Latent Gene Diffusion for Spatial Transcriptomics Completion

Daniela Ruiz; Daniela Vega; Leonardo Manrique; Pablo Arbel\'aez; Paula C\'ardenas

REVIEW 4 major objections 5 minor 31 references

Latent Gene Diffusion for Spatial Transcriptomics Completion

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that LGDiST, a reference-free latent gene diffusion model, completes missing spatial gene expression with an average MSE 18% lower than the prior reference-free method across 26 datasets, and that this completion improves

desk verdict New reference-free ST completion method with broad benchmark results, but the 18% gain rests on random-mask simulation and the partial-dropout protocol is under-specified. read the letter →

arxiv 2509.01864 v1 pith:JX2VUCMT submitted 2025-09-02 cs.CV

Paula Cárdenas , Leonardo Manrique , Daniela Vega , Daniela Ruiz , Pablo Arbeláez This is my paper

classification cs.CV

keywords spatialtranscriptomicsdatadropoutgeneexpressioncompletionlatentdiffusiontransformerMoran'sIreference-freeimputationprediction

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 is trying to establish that the dropout problem in sequencing-based spatial transcriptomics can be solved without external single-cell references, using a generative model that completes missing gene values from the tissue's own spatial structure. It claims an 18% average MSE reduction over the prior reference-free completion method across 26 datasets, and shows that six established gene-expression-from-histology predictors train better on LGDiST-completed data, with up to 10% lower MSE. The reason to care: dropout is a known bottleneck for spatial transcriptomics and for computer-vision models that predict expression from tissue images, and a reference-free completion method would make the technology usable on new tissues without alignment to external data.

What carries the argument

The load-bearing mechanism is a two-stage latent diffusion. An autoencoder with a transformer encoder maps each spot's neighborhood gene-expression matrix—highly spatially associated genes plus 1,024 low-Moran's-I 'context genes'—into a 128-dimensional latent space; a Diffusion Transformer then iteratively denoises a noised row for the central spot, conditioned on the latent vectors of its six nearest neighbors. The context genes, previously discarded as uninformative, supply covariance structure that enriches the latent space, while neighbor conditioning supplies spatial context. Ablations quantify the contribution of each piece: dropping context genes raises MSE by 23%, dropping the autoen

What would settle it

Compare LGDiST against the prior reference-free method on a dataset where the missing values are realistic rather than random—for example, mask contiguous spatial regions or genes with low detection rates inferred from statistical dropout models, then validate recovered values against matched single-cell RNA-seq profiles. If LGDiST's MSE advantage shrinks or reverses under this mask, the random-mask evaluation was the load-bearing artifact.

Watch

Extended reading notes

Core claim

The paper introduces LGDiST, a generative completion model that repairs missing gene expression values in sequencing-based spatial transcriptomics without any external single-cell reference. It claims that completing data in a learned latent gene space, rather than raw expression space, and conditioning the diffusion process on encoded expression of neighboring spots yields an average MSE 18% lower than the previous reference-free state of the art across 26 datasets. The paper further claims that when six existing histology-to-expression prediction models are trained on LGDiST-completed data, their evaluation MSE improves by up to 10%, with stronger spatial-pattern preservation (higher PCC)

Load-bearing premise

The central evaluation assumes that randomly masking already-measured expression values faithfully reproduces real ST dropout; if genuine dropout is structured by space or by gene, the measured 18% improvement may not carry over.

Editorial extensions

If this is right

LGDiST can be applied to sequencing-based spatial transcriptomics samples without a matched single-cell reference, removing batch-effect and alignment-quality risks.
Training data completed by LGDiST improves six existing image-to-expression models, so completion can serve as a preprocessing step that raises the ceiling of current predictors.
The use of low-spatial-autocorrelation context genes demonstrates that genes previously filtered out as uninformative carry usable signal for imputation.
A six-neighbor configuration gives nearly the same accuracy as an 18-neighbor configuration at 171% lower FLOPs, so the deployment cost stays modest.

