The evaluation covers five brain snRNA-seq datasets and compares against Seurat, SciBet, scBERT, scHyena, and several doublet-detection baselines. The authors report that scMamba achieves the highest or near-highest scores in most settings and state in the abstract that it outperforms benchmark methods across downstream tasks.

Significance. If the empirical results hold, scMamba is a useful addition to the single-cell foundation-model literature: it processes full-length continuous expression without HVG selection, is pre-trained on a large neurodegeneration-oriented brain corpus, and is evaluated on downstream datasets that are disjoint from the pre-training data. The simulated-doublet experiments provide a helpful external check for the doublet detection claim. The contribution is primarily empirical rather than architectural or theoretical, and the paper is generally careful in using official baseline implementations. However, the reported performance advantage is uneven across datasets and metrics, and the cell-type and in-vivo doublet labels used for both training and evaluation come from the authors' own computational pipeline, so the broad outperformance claim needs qualification and the label-validity risk should be addressed explicitly.

major comments (3)
  1. [Abstract; Supplementary Table 4; Fig. 6; Discussion] The abstract's claim that 'scMamba outperforms benchmark methods in various downstream tasks' is broader than the reported results. In Supplementary Table 4, scHyena has a higher macro F1 than scMamba on Smajic cell-type classification (0.9843 vs. 0.9742), and scBERT has higher micro F1 than scMamba in Leng subtype classification (0.7703 vs. 0.7641) and in Jung cell-type classification (0.9791 vs. 0.9766). In the zero-imputation experiment, the text states that scHyena 'achieves the highest NMI and ARI metric values.' In doublet detection, the Discussion explicitly concedes that DoubletDetection surpasses scMamba in precision and TNR. The central claim should be qualified to state that scMamba is competitive with, and often better than, the baselines, or the abstract should be revised to reflect the metrics in which it does not dominate.
  2. [Methods: 'Cell type annotation based on unsupervised clustering'; Results: 'scMamba is capable of classifying sub cell…] The cell-type, subtype, and subcluster labels are produced by the authors' own clustering pipeline (Harmony integration, Leiden at resolution 1.8, marker-gene inspection with a 40-gene panel, with one cluster labeled 'unidentified'), and these same pipeline-derived labels are used both for fine-tuning and for computing every reported classification F1 score. There is no external validation of these labels, such as an independent atlas-based annotation, in situ hybridization, or immunohistochemistry. Consequently, the classification results measure agreement with the authors' label-generation pipeline, and if the label noise is correlated with model architecture (for example, through gene-weighting biases), the reported advantages may not transfer to genuinely independent labels. The authors should either add an external label validation or explicitly discuss this limitation and provide a robustness analysis under label perturbation.
  3. [Results: 'scMamba effectively identifies and filters doublets'; Methods: 'Data preparation'] For in-vivo doublet detection, the ground-truth labels appear to be the same Scrublet calls (average doublet score > 0.1) that were used to annotate doublet-enriched clusters during data preparation, and it is not stated whether the same cells were used to fine-tune scMamba for this task. The simulated-doublet experiments provide an external check, but simulated doublets are constructed by averaging pairs of singlets and may not capture the full complexity of real doublets, such as same-type doublets or degraded nuclei. The authors should clarify whether any training labels were derived from the same Scrublet calls and should report the in-vivo and simulated results separately in the main text, rather than only in heatmaps, so that the reader can assess how much of the reported AUROC and AUPRC advantage depends on synthetic doublets.
minor comments (6)
  1. [Methods: 'State space models', Eq. (2)] The expression for the discretized B matrix is improperly parenthesized: it should read B = (ΔA)^{-1}(exp(ΔA) − I)ΔB, and the current text appears to place the factor I·ΔB outside the parentheses.
  2. [Fig. 1c caption] The last sentence of the caption duplicates the preceding sentence: 'Both zero and non-zero values are masked with different masking probabilities' is stated twice and should be reduced to one statement.
  3. [Results: 'scMamba improves robustness in DEG analysis'] The text contains the unresolved placeholder 'Smajic dataset (ref)'; this should be replaced with the proper citation (reference 15).
  4. [General] The manuscript does not include a data or code availability statement. Given that the study introduces a pre-trained model and reports a large-scale benchmark, releasing the pretraining code, fine-tuning code, and trained weights is important for reproducibility and should be stated.
  5. [Supplementary Table 4] The table would be easier to interpret if the best score in each row (or column block) were bolded and if the number of wins/losses against each baseline were summarized in the text, since the current narrative relies on qualitative descriptions of a large table.
  6. [Fig. 5b] The panels labeled 'histograms' show summary values (means over five groups) rather than distributions; 'bar plots' would be a more accurate description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: scMamba's claims are empirical benchmark results with held-out evaluation, separated pre-training data, and an independent simulated-doublet check; the only self-citation (scHyena) appears as a controlled baseline.

