Pith. sign in

REVIEW 4 major objections 5 minor 25 references

CausalGeD: Blending Causality and Diffusion for Spatial Gene Expression Generation

T0 review · 4 major / 5 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 Causality-aware diffusion lifts spatial gene prediction up to 32 percent over baselines

desk verdict A useful empirical imputation framework whose causal interpretation is unsupported: the attention mask is a random triangular partition, so the 'regulatory mechanism' claims are overreach. read the letter →

arxiv 2502.07751 v1 pith:NJ2DIJ3H submitted 2025-02-11 cs.CV q-bio.GN

Rabeya Tus Sadia , Md Atik Ahamed , Qiang Cheng This is my paper · ORCID
classification cs.CVq-bio.GN
keywords CausalrelationshipdiffusionmodelspatialtranscriptomicsdatascRNA-seqtransformerautoregressiongeneexpressiongenerationattentionmask
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

The paper argues that spatial gene expression imputation from single-cell RNA sequencing improves when the model explicitly respects causal, directional relationships among genes rather than only correlations. It presents CausalGeD, a framework that blends a diffusion denoising process with an autoregressive sequence model inside a Causality-Aware Transformer (CAT), and reports that it outperforms nine baselines on ten tissue datasets by 5–32% in Pearson correlation, structural similarity, RMSE, and JS divergence. If those results are right, researchers would gain a more accurate tool for reconstructing missing spatial transcripts and for probing tissue organization and disease states. The paper's own stated limitation is that the spatial genes must be a subset of the single-cell genes.

What carries the argument

The load-bearing mechanism is the causality-aware attention mask built by Algorithm 2, which splits gene tokens into autoregressive steps of random sizes and masks attention so earlier steps are visible to later steps but not vice versa, while condition tokens stay unmasked. This mask is applied inside the Causality-Aware Transformer (CAT) module, the component that merges diffusion and autoregression; the diffusion process supplies a denoising objective and the transformer supplies the causal ordering. The central identity that carries the argument is the joint transition in Eq. 3, which couples each autoregressive step's noisy tokens to clean tokens from previous steps.

What would settle it

Train CausalGeD on the same ten datasets with the causal mask replaced by a random or reversed attention mask and compare PCC/SSIM; if performance does not drop materially, the causal ordering is not what drives the gains. Equivalently, if the model's learned attention weights show no agreement with known regulatory edges on a dataset with validated interactions, the biological interpretation is unsupported.

Watch

Extended reading notes

Core claim

The paper's core assertion is that the standard diffusion-model pipeline for spatial transcriptomics ignores directional gene-gene dependencies, and that adding an autoregressive token order protected by a causal attention mask recovers those dependencies. CausalGeD encodes spatial and single-cell data jointly, adds noise to the spatial latent representation, and denoises it conditioned on scRNA-seq while a causality-aware mask forces earlier gene tokens to influence later ones. Across ten tissue datasets, the paper reports that this design beats state-of-the-art baselines by 5–32% on Pearson correlation, SSIM, RMSE, and JS divergence, including gains attributed to better structural fidelity in tumor and embryo data.

Load-bearing premise

The model's causal mask derives its ordering from a randomized autoregressive split (Algorithm 1) rather than from Granger causality or known regulatory networks, so the claim that CausalGeD captures biological causality rests on the assumption that a random AR order can stand in for true gene-gene dependencies.

Editorial extensions

If this is right

  • Spatial imputation pipelines that adopt causal ordering could see correlation and structural-similarity gains of roughly 5–32% over current diffusion-based tools, with the largest gains on datasets with high dropout.
  • The CAT-style fusion of diffusion and autoregression can be transferred to other paired single-cell datasets where a full-coverage modality guides imputation of a sparse spatial modality.
  • Because the causal mask needs no predefined gene network, the method can be applied to organisms or tissues where regulatory annotations are sparse.
  • The sampling-strategy ablation shows that diffusion timesteps can be cut to as little as one-twentieth of the default while keeping nearly the same accuracy, lowering compute cost substantially.
  • If the biological interpretation is correct, predicted spatial patterns may help locate regulatory niches in tumor microenvironments and in developing tissues.