Reading between the lines

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

If random masking does not reproduce real dropout structure, the 18% advantage may not transfer to genuine missing data; a direct test is to mask contiguous spatial domains or genes with known low detection rates and validate against matched single-cell profiles.
The same latent-diffusion recipe could be lifted to other spatially resolved omics, such as protein or methylation panels, where reference-free dropout completion is needed and only a spatial neighborhood matrix is available.
The downstream gains suggest the image-to-expression task is limited as much by noisy training targets as by model architecture; jointly training completion and prediction might yield larger improvements than either alone.
Because the autoencoder latent space is trained per dataset, cross-tissue generalization is not established; a single latent space shared across datasets would be a testable extension.

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.

Desk Editor's Note

read the letter

LGDiST is a legitimate step forward for reference-free spatial transcriptomics completion. The core idea—latent diffusion over a neighborhood's gene expression, using low-Moran's-I 'context genes' as part of the latent representation—is new and sensible, and the results are broad: 18% lower average MSE than SpaCKLE over the 26 SpaRED datasets, plus a downstream experiment showing that six histology-to-expression prediction models improve when trained on LGDiST-completed data. The ablations are coherent: removing context genes hurts, removing the autoencoder hurts more, and 1-hop neighbors are sufficient. Code is promised public.

The soft spots are mostly about how the evidence is presented. The main comparison has no error bars or significance tests, so it's hard to know whether the 18% is stable across random seeds or dataset subsets. The stDiff* baseline is under-specified—'restructured' to be reference-free, but the exact changes aren't given, leaving the reader to take the comparison on faith. The ablations use six datasets that are not listed, so the reader can't assess their coverage.

The bigger issue is the dropout simulation. The model is trained to complete a fully empty central spot, with the central row masked in Eq. 4. But the real use case is partial dropout, and the paper never explains how observed central-spot values enter the condition C or the noise schedule at inference. It only says to extract the missing values from the generated profile. That's a gap, not a deal-breaker—the method could still work—but it's underspecified for reproduction. More importantly, the evaluation simulates dropout by randomly hiding successfully measured values. Real ST dropout is a detection process: low-abundance transcripts and genes with poor detection efficiency are more likely to be missing, and spatial structure matters. If the synthetic mask over-represents high-expression genes, the reported 18% advantage may not transfer to genuine missing data. This is a threat to external validity, not to the internal comparison under the same mask.

Who is this for? Anyone working on ST imputation or gene-expression prediction from histology. The paper deserves a serious referee. I'd ask the authors to report variance across seeds, specify the stDiff* modification, list the ablation datasets, and—ideally—validate on a real dropout scenario (e.g., hold out genes with high native dropout rates or simulate dropout with a detection model). With those additions, this would be a solid reference for the subfield. Send it to peer review.

Referee Report

4 major / 5 minor

Summary. The paper introduces LGDiST, a reference-free latent diffusion model for gene expression dropout completion in sequencing-based spatial transcriptomics. The method has two stages: a transformer autoencoder builds a latent space over neighborhoods comprising highly spatially associated genes (HSAGs) plus additional low-Moran's-I 'context genes' (CGs); a Diffusion Transformer is then trained to denoise a fully masked central spot conditioned on encoded neighbor information. Experiments on the 26 SpaRED datasets report an average MSE 18% lower than SpaCKLE, the prior reference-free state of the art, and show that training six existing histology-to-expression prediction models on LGDiST-completed data improves their MSE by up to 10%. Ablations on six datasets indicate that removing CGs, removing the autoencoder, or removing neighbor conditioning degrades completion performance.

Significance. If the central claim holds, LGDiST is a meaningful advance: it is the first reference-free generative model for ST completion, it demonstrates that low-spatial-autocorrelation genes carry useful context, and its downstream benefit for gene-expression prediction from histology images is practically relevant. The paper's positive features include a broad 26-dataset benchmark, publicly available code, honest reporting of suboptimal qualitative cases, and a clear two-stage architecture. The main caveat is that the headline improvement is established only on synthetically masked data, and the inference protocol for the actual partial-dropout setting is not fully specified; these issues are fixable but need attention.

major comments (4)

