Pith. sign in

REVIEW 3 major objections 32 references

Spatial transcriptomics becomes croppable multi-channel images, unlocking large-scale pretraining that keeps local tissue structure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Cropping ST slides into fixed multi-channel gene patches preserves local spatial context, multiplies training samples, and beats spot- and slice-based pretraining on domain detection.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid practical recipe for ST pretraining units (fixed patches + gene channels) with consistent gains over the authors' spot/multi-spot baselines; geometry preservation is asserted more than measured, and baselines are incomplete. the 3 major comments →

arxiv 2603.13432 v5 pith:5B3YV6W6 submitted 2026-03-13 cs.CV cs.AI

Spatial Transcriptomics as Images for Large-Scale Pretraining

Yishun Zhu , Jiaxin Qi , Jian Wang , Yuhua Zheng , Jianqiang Huang This is my paper
classification cs.CV cs.AI
keywords spatial transcriptomicslarge-scale pretrainingmulti-channel imagespatch constructiongene subset selectionmasked autoencoderspatial domain detectionvision transformer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 large-scale pretraining for spatial transcriptomics has been stuck between two bad sample units: single spots that throw away spatial relationships, and whole tissue slices that are too big and too few. The authors propose treating each tissue slice as a croppable multi-channel image. Fixed-size spatial patches keep local neighborhood structure while multiplying the number of training examples; an importance-aware gene subset along the channel axis keeps input size manageable and training stable. Models pretrained this way outperform spot-level and neighborhood-augmented baselines on spatial domain detection and masked region reconstruction across several human tissue datasets. The practical payoff is a standardized, image-like recipe that lets ordinary vision-style self-supervision scale to spatial biology.

Core claim

Treating spatial transcriptomics as croppable multi-channel images—fixed spatial patches with variance-weighted gene channels—preserves local spatial context while producing enough samples for effective large-scale pretraining, and this construction consistently improves downstream spatial domain detection and masked reconstruction over conventional spot-based and multi-spot schemes.

What carries the argument

The image-like patch construction pipeline: rank-based coordinate compaction of observed spots into a tight lattice, random fixed-size window cropping, and stochastic variance-weighted selection of a fixed number of genes as channels. This unit feeds a masked autoencoder with a vision-transformer backbone.

Load-bearing premise

Compacting staggered spot coordinates into a tight grid and then randomly cropping fixed windows still preserves the biologically meaningful local neighborhoods that the method claims to exploit.

What would settle it

On held-out Visium-style datasets, replace the proposed patches with pure spot-level samples (or with deliberately geometry-warped windows) under otherwise identical architecture and training; if the accuracy and ARI gains on spatial domain detection disappear, the claim that the image-like unit is what carries the improvement fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A standardized patch-and-channel recipe can turn growing public ST collections into reusable pretraining corpora comparable to image datasets.
  • Vision-style masked autoencoders and other computer-vision self-supervision methods become directly applicable to spatial transcriptomics without graph redesign.
  • Moderate spatial window sizes (around 16×16) with roughly 512 gene channels emerge as a practical default balancing context and noise.
  • Downstream spatial domain detection and gene-expression recovery improve under few-shot frozen-encoder evaluation across heterogeneous tissues.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same construction should transfer, with modest retuning of window size, to higher-resolution platforms whose spot lattices are denser than Visium.
  • If patches become the standard unit, multimodal models that already treat histology as images can share the same spatial grid and jointly pretrain image and gene channels.
  • Failure modes of the method will most likely appear on tissues whose functional domains are larger than the chosen window or whose gene variability is poorly captured by simple variance weighting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper argues that the basic training unit for large-scale spatial transcriptomics (ST) pretraining is ill-defined: spot-level sampling discards spatial structure, while whole-slice sampling yields too few, too-large examples. It proposes treating each ST slice as a croppable multi-channel image: after rank-based compaction of observed spots onto a dense lattice (Eq. 5), fixed-size h imes w windows are randomly cropped and a variance-weighted subset of m genes is selected as channels (Eqs. 6–7, Algorithm 1). A ViT-MAE is pretrained with masked reconstruction on ~1.09M spots and evaluated on spatial domain detection (MLP and k-NN) and masked-region reconstruction across six human Visium datasets. Tables 1, 3 and 7 show consistent gains over raw expression, scGPT and scGPT-spatial; ablations (Tables 4–6) identify 16 imes16 imes512 as a strong default and support both spatial patching and channel design.