full rationale

scMamba's claims are empirical benchmark results, not derivations from first principles. The evaluation protocol separates pre-training data from downstream data: "Among the data, Lau, Leng, Smajic, Zhu, and Jung datasets were selected for downstream task, while the remaining 10 datasets were used for pre-training scMamba." The only self-citation (scHyena, Ref. 38, by overlapping authors) is used as a comparison baseline and is explicitly pre-trained on the same data as scMamba ("scBERT and scHyena were pre-trained on the same dataset used for pre-training scMamba"), so it is a controlled baseline rather than load-bearing support. The doublet experiments directly address the proxy-label concern: "we annotated doublets using Scrublet to establish ground truth for doublet detection. However, the labels generated by Scrublet may not always be accurate. To ensure a fairer comparison, we also conducted experiments using simulated doublets." The simulated-doublet results provide an independent external check. Cell-type labels are computationally derived from a marker-gene-plus-Leiden pipeline, but they are defined independently of scMamba's outputs, so shared labels create a benchmark-validity limitation rather than a circular derivation. No equation in the paper defines a predicted quantity in terms of a fitted input, and no load-bearing argument rests on a self-citation chain. Therefore no significant circularity is present.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

This is an empirical systems paper, not a derivation. The load-bearing choices are hand-set hyperparameters, computational label pipelines, and the MSE loss assumption. No new physical or biological entities are postulated.

free parameters (6)
  • Pre-training mask probability = 0.15
    Set by hand following BERT-style masking; only nonzero expressions are masked. Affects what the model learns.
  • Imputation mask probabilities = 0.4 nonzero, 0.04 zero
    Chosen to balance zero/nonzero masking during imputation fine-tuning; central to imputation results.
  • Number of CLS embeddings = 3 (beginning, middle, end)
    Authors state that three CLS embeddings 'enhanced performance'; no principled derivation.
  • Gene detection threshold = 0.5% of cells (Kamath excluded)
    Filters 61,325 Ensembl IDs to 19,306 genes; arbitrary and affects all downstream inputs.
  • UMI count filter for imputation = 4,000 UMIs
    Only cells with UMI > 4,000 are used for imputation fine-tuning and evaluation, removing low-quality cells.
  • Label-clustering settings = top 2,000 HVGs, 20 PCs, 40 neighbors, Leiden resolution 1.8
    Used with Harmony to produce the cell type labels that define classification ground truth; changes here change all classification scores.
assumptions (4)
  • domain assumption MSE loss is appropriate for log-normalized expression values
    Pretraining and imputation use squared error on log(x+1) normalized counts, implying Gaussian noise, while UMI counts are typically modeled as negative binomial. This choice shapes imputation behavior.
  • domain assumption Scrublet doublet scores provide valid doublet ground truth
    In vivo doublet detection is evaluated against Scrublet labels, the same tool used to annotate doublets during preprocessing (Methods, Data preparation). Simulated doublets partially mitigate this.
  • domain assumption Leiden clusters with Harmony and marker genes provide valid cell type labels
    All classification fine-tuning and evaluation rely on these computationally derived labels; there is no orthogonal experimental validation.
  • domain assumption Higher DEG overlap after imputation indicates improved data quality
    DEG analysis measures reproducibility of DEG lists under subsampling, not agreement with true expression or true DEGs; smoothing could inflate overlap.