Reading between the lines

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

  • The paper's causal mask is generated from a random autoregressive split (Algorithm 1) rather than from the Granger causality analysis it cites; comparing those two orderings would isolate whether the reported gains come from genuine causal structure or merely from the autoregressive inductive bias.
  • A natural extension is to make the attention mask learnable end-to-end, which would turn CausalGeD into a candidate causal-discovery tool for spatial transcriptomics, with learned attention edges checked against known regulatory interactions.
  • The architecture suggests a broader recipe: any paired dataset with full-coverage covariates and sparse targets could benefit from a causality-aware diffusion-autoregressive hybrid, not just gene expression data.
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

4 major / 5 minor

Summary. The paper proposes CausalGeD, a diffusion-plus-autoregressive transformer for imputing missing spatial gene expression by integrating scRNA-seq and spatial transcriptomics data. The method uses a "causality-aware transformer" (CAT) whose attention mask is built from randomly chosen autoregressive split sizes, and the authors claim that this architecture captures gene regulatory mechanisms without predefined relationships. Experiments on ten paired datasets report very large gains over nine baselines (e.g., PCC up to 0.92), together with ablations on decoder training, encoder variation, AR step decay, transformer depth, and diffusion time steps. The paper positions the Granger causality analysis of gene pairs as the biological motivation and claims that CausalGeD provides new insights into regulatory mechanisms.

Significance. The problem addressed is important: spatial transcriptomics imputation from scRNA-seq data is a central task, and the paper targets a large evaluation across ten datasets with multiple baselines. If the causal-attention claims were sound, the work could be a useful contribution to the field. However, the causal mechanism at the core of the paper is not actually implemented: the attention mask is generated from random AR split sizes and is gene-agnostic, so the stated novelty and biological interpretability are unsupported. The empirical headline claim is also contradicted by the paper's own Table I on multiple datasets. The manuscript also does not provide code, which limits reproducibility, although the extensive ablation study and the explicit reporting of per-dataset results are strengths.

major comments (4)
  1. [Appendix VII-A, Algorithms 1 and 2; Section III-C(c)] The causal attention mask is built from AR split sizes randomly generated in Algorithm 1 (with an exponential decay factor alpha) and a purely triangular position-based mask in Algorithm 2. The mask depends only on the positions of tokens, not on the identities of genes or on the Granger causality results presented in Figure 1 and Section I. Therefore the central claim that CausalGeD "captures regulatory mechanisms without predefined relationships" (Abstract, Section III) is unsupported. To support the claim, the authors would need to show that the mask is derived from learned or data-driven causal structures, or at least demonstrate token-order invariance (e.g., shuffling the gene order while keeping the mask and confirming that predictions do not change). As written, the model reduces to an autoregressive diffusion transformer with a random token partition.
  2. [Abstract and Table I] The claim that "Across 10 tissue datasets, CausalGeD outperformed state-of-the-art baselines by 5-32% in key metrics" is contradicted by the data in Table I. On the MH dataset, CausalGeD's PCC is 0.612 while SpaDiT achieves 0.621, so CausalGeD is worse. On RMSE, CausalGeD is worse than SpaDiT on MG (1.155 vs 0.877), MH (1.275 vs 1.103), and MC (1.085 vs 1.004), and only marginally better on MVC (1.115 vs 1.116). Thus the claim of consistent state-of-the-art performance across all four metrics and ten datasets is not supported by the reported numbers.
  3. [Section III-C, Eq. (3)] Equation (3) defines the AR-diffusion joint distribution in a way that is notationally and conceptually unclear. The forward process in Eq. (1) is a standard Markov chain that depends only on the previous diffusion step, but Eq. (3) conditions the forward transitions for AR step kappa_s on clean tokens from previous AR steps (x_hat^{0,kappa_1:s-1}). This mixing of conditioning on clean tokens during the forward process is not derived from the forward diffusion equation and does not define a valid generative process as stated. The authors should provide a precise generative model (e.g., a masked diffusion formulation) and clarify how training samples are constructed.
  4. [Section IV-A and Figure 1] The Granger causality analysis is central to the motivation but is described only as computed on "randomly selected MC data" with no methodological details: no lag order, no preprocessing, no test procedure, and no specification of how the statistically significant F-statistics are obtained. As a result, the reader cannot verify the claim that existing methods overlook "important causal relationships" or evaluate whether the chosen gene pairs are representative. A concrete, reproducible description of the Granger causality analysis is needed.
minor comments (5)
  1. [Abstract] There is a punctuation error: "with structural similarity often below 60%, We attribute" should be "below 60%. We attribute".
  2. [Section IV-A, paragraph 1] The sentence "These datasets come from a variety of biological tissues, mostly tissue samples from human and mouse breast cancer" is inaccurate, as the dataset table includes mouse hippocampus, hypothalamus, liver, gastrulation, visual cortex, hindlimb muscle, embryo, and cerebellum, not mostly breast cancer.
  3. [Table I caption] The dataset abbreviations in Table I (MG, MH, MHPR, MVC, MHM, HBC, ME, MPMC, MC, ML) are not defined in the caption or in Section IV-A; the reader must wait until Appendix Table III to decode them. The caption should define the abbreviations.
  4. [Figure 5] The hierarchical clustering figure legend says "the first row of the figure shows the true gene labels after clustering," but the figure itself appears to show heatmaps for different methods. The description does not specify how the similarity between predicted and true labels is quantified, so the claimed "closely align" is not supported by a numeric measure.
  5. [References [14]] Reference [14] appears as "C. Deng, D. Zh, K. Li, S. Guan, and H. Fan" — the author name appears truncated ("D. Zh" should likely be "D. Zhang" or similar) and the title is given without a journal or venue. This reference should be checked and corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical gains are measured against external baselines; the unsupported causal interpretation is an evidence gap, not a circular derivation.