Significance. If the empirical gains hold under fairer geometry-aware and architecture-matched controls, the work supplies a practical, reproducible recipe for turning heterogeneous ST corpora into image-like pretraining sets that can directly reuse CV backbones (ViT/MAE). That would lower the barrier to large-scale ST foundation models and clarify the sample-unit design space between pure spot and pure slice formulations. Strengths include multi-dataset evaluation, two classifier heads, patch-size/channel ablations, a reconstruction task, and an explicit comparison to a much larger slice-level model (SToFM, Table 2). The contribution is primarily empirical and engineering rather than theoretical.

major comments (3)
  1. §3.2 Sample Construction (Eq. 5) and Q9 assert that rank-based compaction preserves relative positions and does not distort Visium geometry, yet no quantitative check is provided: no comparison of original Euclidean neighbor sets vs. post-compaction neighbors, no adjacency-preservation or distance-correlation metric, and no ablation that freezes the irregular lattice (or uses true physical coordinates) while keeping the same gene-channel and MAE recipe. Because the central claim attributes gains over spot/multi-spot baselines to preserved local spatial context (Tables 1, 3, 7), this missing control is load-bearing; without it the lift could be explained by denser sampling or channel selection alone.
  2. Baselines and architecture isolation (Q8, Tables 1–3): graph-based ST methods (GraphST, STAGATE) are excluded as task-specific, and the only slice-level comparison is a frozen SToFM checkpoint trained on 88M spots with a different architecture. To isolate the effect of the proposed sample unit, the manuscript needs at least one architecture-matched control that applies the same ViT-MAE encoder–decoder to spot-level or multi-spot inputs (and, if feasible, to non-compacted windows). Without that, the contribution of patch construction versus model capacity remains entangled.
  3. Pretraining corpus vs. evaluation protocol (§4.1–4.2): the pretraining set is assembled following scGPT-spatial (~1.09M spots from CELLXGENE/GEO). The paper should state explicitly whether any of the six downstream datasets (or their donors/slices) appear in that corpus, and report a leave-one-dataset-out or donor-held-out check. Overlap would inflate the reported transfer gains and weaken the claim of a general pretraining paradigm.

Circularity Check

0 steps flagged

No circularity: empirical ML methods paper; downstream Acc/ARI/MSE are external held-out metrics, not quantities forced by the pretraining construction or self-citation.

full rationale

The paper proposes a data-organization recipe (rank-based compaction, fixed h×w crops, variance-weighted gene channels) and a standard MAE/ViT masked-reconstruction objective (Eq. 8). Its load-bearing claims are comparative empirical results on external labeled ST benchmarks (spatial domain detection Acc/ARI in Tables 1–6; masked-region MSE/MAE in Table 7) against independently defined baselines (raw expression, scGPT, scGPT-spatial, and a supplied SToFM checkpoint). Nothing in the derivation chain equates a reported metric to a fitted input by construction: patch size and channel count are chosen by ordinary ablation (Tables 4–5), not presented as first-principles predictions of those same quantities. Gene selection (Eq. 7) and the pretraining loss do not define the downstream classifiers or reconstruction scores. Citations to ViT/MAE and prior ST models are external methods literature, not self-authored uniqueness theorems that force the result. Concerns about whether rank compaction (Eq. 5) preserves physical neighborhoods are validity/assumption issues, not circular reductions. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard ML practice plus domain choices about how to rasterize ST and which genes form channels. Free parameters are the usual pretraining hyperparameters selected by ablation. No new physical entities are postulated; the 'image-like patch' is a data representation, not a biological object. Load-bearing domain assumptions are that Visium-like grids can be compacted and cropped without destroying the spatial signal the model is supposed to learn, and that variance-based gene sampling is a valid importance proxy.