how reviews work

0 comments
Cite this review

Pith. "Pith review of scMamba: A Pre-Trained Model for Single-Nucleus RNA Sequencing Analysis in Neurodegenerative Disorders." pith.science (2026). https://pith.science/paper/22CGABMD

@misc{pith2026250219429,
  author       = {Pith},
  title        = {Pith review of: scMamba: A Pre-Trained Model for Single-Nucleus RNA Sequencing Analysis in Neurodegenerative Disorders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22CGABMD}},
  note         = {Machine review of arXiv:2502.19429}
}
read the original abstract

Single-nucleus RNA sequencing (snRNA-seq) has significantly advanced our understanding of the disease etiology of neurodegenerative disorders. However, the low quality of specimens derived from postmortem brain tissues, combined with the high variability caused by disease heterogeneity, makes it challenging to integrate snRNA-seq data from multiple sources for precise analyses. To address these challenges, we present scMamba, a pre-trained model designed to improve the quality and utility of snRNA-seq analysis, with a particular focus on neurodegenerative diseases. Inspired by the recent Mamba model, scMamba introduces a novel architecture that incorporates a linear adapter layer, gene embeddings, and bidirectional Mamba blocks, enabling efficient processing of snRNA-seq data while preserving information from the raw input. Notably, scMamba learns generalizable features of cells and genes through pre-training on snRNA-seq data, without relying on dimension reduction or selection of highly variable genes. We demonstrate that scMamba outperforms benchmark methods in various downstream tasks, including cell type annotation, doublet detection, imputation, and the identification of differentially expressed genes.

Figures

Figures reproduced from arXiv: 2502.19429 by the authors.

Figure 1
Figure 1. Overall framework of scMamba. a. The scMamba model comprises a linear layer for expression embeddings, gene embeddings, and bidirectional Mamba blocks. During pre-training, a subset of input data is masked, and the model predicts the expression levels at the masked positions. b. For fine-tuning classifica￾tion tasks, three [CLS] embeddings are inserted into the input embeddings. These embeddings are processed throug… view at source ↗
Figure 2
Figure 2. a. To generate cell embeddings from the pre-trained model, snRNA-seq data is input into the model, and the resulting output features are averaged along the sequence length. b. UMAP visualization of cell embed￾dings from the pre-trained scMamba model. Each UMAP is colored based on 8 major cell types or 72 subtypes. c. UMAP visualization of gene embeddings from pre-trained scMamba model. Marker genes of 4 distinct cel… view at source ↗
Figure 3
Figure 3. a. The dataset is labeled with 8 major cell types and 72 detailed subtypes. b. F1 score distribution across 8 cell types, with each box plot representing results for individual datasets. c. F1 score distribution across 72 subtypes, with each box plot representing results for individual datasets. d. F1 score distribution across 127 subclusters, with each box plot representing results for individual datasets. cal leve… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: a. Simulated doublets are generated by averaging the UMI counts of two randomly selected singlets. b. Heatmap of evaluation metric scores for in vivo doublet detection by each method across datasets. White squares indicate where the method failed to execute. c. Heatmap…
Figure 5
Figure 5. Figure 5: Results of nonzero imputation experiments. a. Joint plots comparing true and imputed values predicted by various imputation methods on the Jung dataset. The x and y-axis represent true values and imputed values, respectively. Values in the upper left corner of each joi…
Figure 6
Figure 6. Figure 6: Results of zero imputation experiments. UMAP visualizations of raw and imputed snRNA-seq data for the Leng dataset. The figures are labeled with cell type, subtype, and batch (patient), respectively. in imputation tasks. To further evaluate the imputation performance o…
Figure 7
Figure 7. Figure 7: Evaluating scMamba performance in single-cell differential expression analysis. a. The overlap ratio represents the intersection of DEGs between 50% subsampled data and the original 100% dataset. -log10 P-values were calculated using a paired t-test for each overlap ra…
Figure 8
Figure 8. Figure 8: State space models (SSMs) and Mamba. a. Selective SSMs dynamically adjust their parameters based on input data, enabling selective retention of information relevant to the input. It allows them to retain information based on input selectively. b. Selective state space …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Mamba for Wireless Communications and Networking: Principles and Opportunities

    cs.NI 2025-08 conditional novelty 4.0 of 10

    The paper argues Mamba can improve efficiency and performance in wireless tasks, backed by two small case studies with mixed results.