[§3.3, Eq. (4)] The training objective in Eq. (4) uses a binary mask M that zeroes the entire central-spot row, and conditioning C = M ⊙ XE contains only neighbor information. The paper then states that for real partial dropout, during inference one 'extract[s] only the values corresponding to the missing data,' but the protocol is unspecified: how are the observed central-spot genes incorporated into the condition C, and what is the noising schedule for the partially observed central row? Without this, the central real-world use case is not reproducible. Please provide the exact inference algorithm for arbitrary masks, including whether observed values are kept clean or re-noised, and how the decoder output is combined with observed values.
[§3.3, §4.2, Fig. 4] The evaluation simulates dropout by uniformly hiding measured values, but real ST dropout is a detection process that depends on expression abundance, gene detection efficiency, and local sequencing depth; it is neither uniform across genes nor spatially independent. The 18% average MSE improvement is measured solely on random masks, so the external validity to genuine dropout is asserted, not demonstrated. Please either add experiments with more realistic dropout models (e.g., abundance-dependent dropout, spatially structured masks) or explicitly stratify the current results by gene expression level and spatial pattern. At minimum, this limitation should be discussed in the text.
[§4.2, stDiff* baseline] The comparison against stDiff* is a key part of the benchmark, but the modification that makes stDiff reference-free is described only as 'restructuring its diffusion conditioning.' No details are given about what changed, what hyperparameters were used, or how the scRNA-seq reference was removed. Since LGDiST is compared against this baseline on all 26 datasets, the reader cannot assess whether the comparison is fair or reproduce the baseline. Please specify the exact modifications, or provide the baseline code, and report any tuning performed.
[§4.4, Tables 1 and 2] The ablation experiments are performed on only six datasets, but the datasets are not named, and no statistical tests are provided. The reported standard deviations are large relative to the differences (e.g., in Table 1, 0.573 ± 0.136 vs 0.710 ± 0.214; Table 2, 0.573 ± 0.136 vs 0.588 ± 0.155). Without paired tests or confidence intervals across the ablated datasets, the claims that context genes and 1-hop neighborhoods are essential are not strongly supported. Please list the datasets used, provide per-dataset results, and report a paired significance test or effect size.

minor comments (5)

[§3.3, Eq. (4)] The mask M is defined in {0,1}^((n+1)×d), but earlier in the same section M is used as a mask over gene expression values in the input space. Please clarify the notation and distinguish the latent-space mask from the gene-space mask.
[Figure 3 caption] The caption contains the typo 'HSGAs'; it should be 'HSAGs'.
[Figure 4] The y-axis label reads 'MSE for all imputation methods' while the legend uses 'LGDist' (should be LGDiST). Also, the figure shows distributions but does not explain whether these are box plots or violin plots; please describe the plot elements in the caption.
[Figure 7] The line plot reports completion MSE for a single dataset (10XGMBSP). It would be helpful to state whether the curves are averaged over multiple random masks and to include error bars or confidence intervals.
[§4.3, Figure 8] The right Y-axis of Figure 8 reports 'improvement percentage' with no numerical table. Given that the PCC improvement for SEPAL is described as 188%, please include a table with the exact per-method percentages and the number of datasets where improvement occurred, so the reader can verify the claims.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are empirical comparisons on held-out masks; self-citations are present but non-load-bearing.

full rationale

The paper's central claims are empirical, not derivational. LGDiST is trained to reconstruct fully masked central spots from neighbor information (Eqs. 4–5), and evaluation is performed by randomly masking successfully measured values and comparing MSE/PCC against SpaCKLE and a modified stDiff on the held-out values. The headline 18% average MSE reduction is a benchmark result, not a quantity recovered from the model's own inputs. The ablation experiments compare retrained variants, so the context-gene and neighbor-conditioning claims are also empirical, not definitional. The paper does cite the authors' prior SpaRED/SpaCKLE/SEPAL work for the benchmark, the median pre-completion, and the evaluation protocol; these are self-citations, but they are not load-bearing in a circular sense because both LGDiST and the SpaCKLE baseline use the same pre-completion and are evaluated on the same held-out masks. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The underspecified partial-mask inference protocol and the realism of random masking as a dropout model are substantive scientific concerns, but they concern external validity and reproducibility, not circularity of the derivation.