free parameters (5)
  • patch spatial size (h×w) = 16×16 (recommended)
    Chosen by ablation (Table 4); 16×16 recommended as default. Directly affects sample count, context, and reported accuracy.
  • gene channel count m = 512 (default)
    Fixed channel dimensionality for pretraining; ablated at 128/256/512 (Table 5). Controls capacity and stability.
  • masking ratio = 30%
    MAE pretraining hyperparameter set to 30% without extensive sensitivity analysis in the main text.
  • number of random windows per slice n_win / n_s
    Controls effective pretraining sample multiplicity; listed as a hyperparameter in Algorithm 1 but not fully reported as a sensitivity study.
  • pretraining epochs / LR / batch size = 40 epochs, lr=1e-4, bs=32
    40 epochs, Adam 1e-4, batch 32; standard training knobs that affect final embeddings used in all tables.
axioms (5)
  • domain assumption Rank-based compaction of observed Visium-style coordinates preserves local adjacency sufficiently for spatial representation learning without imputation.
    Invoked in §3.2 Sample Construction step (1) and defended in Q9; if false, patch geometry is distorted.
  • domain assumption Local fixed windows capture the spatial dependencies that matter for domain detection better than isolated spots or whole-slice samples under fixed compute.
    Core modeling premise of §1 and §3.2; the experiments test consequences but do not prove uniqueness of this scale.
  • domain assumption Per-window gene variance is a valid proxy for gene importance when subsampling channels.
    Eq. (7) and Table 6; weighted sampling vs HVG vs random is empirical, not theoretically derived.
  • domain assumption Masked gene/expression reconstruction with a ViT encoder yields transferable spatial-molecular features for frozen downstream classifiers.
    Standard self-supervised transfer assumption used throughout §4; common in MAE literature but still an assumption.
  • standard math Transformer attention and MSE reconstruction losses are well-defined on real-valued multi-channel tensors.
    Background ML machinery for Eqs. (2)–(8).
invented entities (1)
  • Image-like ST patch unit (h×w×m multi-channel crop with variance-weighted gene channels) no independent evidence
    purpose: Serve as the basic pretraining sample that balances spatial context and sample count.
    Not a new biological object; a data-construction abstraction. Independent evidence is only the paper's own downstream gains, not an external measurement of the entity itself.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial Transcriptomics as Images for Large-Scale Pretraining." pith.science (2026). https://pith.science/paper/5B3YV6W6

@misc{pith2026260313432,
  author       = {Pith},
  title        = {Pith review of: Spatial Transcriptomics as Images for Large-Scale Pretraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5B3YV6W6}},
  note         = {Machine review of arXiv:2603.13432}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Spatial Transcriptomics (ST) profiles thousands of gene expression values at discrete spots with precise coordinates on tissue sections, preserving spatial context essential for clinical and pathological studies. With rising sequencing throughput and advancing platforms, the expanding data volumes motivate large-scale ST pretraining. However, the fundamental unit for pretraining, i.e., what constitutes a single training sample, remains ill-posed. Existing choices fall into two camps: (1) treating each spot as an independent sample, which discards spatial dependencies and collapses ST into single-cell transcriptomics; and (2) treating an entire slide as a single sample, which produces prohibitively large inputs and drastically fewer training examples, undermining effective pretraining. To address this gap, we propose treating spatial transcriptomics as croppable images. Specifically, we define a multi-channel image representation with fixed spatial size by cropping patches from raw slides, thereby preserving spatial context while substantially increasing the number of training samples. Along the channel dimension, we define gene subset selection rules to control input dimensionality and improve pretraining stability. Extensive experiments show that the proposed image-like dataset construction for ST pretraining consistently improves downstream performance, outperforming conventional pretraining schemes. Ablation studies verify that both spatial patching and channel design are necessary, establishing a unified, practical paradigm for organizing ST data and enabling large-scale pretraining.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 1 canonical work pages

  1. [1]

    A spatiotemporal organ-wide gene expression and cell atlas of the developing human heart.Cell, 179(7):1647–1660.e19,

    Michaela Asp, Stefania Giacomello, Ludvig Larsson, Chenglin Wu, Daniel F ¨urth, Xiaoyan Qian, Eva Wrdell, Joaquin Custodio, Johan Reimegrd, and Fredrik Salm ´en. A spatiotemporal organ-wide gene expression and cell atlas of the developing human heart.Cell, 179(7):1647–1660.e19,

  2. [2]