Reference graph

Works this paper leans on

62 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    J., Gorski, S

    Saliba, A.-E., Westermann, A. J., Gorski, S. A. & V ogel, J. Single-cell RNA-seq: Advances and future challenges. Nucleic acids research 42, 8845–8860 (2014)

    work page 2014

  2. [2]

    E., Maartens, A., Hupalowska, A., Teichmann, S

    Rood, J. E., Maartens, A., Hupalowska, A., Teichmann, S. A. & Regev, A. Impact of the human cell atlas on medicine. Nature medicine 28, 2486–2496 (2022). 26

    work page 2022

  3. [3]

    Li, C. et al. SciBet as a portable and fast single cell type identifier. Nature communications 11, 1818 (2020)

    work page 2020

  4. [4]

    Integrated analysis of multimodal single-cell data

    Hao, Y .et al. Integrated analysis of multimodal single-cell data. Cell 184, 3573–3587 (2021)

    work page 2021

  5. [5]

    Villani, A.-C. et al. Single-cell RNA-seq reveals new types of human blood dendritic cells, monocytes, and progenitors. Science 356, eaah4573 (2017)

    work page 2017

  6. [6]

    Jaitin, D. A. et al. Massively parallel single-cell RNA-seq for marker-free decomposition of tissues into cell types. Science 343, 776–779 (2014)

    work page 2014

  7. [7]

    & Satija, R

    Papalexi, E. & Satija, R. Single-cell RNA sequencing to explore immune cell heterogeneity. Nature Reviews Immunology 18, 35–45 (2018)

    work page 2018

  8. [8]

    Kinker, G. S. et al. Pan-cancer single-cell RNA-seq identifies recurring programs of cellular heterogeneity. Nature genetics 52, 1208–1218 (2020)

    work page 2020

Show all 62 references
  1. [9]

    Saunders, A. et al. Molecular diversity and specializations among the cells of the adult mouse brain. Cell 174, 1015–1030 (2018)

    2018

  2. [10]

    Hodge, R. D. et al. Conserved cell types with divergent features in human versus mouse cortex. Nature 573, 61–68 (2019)

    2019

  3. [11]

    Keren-Shaul, H. et al. A unique microglia type associated with restricting development of Alzheimer’s disease. Cell 169, 1276–1290 (2017)

    2017

  4. [12]

    Mathys, H. et al. Single-cell transcriptomic analysis of Alzheimer’s disease. Nature 570, 332–337 (2019)

    2019

  5. [13]

    Habib, N. et al. Disease-associated astrocytes in Alzheimer’s disease and aging. Nature neuroscience 23, 701–706 (2020). 27

    2020

  6. [14]

    Leng, K. et al. Molecular characterization of selectively vulnerable neurons in Alzheimer’s disease. Nature neuroscience 24, 276–287 (2021)

    2021

  7. [15]

    Smaji ´c, S. et al. Single-cell sequencing of human midbrain reveals glial activation and a Parkinson-specific neuronal state. Brain 145, 964–978 (2022)

    2022

  8. [16]

    Huang, M. et al. SA VER: Gene expression recovery for single-cell RNA sequencing.Nature methods 15, 539–542 (2018)

    2018

  9. [17]

    Li, W. V . & Li, J. J. An accurate and robust imputation method scImpute for single-cell RNA- seq data. Nature communications 9, 997 (2018)

    2018

  10. [18]

    Van Dijk, D. et al. Recovering gene interactions from single-cell data using data diffusion. Cell 174, 716–729 (2018)