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

The central claim is an empirical performance gain; it rests on the choice of gene subsets, pre-completion, mask simulation, and neighborhood size. These are design choices, not derived from biology, so they are listed as free parameters or domain assumptions.

free parameters (4)

alpha = not reported
Weighting factor in the reconstruction loss between HSAGs and context genes; no value or tuning procedure is given, and it drives the latent space design.
latent_dim = 128
Dimension of the latent space; chosen by hand, not justified.
n_neighbors = 6
Number of neighboring spots used for conditioning; ablation shows 6 beats 0 but 18 does not improve, so 6 is a hand-chosen default.
n_context_genes = 1024
Number of low-Moran's-I context genes; chosen by adjusting SpaRED's filtering threshold, no sensitivity analysis.

assumptions (4)

domain assumption Moran's I based gene selection is sufficient to define biologically meaningful gene sets
HSAGs are the top 32/128 Moran's I genes; CGs are the next ones up to 1024. The paper assumes these capture the relevant gene programs.
domain assumption Median pre-completion provides a valid initialization for training
The model is trained on data pre-filled with median expression, following [18]; if this initialization is biased, the model inherits the bias.
domain assumption Random masking of measured values simulates realistic dropout
Evaluation masks random measured entries; true dropout may be non-uniform and technology-specific.
domain assumption Neighborhood of n nearest spots contains sufficient spatial context
Default 6 neighbors (1-hop) chosen; no theoretical justification beyond ablation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Latent Gene Diffusion for Spatial Transcriptomics Completion." pith.science (2026). https://pith.science/paper/JX2VUCMT

@misc{pith2026250901864,
  author       = {Pith},
  title        = {Pith review of: Latent Gene Diffusion for Spatial Transcriptomics Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JX2VUCMT}},
  note         = {Machine review of arXiv:2509.01864}
}

read the original abstract

Computer Vision has proven to be a powerful tool for analyzing Spatial Transcriptomics (ST) data. However, current models that predict spatially resolved gene expression from histopathology images suffer from significant limitations due to data dropout. Most existing approaches rely on single-cell RNA sequencing references, making them dependent on alignment quality and external datasets while also risking batch effects and inherited dropout. In this paper, we address these limitations by introducing LGDiST, the first reference-free latent gene diffusion model for ST data dropout. We show that LGDiST outperforms the previous state-of-the-art in gene expression completion, with an average Mean Squared Error that is 18% lower across 26 datasets. Furthermore, we demonstrate that completing ST data with LGDiST improves gene expression prediction performance on six state-of-the-art methods up to 10% in MSE. A key innovation of LGDiST is using context genes previously considered uninformative to build a rich and biologically meaningful genetic latent space. Our experiments show that removing key components of LGDiST, such as the context genes, the ST latent space, and the neighbor conditioning, leads to considerable drops in performance. These findings underscore that the full architecture of LGDiST achieves substantially better performance than any of its isolated components.

Figures

Figures reproduced from arXiv: 2509.01864 by the authors.

**Figure 1.** Spatial Transcriptomics Completion Pipeline. LGDiST processes ST neighborhoods to infer missing values, generating a completed expression map with high-quality data. the models highly sensitive to the quality of external data but also increases the challenge of addressing dropout in sequencing-based ST, given its lower resolution. To overcome the limitations mentioned above, this paper introduces Latent Gene Diffus… view at source ↗

**Figure 2.** AutoEncoder training framework. We add a 2D positional encoding to the gene expression neighborhood X before passing it through an encoder Eθ that maps it to a Latent Space Representation XE. The decoder Dϕ then reconstructs the input from the latent space, aiming to minimize reconstruction loss. the completion process significantly increases the computational resources and experimental data requirements, since hig… view at source ↗

**Figure 3.** (a) LGDiST Training: The encoder Eθ maps the gene expression neighborhood X to XE. We feed this representation to the DiT with the condition C and t. Through an iterative noising-denoising process, the DiT learns to predict the noise in X˜E,t. After learning how to remove the noise, the DiT outputs XE, t ˜ = 0 and Dϕ decodes the result to obtain the final gene expression matrix. (b) LGDiST Inference: After obtaining… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: MSE of the three completion methods on each dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Optimal qualitative results for data completion with LGDiST vs with SpaCKLE [ [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Suboptimal qualitative results for data completion with LGDiST vs with SpaCKLE [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Line plot displaying completion MSE for LGDiST and [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: (a) MSE and (b) PCC of gene-expression prediction models when trained on LGDiST-completed data compared to training on [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

[1]