    Spatially resolved transcriptomes—next generation tools for tissue exploration.BioEssays, 42:1900221, 2020

    Michaela Asp, Joseph Bergenstr ˚ahle, and Joakim Lunde- berg. Spatially resolved transcriptomes—next generation tools for tissue exploration.BioEssays, 42:1900221, 2020. 1

    2020

  3. [3]

    Darren J. Burgess. Spatial transcriptomics coming of age. Nature Reviews Genetics, 20(6):317–317, 2019. 1

    2019

  4. [4]

    Boettiger, Jeffrey R

    Kok Hao Chen, Alistair N. Boettiger, Jeffrey R. Moffitt, Siyuan Wang, and Xiaowei Zhuang. Spatially resolved, highly multiplexed rna profiling in single cells.Science, 348 (6233):aaa6090, 2015. 2

    2015

  5. [5]

    Youngmin Chung, Ji Hun Ha, Kyeong Chan Im, and Joo Sang Lee. Accurate spatial gene expression prediction by integrating multi-resolution features.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11591–11600, 2024. 1

    2024

  6. [6]

    scgpt: Towards building a foundation model for single-cell multi-omics using genera- tive ai.bioRxiv, 2023

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, and Bo Wang. scgpt: Towards building a foundation model for single-cell multi-omics using genera- tive ai.bioRxiv, 2023. 6, 7

    2023

  7. [7]

    Deciphering spatial do- mains from spatially resolved transcriptomics with an adap- tive graph attention auto-encoder.Nature Communications, 13(1), 2022

    Kangning Dong and Shihua Zhang. Deciphering spatial do- mains from spatially resolved transcriptomics with an adap- tive graph attention auto-encoder.Nature Communications, 13(1), 2022. 3, 9

    2022

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale.ArXiv, abs/2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale.ArXiv, abs/2010.11929, 2020. 9

    Pith/arXiv arXiv 2010

  9. [9]

    Bryan He, Ludvig Bergenstr ˚ahle, Linnea Stenbeck, Abubakar Abid, Alma Andersson, ˚Ake Borg, Jonas Maaskola, Joakim Lundeberg, and James Y . Zou. Integrat- ing spatial gene expression and breast tumour morphology via deep learning.Nature Biomedical Engineering, 4:827 – 834, 2020. 2

    2020

  10. [10]

    Girshick

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Pi- otr Doll’ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988, 2021. 9

    2022

  11. [11]

    Gottscho, Florian Wagner, Stephen R

    Amanda Janesick, Robert Shelansky, Andrew D. Gottscho, Florian Wagner, Stephen R. Williams, Morgane Rouault, Ghezal Beliakoff, Carolyn A. Morrison, Michelli F. Oliveira, Jordan T. Sicherman, Andrew Kohlway, Jawad Abousoud, Tingsheng Yu Drennon, Seayar H. Mohabbat, 10x Develop- ment Teams, and Sarah E. B. Taylor. High resolution map- ping of the tumor micr...

    2023

  12. [12]

    Enhanc- ing spatial domain detection in spatial transcriptomics with ensdd.Communications Biology, 7(1), 2024

    Hui-Sheng Li, Yu-Ting Tan, and Xiao-Fei Zhang. Enhanc- ing spatial domain detection in spatial transcriptomics with ensdd.Communications Biology, 7(1), 2024. 6

    2024

  13. [13]

    St-align: A multimodal foundation model for image-gene alignment in spatial transcriptomics.ArXiv, abs/2411.16793,

    Yuxiang Lin, Ling Luo, Ying Chen, Xushi Zhang, Zihui Wang, Wenxian Yang, Mengsha Tong, and Rongshan Yu. St-align: A multimodal foundation model for image-gene alignment in spatial transcriptomics.ArXiv, abs/2411.16793,

    Pith/arXiv arXiv

  14. [14]

    Joint cell segmentation and cell type annotation for spatial transcriptomics.Molecu- lar Systems Biology, 17(6), 2021

    Russell Littman, Zachary Hemminger, Robert Foreman, Douglas Arneson, Guanglin Zhang, Fernando G ´omez- Pinilla, Xia Yang, and Roy Wollman. Joint cell segmentation and cell type annotation for spatial transcriptomics.Molecu- lar Systems Biology, 17(6), 2021. 6