full rationale

No circular derivation is present. CausalGeD's central quantitative claim, outperforming nine baselines on ten tissue datasets by 5-32% in PCC, SSIM, RMSE, and JS, is an external empirical comparison: performance is reported on a held-out test set, hyperparameters are tuned on validation metrics, and the baselines are independently published methods. The causal vocabulary is motivated by a Granger causality analysis in Figure 1, but the actual causal attention mask (Algorithms 1-2 in Appendix VII-A) is built from randomly generated AR split sizes and position-based triangular masks, not from the Granger results or gene identities. This means the paper's claim that the model 'captures regulatory mechanisms' is not justified by the described architecture, but that is a support/correctness weakness, not a circularity: no fitted parameter is renamed as a prediction, no equation reduces to its own input, and no load-bearing self-citation or imported uniqueness theorem is used. The empirical comparison stands independently of the causal narrative, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard diffusion math and on an unstated, untested premise that a random AR ordering captures gene causality. Most parameters are engineering choices rather than derived quantities.

free parameters (4)
  • AR step decay alpha = 0.8
    Chosen as 'optimal decay ratio' via ablation in Section V.C.
  • Diffusion timesteps T = 2000
    Default setting; robustness tested but not derived.
  • Number of transformer blocks = 3 or more
    Selected based on ablation in Section V.D; no principled derivation.
  • Diffusion variance schedule beta_t = not reported
    Needed for forward diffusion (Equation 1) but values are not given in the paper.
assumptions (5)
  • domain assumption Granger causality on expression matrices indicates gene regulatory relationships
    Section I and Figure 1 use F-statistics to motivate causality, but no evidence links Granger causality to a true regulatory mechanism.
  • domain assumption The causal attention mask of [14] can be transferred from images to continuous gene expression values
    Section III.B.c claims fundamental innovations but the core mask is from image transformers.
  • ad hoc to paper A random autoregressive ordering of genes is sufficient to learn regulatory dependencies
    Algorithm 1 determines split sizes and token order randomly; the paper offers no justification for why this captures causal structure.
  • standard math Standard DDPM forward diffusion process (Equation 1) applies to latent representations
    Assumes Gaussian noising is valid for the latent space.
  • domain assumption Top 25% highly variable genes are representative for evaluation
    Section IV.A.c selects HVGs; results may not generalize to low-variability genes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CausalGeD: Blending Causality and Diffusion for Spatial Gene Expression Generation." pith.science (2026). https://pith.science/paper/NJ2DIJ3H

@misc{pith2026250207751,
  author       = {Pith},
  title        = {Pith review of: CausalGeD: Blending Causality and Diffusion for Spatial Gene Expression Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJ2DIJ3H}},
  note         = {Machine review of arXiv:2502.07751}
}
read the original abstract

The integration of single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) data is crucial for understanding gene expression in spatial context. Existing methods for such integration have limited performance, with structural similarity often below 60\%, We attribute this limitation to the failure to consider causal relationships between genes. We present CausalGeD, which combines diffusion and autoregressive processes to leverage these relationships. By generalizing the Causal Attention Transformer from image generation to gene expression data, our model captures regulatory mechanisms without predefined relationships. Across 10 tissue datasets, CausalGeD outperformed state-of-the-art baselines by 5- 32\% in key metrics, including Pearson's correlation and structural similarity, advancing both technical and biological insights.