    2018

  11. [19]

    & Garmire, L

    Arisdakessian, C., Poirion, O., Yunits, B., Zhu, X. & Garmire, L. X. DeepImpute: An accurate, fast, and scalable deep neural network method to impute single-cell RNA-seq data. Genome biology 20, 1–14 (2019)

    2019

  12. [20]

    M., Mircea, M., Mueller, N

    Eraslan, G., Simon, L. M., Mircea, M., Mueller, N. S. & Theis, F. J. Single-cell RNA-seq denoising using a deep count autoencoder. Nature communications 10, 390 (2019)

    2019

  13. [21]

    Aran, D. et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nature immunology 20, 163–172 (2019)

    2019

  14. [22]

    Yang, F. et al. scBERT as a large-scale pretrained deep language model for cell type annotation of single-cell RNA-seq data. Nature Machine Intelligence 4, 852–866 (2022)

    2022

  15. [23]

    Bert: Pre-training of deep bidirectional Transformers for language understanding

    Devlin, J. Bert: Pre-training of deep bidirectional Transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018). 28

    2018 arXiv

  16. [24]

    Brown, T. et al. Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901 (2020)

    2020

  17. [25]

    Bommasani, R. et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    2021 arXiv

  18. [26]

    Ramesh, A. et al. Zero-shot text-to-image generation. In International conference on machine learning, 8821–8831 (Pmlr, 2021)

    2021

  19. [27]

    Theodoris, C. V . et al. Transfer learning enables predictions in network biology. Nature 618, 616–624 (2023)

    2023

  20. [28]

    Cui, H. et al. scgpt: toward building a foundation model for single-cell multi-omics using generative ai. Nature Methods 1–11 (2024)

    2024

  21. [29]

    Hao, M. et al. Large-scale foundation model on single-cell transcriptomics. Nature Methods 1–11 (2024)

    2024

  22. [30]

    & Dao, T

    Gu, A. & Dao, T. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling (2024)

    2024

  23. [31]

    Huang, J. et al. MambaMIR: An arbitrary-masked mamba for joint medical image reconstruc- tion and uncertainty estimation. arXiv preprint arXiv:2402.18451 (2024)

    2024 arXiv

  24. [32]

    Liu, Y . et al. VMamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166 (2024)

    2024 arXiv

  25. [33]

    Qiao, Y . et al. VL-Mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600 (2024)

    2024 arXiv

  26. [34]

    Caduceus: Bi-directional equivariant long-range DNA sequence modeling

    Schiff, Y .et al. Caduceus: Bi-directional equivariant long-range DNA sequence modeling. In First Workshop on Long-Context Foundation Models @ ICML 2024(2024). 29

    2024

  27. [35]

    Zhu, L. et al. Vision Mamba: Efficient visual representation learning with bidirectional state space model. In Forty-first International Conference on Machine Learning(2024)

    2024

  28. [36]

    Guo, H. et al. MambaIR: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, 222–241 (Springer, 2025)

    2025

  29. [37]

    & Melville, J

    McInnes, L., Healy, J. & Melville, J. UMAP: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  30. [38]

    Oh, G., Choi, B., Jung, I. & Ye, J. C. schyena: Foundation model for full-length single-cell rna-seq analysis in brain. arXiv preprint arXiv:2310.02713 (2023)

    2023 arXiv

  31. [39]

    Choromanski, K. M. et al. Rethinking attention with Performers. In International Conference on Learning Representations (2021)

    2021

  32. [40]

    Poli, M. et al. Hyena hierarchy: Towards larger convolutional language models. In Interna- tional Conference on Machine Learning, 28043–28078 (PMLR, 2023)

    2023

  33. [41]

    L., Lopez, R

    Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: Computational identification of cell dou- blets in single-cell transcriptomic data. Cell systems 8, 281–291 (2019)

    2019

  34. [42]

    Xi, N. M. & Li, J. J. Benchmarking computational doublet-detection methods for single-cell rna sequencing data. Cell systems 12, 176–194 (2021)

    2021

  35. [43]