    2021

  15. [15]

    Spatially informed clustering, integration, and de- convolution of spatial transcriptomics with graphst.Nature Communications, 14(1), 2023

    Yahui Long, Kok Siong Ang, Mengwei Li, Kian Long Kelvin Chong, Raman Sethi, Chengwei Zhong, Hang Xu, Zhiwei Ong, Karishma Sachaphibulkij, Ao Chen, Li Zeng, Huazhu Fu, Min Wu, Lina Hsiu Kim Lim, Longqi Liu, and Jinmiao Chen. Spatially informed clustering, integration, and de- convolution of spatial transcriptomics with graphst.Nature Communications, 14(1),...

    2023

  16. [16]

    Yahui Long, Kok Siong Ang, Raman Sethi, Sha Liao, Yang Heng, Lynn van Olst, Shuchen Ye, Chengwei Zhong, Hang Xu, Di Zhang, Immanuel Kwok, Nazihah Husna, Min Jian, Lai Guan Ng, Ao Chen, Nicholas R. J. Gascoigne, David Gate, Rong Fan, Xun Xu, and Jinmiao Chen. Deciphering spatial domains from spatial multi-omics with spatialglue. Nature Methods, 21(9):1658–...

    2024

  17. [17]

    Spatiotemporal dy- namics of molecular pathology in amyotrophic lateral scle- rosis.Science, 364(6435):89–93, 2019

    Silas Maniatis, Tarmo ¨Aij¨o, Sanja Vickovic, Catherine Braine, Kristy Kang, Annelie Mollbrink, Delphine Fage- galtier, ˇZaneta Andrusivov´a, Sami Saarenp¨a¨a, Gonzalo Saiz- Castro, Miguel Cuevas, Aaron Watters, Joakim Lundeberg, Richard Bonneau, and Hemali Phatnani. Spatiotemporal dy- namics of molecular pathology in amyotrophic lateral scle- rosis.Scien...

    2019

  18. [18]

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

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

    2021

  19. [19]

    Transcriptome-scale spatial gene expres- sion in the human dorsolateral prefrontal cortex.Nature neu- roscience, 24(3):425–436, 2021

    Kristen R Maynard, Leonardo Collado-Torres, Lukas M We- ber, Cedric Uytingco, Brianna K Barry, Stephen R Williams, Joseph L Catallini, Matthew N Tran, Zachary Besich, Mad- havi Tippani, et al. Transcriptome-scale spatial gene expres- sion in the human dorsolateral prefrontal cortex.Nature neu- roscience, 24(3):425–436, 2021. 6

    2021

  20. [20]

    Miller, Jean Fan, and Daniel Jacob Tward

    Michael I. Miller, Jean Fan, and Daniel Jacob Tward. Multi scale diffeomorphic metric mapping of spatial transcrip- tomics datasets.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4467–4475, 2021. 1

    2021

  21. [21]

    Moffitt, Dhananjay Bambah-Mukku, Stephen W

    Jeffrey R. Moffitt, Dhananjay Bambah-Mukku, Stephen W. Eichhorn, Eric Vaughn, Karthik Shekhar, Julio D. Perez, Nimrod D. Rubinstein, Junjie Hao, Aviv Regev, Catherine Dulac, and Xiaowei Zhuang. Molecular, spatial, and func- tional single-cell profiling of the hypothalamic preoptic re- gion.Science, 362(6416):eaau5324, 2018. 2

    2018

  22. [22]

    St ˚ahl, Fredrik Salm ´en, Sanja Vickovic, Anna Lundmark, Jos ´e Fern ´andez Navarro, Jens Magnusson, Ste- fania Giacomello, Michaela Asp, Jakub O

    Patrik L. St ˚ahl, Fredrik Salm ´en, Sanja Vickovic, Anna Lundmark, Jos ´e Fern ´andez Navarro, Jens Magnusson, Ste- fania Giacomello, Michaela Asp, Jakub O. Westholm, Mikael Huss, Annelie Mollbrink, Sten Linnarsson, Simone Codeluppi, ˚Ake Borg, Fredrik Pont ´en, Paul Igor Costea, Pelin Sahl ´en, Jan Mulder, Olaf Bergmann, Joakim Lunde- berg, and Jonas Fr...