Figures

Figures reproduced from arXiv: 2502.07751 by the authors.

Figure 1
Figure 1. Top 5 gene pairs with strongest Granger causality [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture overview of CausalGeD. The framework comprises both training and inference processes, implemented [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Diffusion blended with autoregression. performance. The forward diffusion process spans T timesteps (default T = 2000), where noise is introduced at each step t according to variance schedule βt. Instead of a fixed timestep for all autoregressive (AR) steps, we adopt a flexible sampling approach: 1. Base Strategy: For each AR step s, multiple diffusion timesteps are sampled to capture varying noise levels, enabling … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Low-dimensional UMAP visualizations of predicted gene expression vs. real gene expression for the proposed method [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Hierarchical clustering method to visualize the similarity between the predicted genes and the true gene labels in the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation study of decoder training module [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: The effect of AR step decay D. Transformer Blocks In our CausalGeD model we investigated the impact of transformer block depth on performance-efficiency trade￾offs, dataset-specific variability, and generalization robustness across diverse datasets [PITH_FULL_IMAGE:fi…
Figure 7
Figure 7. Figure 7: Comparison of PCC values across different datasets [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Effect of varying transformer blocks on PCC values [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Illustration of the different parts in the process of causal attention mask formation. White indicates allowed regions, [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    Spage: spatial gene enhancement using scrna-seq,

    T. Abdelaal, S. Mourragui et al., “Spage: spatial gene enhancement using scrna-seq,” Nucleic Acids Research , vol. 48, no. 18, pp. e107–e107, 2020

    work page 2020

  2. [2]

    Integrating spatial and single-cell transcrip- tomics data using deep generative models with spatialscope,

    X. Wan, J. Xiao et al. , “Integrating spatial and single-cell transcrip- tomics data using deep generative models with spatialscope,” Nature Communications, vol. 14, no. 1, p. 7848, 2023

    work page 2023

  3. [3]

    Novosparc: flexible spatial reconstruction of single-cell gene expression with optimal transport,

    N. Moriel, E. Senel et al., “Novosparc: flexible spatial reconstruction of single-cell gene expression with optimal transport,” Nature Protocols, vol. 16, no. 9, pp. 4177–4200, 2021

    work page 2021

  4. [4]

    stdiff: a diffusion model for imputing spatial transcrip- tomics through single-cell transcriptomics,

    K. Li, J. Li et al., “stdiff: a diffusion model for imputing spatial transcrip- tomics through single-cell transcriptomics,” Briefings in Bioinformatics , vol. 25, no. 3, p. bbae171, 2024

    work page 2024

  5. [5]

    Spadit: diffusion transformer for spatial gene expression prediction using scrna-seq,

    X. Li, F. Zhu, and W. Min, “Spadit: diffusion transformer for spatial gene expression prediction using scrna-seq,” Briefings in Bioinformatics, vol. 25, no. 6, p. bbae571, 2024

    work page 2024

  6. [6]

    Prior knowledge driven granger causality analysis on gene regulatory network discovery,

    S. Yao, S. Yoo, and D. Yu, “Prior knowledge driven granger causality analysis on gene regulatory network discovery,” BMC bioinformatics , vol. 16, pp. 1–18, 2015

    work page 2015

  7. [7]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  8. [8]

    Diffusion-lm improves controllable text generation,

    X. Li, J. Thickstun, I. Gulrajani, P. S. Liang, and T. B. Hashimoto, “Diffusion-lm improves controllable text generation,” Advances in Neu- ral Information Processing Systems , vol. 35, pp. 4328–4343, 2022

    work page 2022

Show all 25 references
  1. [9]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  2. [10]

    Deep learning and alignment of spatially resolved single-cell transcriptomes with tangram,

    T. Biancalani, G. Scalia et al., “Deep learning and alignment of spatially resolved single-cell transcriptomes with tangram,” Nature Methods , vol. 18, no. 11, pp. 1352–1362, 2021

    2021

  3. [11]

    Deep generative modeling for single-cell transcriptomics,

    R. Lopez, J. Regier et al. , “Deep generative modeling for single-cell transcriptomics,” Nature Methods, vol. 15, no. 12, pp. 1053–1058, 2018

    2018

  4. [12]

    Inferring spatial and signaling relationships between cells from single cell transcriptomic data,