Spage: spatial gene enhancement using scrna-seq

Tamim Abdelaal, Soufiane Mourragui, Ahmed Mahfouz, and Marcel JT Reinders. Spage: spatial gene enhancement using scrna-seq. Nucleic acids research, 48(18):e107–e107, 2020. 1, 2

work page 2020
[2]

Autoencoders and their applications in machine learning: a survey

Kamal Berahmand, Fatemeh Daneshfar, Elaheh Sadat Salehi, Yuefeng Li, and Yue Xu. Autoencoders and their applications in machine learning: a survey. Artificial Intelli- gence Review, 57(2):28, 2024. 2

work page 2024
[3]

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

Tommaso Biancalani, Gabriele Scalia, Lorenzo Buffoni, Raghav Avasthi, Ziqing Lu, Aman Sanger, Neriman Tokcan, Charles R Vanderburg, ˚Asa Segerstolpe, Meng Zhang, et al. Deep learning and alignment of spatially resolved single-cell transcriptomes with tangram. Nature methods, 18(11):1352– 1362, 2021. 1

work page 2021
[4]

Spatially resolved transcrip- tomics: a comprehensive review of their technological ad- vances, applications, and challenges

Mengnan Cheng, Yujia Jiang, Jiangshan Xu, Alexios- Fotios A Mentis, Shuai Wang, Huiwen Zheng, Sunil Kumar Sahu, Longqi Liu, and Xun Xu. Spatially resolved transcrip- tomics: a comprehensive review of their technological ad- vances, applications, and challenges. Journal of Genetics and Genomics, 50(9):625–640, 2023. 1

work page 2023
[5]

A point cloud segmentation frame- work for image-based spatial transcriptomics

Thomas Defard, Hugo Laporte, Mallick Ayan, Juliette Soulier, Sandra Curras-Alonso, Christian Weber, Florian Massip, Jos ´e-Arturo Londo ˜no-Vallejo, Charles Fouillade, Florian Mueller, et al. A point cloud segmentation frame- work for image-based spatial transcriptomics. Communica- tions Biology, 7(1):823, 2024. 1

work page 2024
[6]

A practical guide to single-cell rna- sequencing for biomedical research and clinical applications

Ashraful Haque, Jessica Engel, Sarah A Teichmann, and Tapio L ¨onnberg. A practical guide to single-cell rna- sequencing for biomedical research and clinical applications. Genome medicine, 9:1–12, 2017. 1

work page 2017
[7]

Integrating spatial gene expression and breast tumour morphology via deep learning

Bryan He, Ludvig Bergenstr ˚ahle, Linnea Stenbeck, Abubakar Abid, Alma Andersson, ˚Ake Borg, Jonas Maaskola, Joakim Lundeberg, and James Zou. Integrating spatial gene expression and breast tumour morphology via deep learning. Nature biomedical engineering , 4(8): 827–834, 2020. 1, 6

work page 2020
[8]

Denoising dif- fusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2

2020

Show all 31 references

[9]

Advances in spatial transcriptomics and its applications in cancer re- search

Yang Jin, Yuanli Zuo, Gang Li, Wenrong Liu, Yitong Pan, Ting Fan, Xin Fu, Xiaojun Yao, and Yong Peng. Advances in spatial transcriptomics and its applications in cancer re- search. Molecular Cancer, 23(1):129, 2024. 1

2024
[10]

Adam: A method for stochastic optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

arXiv
[11]

Ma- chine learning methods for histopathological image analysis: Updates in 2024

Daisuke Komura, Mieko Ochi, and Shumpei Ishikawa. Ma- chine learning methods for histopathological image analysis: Updates in 2024. Computational and Structural Biotechnol- ogy Journal, 2024. 1

2024
[12]

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

Kongming Li, Jiahao Li, Yuhao Tao, and Fei Wang. stdiff: a diffusion model for imputing spatial transcriptomics through single-cell transcriptomics. Briefings in Bioinformatics , 25 (3):bbae171, 2024. 1, 2, 5, 6

2024
[13]

Emerging artificial intelligence applications in spatial transcriptomics analysis

Yijun Li, Stefan Stanojevic, and Lana X Garmire. Emerging artificial intelligence applications in spatial transcriptomics analysis. Computational and Structural Biotechnology Jour- nal, 20:2895–2908, 2022. 1