    2016

  23. [23]

    Alberto Valdeolivas, Bettina Amberg, Nicolas Giroud, Mar- ion Richardson, Eric J. C. G ´alvez, Solveig Badillo, Alice Julien-Laferri`ere, Demeter T ´ur´os, Lena V oith von V oithen- berg, Isabelle Wells, Benedek Pesti, Amy A. Lo, Emilio Y´ang¨uez, Meghna Das Thakur, Michael Bscheider, Marc Sultan, Nadine Kumpesa, Bj ¨orn Jacobsen, Tobias Bergauer, Julio S...

    2024

  24. [24]

    Integrating spatial and single-cell transcriptomics data using deep gen- erative models with spatialscope.Nature Communications, 14(1), 2023

    Xiaomeng Wan, Jiashun Xiao, Sindy Sing Ting Tam, Mingx- uan Cai, Ryohichi Sugimura, Yang Wang, Xiang Wan, Zhix- iang Lin, Angela Ruohao Wu, and Can Yang. Integrating spatial and single-cell transcriptomics data using deep gen- erative models with spatialscope.Nature Communications, 14(1), 2023. 2

    2023

  25. [25]

    scgpt-spatial: Continual pretraining of single-cell foundation model for spatial transcriptomics.bioRxiv, pages 2025–02, 2025

    Chloe Xueqi Wang, Haotian Cui, Andrew Hanzhuo Zhang, Ronald Xie, Hani Goodarzi, and Bo Wang. scgpt-spatial: Continual pretraining of single-cell foundation model for spatial transcriptomics.bioRxiv, pages 2025–02, 2025. 2, 3, 5, 6, 7

    2025

  26. [26]

    Construction of a 3d whole organ- ism spatial atlas by joint modelling of multiple slices with deep neural networks.Nature Machine Intelligence, 5(11): 1200–1213, 2023

    Gefei Wang, Jia Zhao, Yan Yan, Yang Wang, Angela Ruo- hao Wu, and Can Yang. Construction of a 3d whole organ- ism spatial atlas by joint modelling of multiple slices with deep neural networks.Nature Machine Intelligence, 5(11): 1200–1213, 2023. 3

    2023

  27. [27]

    Fmh2st: foundation model-based spatial transcriptomics generation from histological images

    Yuequn Wang, Jun Wang, Yanyu Xu, Ning Liu, Bin Liu, Yu- liang Li, and Guoxian Yu. Fmh2st: foundation model-based spatial transcriptomics generation from histological images. Nucleic Acids Research, 53(17):gkaf865, 2025. 3

    2025

  28. [28]

    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

    2024

  29. [29]

    Ritchie, Xiaodong Liu, and Luyi Tian

    Yue You, Yuting Fu, Lanxiang Li, Zhongmin Zhang, Shikai Jia, Shihong Lu, Wenle Ren, Yifang Liu, Yang Xu, Xiaojing Liu, Fuqing Jiang, Guangdun Peng, Abhishek Sampath Ku- mar, Matthew E. Ritchie, Xiaodong Liu, and Luyi Tian. Sys- tematic comparison of sequencing-based spatial transcrip- tomic methods.Nature Methods, 21(9):1743–1754, 2024. 1

    2024

  30. [30]

    Stofm: a multi-scale founda- tion model for spatial transcriptomics.arXiv preprint arXiv:2507.11588, 2025

    Suyuan Zhao, Yizhen Luo, Ganbo Yang, Yan Zhong, Hao Zhou, and Zaiqing Nie. Stofm: a multi-scale founda- tion model for spatial transcriptomics.arXiv preprint arXiv:2507.11588, 2025. 3, 9

    Pith/arXiv arXiv 2025

  31. [31]

    X. Zheng. Datasets used in smart’s experiments. Zenodo dataset, DOI:10.5281/zenodo.17093158, 2025. 6

    work page doi:10.5281/zenodo.17093158 2025

  32. [32]

    Uni-mol: A universal 3d molecular representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. InInternational Conference on Learning Repre- sentations, 2023. 3

    2023

This paper was first reviewed by grok-4.5 on July 14, 2026.