    Z. Cang and Q. Nie, “Inferring spatial and signaling relationships between cells from single cell transcriptomic data,” Nature Communi- cations, vol. 11, no. 1, p. 2084, 2020

    2020

  5. [13]

    stplus: a reference-based method for the accurate enhancement of spatial transcriptomics,

    C. Shengquan, Z. Boheng et al., “stplus: a reference-based method for the accurate enhancement of spatial transcriptomics,” Bioinformatics, vol. 37, no. Supplement 1, pp. i299–i307, 2021

    2021

  6. [14]

    Causal diffusion transformers for generative modeling,

    C. Deng, D. Zh, K. Li, S. Guan, and H. Fan, “Causal diffusion transformers for generative modeling,”arXiv preprint arXiv:2412.12095, 2024

    2024 arXiv

  7. [15]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems , vol. 32, 2019

    2019

  8. [16]

    A spatially resolved brain region-and cell type-specific isoform atlas of the postnatal mouse brain,

    A. Joglekar, A. Prjibelski et al., “A spatially resolved brain region-and cell type-specific isoform atlas of the postnatal mouse brain,” Nature Communications, vol. 12, no. 1, p. 463, 2021

    2021

  9. [17]

    Molecular, spatial, and func- tional single-cell profiling of the hypothalamic preoptic region,

    J. R. Moffitt, D. Bambah-Mukku et al. , “Molecular, spatial, and func- tional single-cell profiling of the hypothalamic preoptic region,” Science, vol. 362, no. 6416, p. eaau5324, 2018

    2018

  10. [18]

    Single-cell transcriptomics of 20 mouse organs creates a tabula muris: The tabula muris consortium,

    N. Schaum, J. Karkanias et al., “Single-cell transcriptomics of 20 mouse organs creates a tabula muris: The tabula muris consortium,”Nature, vol. 562, no. 7727, p. 367, 2018

    2018

  11. [19]

    Three-dimensional intact-tissue sequencing of single-cell transcriptional states,

    X. Wang, W. E. Allen et al., “Three-dimensional intact-tissue sequencing of single-cell transcriptional states,” Science, vol. 361, no. 6400, p. eaat5691, 2018

    2018

  12. [20]

    Large-scale integration of single- cell transcriptomic data captures transitional progenitor states in mouse skeletal muscle regeneration,

    D. W. McKellar, L. D. Walter et al., “Large-scale integration of single- cell transcriptomic data captures transitional progenitor states in mouse skeletal muscle regeneration,” Communications Biology , vol. 4, no. 1, p. 1280, 2021

    2021

  13. [21]

    A single-cell and spatially resolved atlas of human breast cancers,

    S. Z. Wu, G. Al-Eryani et al., “A single-cell and spatially resolved atlas of human breast cancers,” Nature Genetics , vol. 53, no. 9, pp. 1334– 1347, 2021

    2021

  14. [22]

    A coordinated progression of progenitor cell states initiates urinary tract development,

    O. Sanchez-Ferras, A. Pacis et al. , “A coordinated progression of progenitor cell states initiates urinary tract development,” Nature Com- munications, vol. 12, no. 1, p. 2627, 2021

    2021

  15. [23]

    Single-cell rna landscape of intratumoral heterogeneity and immunosuppressive microenvironment in advanced osteosarcoma,

    Y . Zhou, D. Yang et al. , “Single-cell rna landscape of intratumoral heterogeneity and immunosuppressive microenvironment in advanced osteosarcoma,” Nature Communications, vol. 11, no. 1, p. 6322, 2020

    2020

  16. [24]

    Sequencing of individual bar- coded cdnas using pacific biosciences and oxford nanopore technologies reveals platform-specific error patterns,

    A. Mikheenko, A. D. Prjibelski et al. , “Sequencing of individual bar- coded cdnas using pacific biosciences and oxford nanopore technologies reveals platform-specific error patterns,” Genome Research , vol. 32, no. 4, pp. 726–737, 2022. VII. A PPENDIX A. Causal Mask formation...

    2022

  17. [25]

    + [S] 10: sz ← [cs[i + 1]− cs[i] for i ∈ [0, N− 1]] 11: return sz, cs Algorithm 2 Generate Causal Attention Mask Input: s (sample length), c (conditional length), sz (split sizes), cs (cumulative sum of split sizes) 1: v ← s − sz[−1] % Visible length 2: ctx ← c + v % Context l...

    1939

Pith tools

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