2022
[14]

scdiffusion: conditional generation of high-quality single- cell data using diffusion model

Erpai Luo, Minsheng Hao, Lei Wei, and Xuegong Zhang. scdiffusion: conditional generation of high-quality single- cell data using diffusion model. Bioinformatics, 40(9): btae518, 2024. 1

2024
[15]

Navigating the complexity of data imputation in spatial transcriptomics: Strategies, challenges, and future directions

Ricfrid van der Marel. Navigating the complexity of data imputation in spatial transcriptomics: Strategies, challenges, and future directions. 2024. 3

2024
[16]

Method of the year: spatially resolved tran- scriptomics

Vivien Marx. Method of the year: spatially resolved tran- scriptomics. Nature methods, 18(1):9–14, 2021. 1

2021
[17]

Sepal: Spatial gene expres- sion prediction from local graphs

Gabriel Mejia, Paula C ´ardenas, Daniela Ruiz, Angela Castillo, and Pablo Arbel ´aez. Sepal: Spatial gene expres- sion prediction from local graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2294–2303, 2023. 3, 5, 6

2023
[18]

Enhancing gene expression prediction from histology images with spa- tial transcriptomics completion

Gabriel Mejia, Daniela Ruiz, Paula C ´ardenas, Leonardo Manrique, Daniela Vega, and Pablo Arbel ´aez. Enhancing gene expression prediction from histology images with spa- tial transcriptomics completion. In International Conference on Medical Image Computing and Computer-Assis...

2024
[19]

Notes on continuous stochastic phenom- ena

Patrick AP Moran. Notes on continuous stochastic phenom- ena. Biometrika, 37(1/2):17–23, 1950. 3

1950
[20]

Leverag- ing information in spatial transcriptomics to predict super- resolution gene expression from histology images in tumors

Minxing Pang, Kenong Su, and Mingyao Li. Leverag- ing information in spatial transcriptomics to predict super- resolution gene expression from histology images in tumors. BioRxiv, pages 2021–11, 2021. 6

2021
[21]

Scalable diffusion models with transformers

William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 4195–4205,
[22]

stlearn: integrating spatial location, tissue morphol- ogy and gene expression to find cell types, cell-cell inter- actions and spatial trajectories within undissociated tissues

Duy Pham, Xiao Tan, Jun Xu, Laura F Grice, Pui Yeng Lam, Arti Raghubar, Jana Vukovic, Marc J Ruitenberg, and Quan Nguyen. stlearn: integrating spatial location, tissue morphol- ogy and gene expression to find cell types, cell-cell inter- actions and spatial trajectories within...

2020
[23]

High-resolution image synthesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 4

2022
[24]

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

Chen Shengquan, Zhang Boheng, Chen Xiaoyang, Zhang Xuegong, and Jiang Rui. stplus: a reference-based method for the accurate enhancement of spatial transcriptomics. Bioinformatics, 37(Supplement 1):i299–i307, 2021. 1, 2

2021
[25]

Structure-based protein and small molecule generation using egnn and diffusion models: A comprehensive review

Farzan Soleymani, Eric Paquet, Herna Lydia Viktor, and Wojtek Michalowski. Structure-based protein and small molecule generation using egnn and diffusion models: A comprehensive review. Computational and Structural Biotechnology Journal, 2024. 2

2024
[26]

Spatial transcriptomics: Technologies, applications and ex- perimental considerations

Ye Wang, Bin Liu, Gexin Zhao, YooJin Lee, Anton Buzdin, Xiaofeng Mu, Joseph Zhao, Hong Chen, and Xinmin Li. Spatial transcriptomics: Technologies, applications and ex- perimental considerations. Genomics, 115(5):110671, 2023. 1

2023
[27]

Spatially resolved gene expression prediction from histology images via bi- modal contrastive learning

Ronald Xie, Kuan Pang, Sai Chung, Catia Perciani, Sonya MacParland, Bo Wang, and Gary Bader. Spatially resolved gene expression prediction from histology images via bi- modal contrastive learning. Advances in Neural Information Processing Systems, 36:70626–70637, 2023. 1, 6

2023
[28]

Integration tools for scrna-seq data and spatial transcriptomics sequencing data

Chaorui Yan, Yanxu Zhu, Miao Chen, Kainan Yang, Feifei Cui, Quan Zou, and Zilong Zhang. Integration tools for scrna-seq data and spatial transcriptomics sequencing data. Briefings in Functional Genomics, 23(4):295–302, 2024. 3