    S., Murrow, L

    McGinnis, C. S., Murrow, L. M. & Gartner, Z. J. DoubletFinder: Doublet detection in single- cell RNA sequencing data using artificial nearest neighbors. Cell systems 8, 329–337 (2019)

    2019

  36. [44]

    J., Sharma, R

    Gayoso, A., Shor, J., Carr, A. J., Sharma, R. & Pe’er, D. DoubletDetection (version v3.0) (2020)

    2020

  37. [45]

    Bais, A. S. & Kostka, D. scds: Computational annotation of doublets in single-cell RNA sequencing data. Bioinformatics 36, 1150–1158 (2020). 30

    2020

  38. [46]

    Bernstein, N. J. et al. Solo: Doublet identification in single-cell RNA-seq via semi-supervised deep learning. Cell systems 11, 95–101 (2020)

    2020

  39. [47]

    Attention is all you need

    Vaswani, A. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  40. [48]

    Agarwal, D. et al. A single-cell atlas of the human substantia nigra reveals cell-specific path- ways associated with neurological disorders. Nature communications 11, 4183 (2020)

    2020

  41. [49]

    Gerrits, E. et al. Distinct amyloid- β and tau-associated microglia profiles in Alzheimer’s disease. Acta neuropathologica 141, 681–696 (2021)

    2021

  42. [50]

    Kamath, T. et al. Single-cell genomic profiling of human dopamine neurons identifies a pop- ulation that selectively degenerates in Parkinson’s disease. Nature neuroscience 25, 588–595 (2022)

    2022

  43. [51]

    Lau, S.-F., Cao, H., Fu, A. K. & Ip, N. Y . Single-nucleus transcriptome analysis reveals dysregulation of angiogenic endothelial cells and neuroprotective glia in Alzheimer’s disease. Proceedings of the National Academy of Sciences 117, 25800–25809 (2020)

    2020

  44. [52]

    Otero-Garcia, M. et al. Molecular signatures underlying neurofibrillary tangle susceptibility in Alzheimer’s disease. Neuron 110, 2929–2948 (2022)

    2022

  45. [53]

    Morabito, S. et al. Single-nucleus chromatin accessibility and transcriptomic characterization of Alzheimer’s disease. Nature genetics 53, 1143–1155 (2021)

    2021

  46. [54]

    Sadick, J. S. et al. Astrocytes and oligodendrocytes undergo subtype-specific transcriptional changes in Alzheimer’s disease. Neuron 110, 1788–1805 (2022)

    2022

  47. [55]

    Smith, A. M. et al. Diverse human astrocyte and microglial transcriptional responses to Alzheimer’s pathology. Acta Neuropathologica 143, 75–91 (2022). 31

    2022

  48. [56]

    Wang, Q. et al. Single-cell transcriptomic atlas of the human substantia nigra in Parkinson’s disease. Biorxiv 2022–03 (2022)

    2022

  49. [57]

    Yang, A. C. et al. A human brain vascular atlas reveals diverse mediators of Alzheimer’s risk. Nature 603, 885–892 (2022)

    2022

  50. [58]

    Zhang, L. et al. Single-cell transcriptomic atlas of Alzheimer’s disease middle temporal gyrus reveals region, cell type and sex specificity of gene expression with novel genetic risk for MERTK in female. medRxiv 2023–02 (2023)

    2023

  51. [59]

    Zhu, B. et al. Single-cell transcriptomic and proteomic analysis of Parkinson’s disease brains. Science Translational Medicine 16, eabo1997 (2024)

    2024

  52. [60]

    A., Angerer, P

    Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: Large-scale single-cell gene expression data analysis. Genome biology 19, 1–5 (2018)

    2018

  53. [61]

    & Kobak, D

    Lause, J., Berens, P. & Kobak, D. Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data. Genome biology 22, 1–20 (2021)

    2021

  54. [62]

    Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nature methods 16, 1289–1296 (2019). Correspondence Correspondence to Jong Chul Ye or Inkyung Jung. Acknowledgements This research was supported by National Research Foundation of K...

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.