2024
[29]

Exemplar guided deep neural network for spatial tran- scriptomics analysis of gene expression prediction

Yan Yang, Md Zakir Hossain, Eric A Stone, and Shafin Rah- man. Exemplar guided deep neural network for spatial tran- scriptomics analysis of gene expression prediction. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 5039–5048, 2023. 1, 6

2023
[30]

Spatial transcriptomics analysis of gene expression pre- diction using exemplar guided graph neural network.Pattern Recognition, 145:109966, 2024

Yan Yang, Md Zakir Hossain, Eric Stone, and Shafin Rah- man. Spatial transcriptomics analysis of gene expression pre- diction using exemplar guided graph neural network.Pattern Recognition, 145:109966, 2024. 1, 6

2024
[31]

Dif- fusion generative modeling for spatially resolved gene ex- pression inference from histology images

Sichen Zhu, Yuchen Zhu, Molei Tao, and Peng Qiu. Dif- fusion generative modeling for spatially resolved gene ex- pression inference from histology images. arXiv preprint arXiv:2501.15598, 2025. 2

2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Spage: spatial gene enhancement using scrna-seq

Tamim Abdelaal, Soufiane Mourragui, Ahmed Mahfouz, and Marcel JT Reinders. Spage: spatial gene enhancement using scrna-seq. Nucleic acids research, 48(18):e107–e107, 2020. 1, 2

work page 2020

[2] [2]

Autoencoders and their applications in machine learning: a survey

Kamal Berahmand, Fatemeh Daneshfar, Elaheh Sadat Salehi, Yuefeng Li, and Yue Xu. Autoencoders and their applications in machine learning: a survey. Artificial Intelli- gence Review, 57(2):28, 2024. 2

work page 2024

[3] [3]

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

Tommaso Biancalani, Gabriele Scalia, Lorenzo Buffoni, Raghav Avasthi, Ziqing Lu, Aman Sanger, Neriman Tokcan, Charles R Vanderburg, ˚Asa Segerstolpe, Meng Zhang, et al. Deep learning and alignment of spatially resolved single-cell transcriptomes with tangram. Nature methods, 18(11):1352– 1362, 2021. 1

work page 2021

[4] [4]

Spatially resolved transcrip- tomics: a comprehensive review of their technological ad- vances, applications, and challenges

Mengnan Cheng, Yujia Jiang, Jiangshan Xu, Alexios- Fotios A Mentis, Shuai Wang, Huiwen Zheng, Sunil Kumar Sahu, Longqi Liu, and Xun Xu. Spatially resolved transcrip- tomics: a comprehensive review of their technological ad- vances, applications, and challenges. Journal of Genetics and Genomics, 50(9):625–640, 2023. 1

work page 2023

[5] [5]

A point cloud segmentation frame- work for image-based spatial transcriptomics

Thomas Defard, Hugo Laporte, Mallick Ayan, Juliette Soulier, Sandra Curras-Alonso, Christian Weber, Florian Massip, Jos ´e-Arturo Londo ˜no-Vallejo, Charles Fouillade, Florian Mueller, et al. A point cloud segmentation frame- work for image-based spatial transcriptomics. Communica- tions Biology, 7(1):823, 2024. 1

work page 2024

[6] [6]

A practical guide to single-cell rna- sequencing for biomedical research and clinical applications

Ashraful Haque, Jessica Engel, Sarah A Teichmann, and Tapio L ¨onnberg. A practical guide to single-cell rna- sequencing for biomedical research and clinical applications. Genome medicine, 9:1–12, 2017. 1

work page 2017

[7] [7]

Integrating spatial gene expression and breast tumour morphology via deep learning

Bryan He, Ludvig Bergenstr ˚ahle, Linnea Stenbeck, Abubakar Abid, Alma Andersson, ˚Ake Borg, Jonas Maaskola, Joakim Lundeberg, and James Zou. Integrating spatial gene expression and breast tumour morphology via deep learning. Nature biomedical engineering , 4(8): 827–834, 2020. 1, 6

work page 2020

[8] [8]

Denoising dif- fusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2

2020