Pith. sign in

REVIEW 4 major objections 6 minor 72 references

SPATIA: Multimodal Generation and Prediction of Spatial Cell Phenotypes

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SPATIA claims that confidence-reweighted flow matching over optimal-transport pseudo-pairs can generate target-state cell morphologies from unpaired spatial transcriptomics, reporting an 8% gain in generative fidelity and up to 3% in…

desk verdict The MIST atlas and fusion architecture are solid; the perturbation-generation claim is only distribution-level, and the paper overstates it. read the letter →

arxiv 2507.04704 v3 pith:WXJM7D3D submitted 2025-07-07 q-bio.QM cs.AIcs.CV

classification q-bio.QMcs.AIcs.CV
keywords spatialtranscriptomicscellmorphologyflowmatchingoptimaltransportmultimodalrepresentationlearninggenerativemodelingphenotypeprediction
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 sets out to show that one model can represent cell morphology, gene expression, and spatial context at three scales — single cell, local niche, and whole tissue — and can use those representations both for prediction and for generating how cell morphology changes under a biological perturbation. The obstacle is that destructive spatial transcriptomics never images the same cell before and after a transition, so the model builds weak control–target pairs by entropy-regularized optimal transport in gene-expression space between same-lineage cells in similar niches, then learns a confidence-weighted flow-matching objective that transports control latents toward target-state distributions. A morphology-profile alignment loss anchors generated images to interpretable phenotypic features, and a condition-contrastive loss keeps distinct transitions separable. Trained on a newly assembled atlas of 25.9 million cell–gene pairs across 17 tissues, the model reports an 8% improvement in generative fidelity and up to 3% predictive gains over 18 baselines on 12 tasks. If correct, this means unpaired spatial data suffice to simulate microenvironment-dependent transitions such as DCIS-to-invasive progression.

What carries the argument

The load-bearing object is the confidence-aware conditional flow-matching objective. For a control–target pair in latent space, the linear bridge $\ell_\lambda = (1-\lambda)\ell_{\mathrm{ctrl}} + \lambda \ell_{\mathrm{tgt}}$ is sampled with $\lambda \sim U(0,1)$, and a velocity field $v_\theta(\ell_\lambda, \lambda \mid z_{\mathrm{cond}})$ is trained against the endpoint displacement $u = \ell_{\mathrm{tgt}} - \ell_{\mathrm{ctrl}}$, weighted by $w(x_{\mathrm{ctrl}}) = c^\gamma / \mathbb{E}[c^\gamma]$, where $c = \max_j P^*_{ij}$ is the coupling strength of the optimal-transport plan. Three adjoints carry the biological content: the spatial perturbation embedding built from mean gene- and morphology-shift signatures per transition; the morphology-profile alignment loss, a sliced Wasserstein distance between generated and real target CellProfiler-feature distributions in a frozen encoder space; and the condition-contrastive loss that prefers the true transition condition over a replaced one. Together they convert distribution-level, unpaired supervision into a generative model of phenotype change.

What would settle it

A lineage-tracking or live-imaging experiment that records the same cell before and after a genuine transition would settle it: run SPATIA's control-to-target generator on the pre-transition cohort and compare the predicted endpoint morphology distribution with the actually observed post-transition morphology distribution of the tracked cells; if the two diverge beyond the reported KS and Wasserstein margins, the weak-pair assumption is the point of failure.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single model can learn both predictive and generative representations of spatial cell phenotypes, and that the generative half in particular can produce target-state morphology distributions for perturbations such as DCIS-to-invasive tumor progression and immune-cold-to-immune-hot remodeling, even though the same cell is never observed before and after the transition. The construction forms weak control–target pairs through entropy-regularized optimal transport in PCA gene-expression space, restricted to same-lineage cells in spatially similar niches, and trains a conditional velocity field on latent bridges between control and target embeddings, weighting each pair by its optimal-transport coupling strength so that uncertain matches contribute less. A morphology-profile alignment term matches generated cells to target-state CellProfiler feature distributions via sliced Wasserstein distance, and a condition-contrastive margin loss keeps different transition conditions distinguishable. Benchmarked on a 25.9-million-cell atlas across 12 tasks with donor-disjoint splits, SPATIA reports 8% higher generative fidelity and up to 3% predictive gains over 18 baseline models.

Load-bearing premise

The generation pipeline assumes that an optimal-transport match in gene-expression space between two same-lineage cells in similar niches points along the real direction of the biological transition, even though the true before-and-after transition is never observed for any single cell.

Editorial extensions

If this is right

  • Weak pseudo-pairs plus confidence-reweighted flow matching can approximate target-state morphology distributions without any paired before/after observations.
  • A single hierarchical model can serve both prediction and controllable generation from spatial transcriptomics; the reported predictive gains persist alongside the generative ones.
  • Spatial context carries real signal for generation: removing the niche level raises FID from 59.2 to 60.3 and KID from 2.04 to 2.24, so the flow is not merely exploiting dataset co-occurrence.
  • The generator degrades gracefully under pairing noise (baseline FID 59.1 rising to 69.4 under 40% within-slide corruption), indicating it does not depend on brittle one-to-one optimal-transport matches.
  • Aligning generation to an interpretable phenotype-feature space rather than raw pixels improves biological correctness as measured by KS and Wasserstein statistics on CellProfiler features.

Reading between the lines

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

  • The confidence-reweighted flow objective is not tied to spatial transcriptomics: any unpaired distributional-transition problem with a noisy matching step, such as drug-response morphology prediction from bulk transcriptomic shift signatures, could use the same recipe.
  • The reported FID, KS, and Wasserstein metrics certify that generated and real target-state distributions look alike, but they do not certify that each control cell was transported along the true biological path; that per-cell claim would need lineage-resolved or time-lapse data the current atlas does not contain.
  • Because transition descriptors are averages over training pairs, the generator likely extrapolates poorly to transitions absent from the atlas; a transition-similarity or meta-learning layer is a natural extension the paper does not explore.
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 / 6 minor

Summary. The paper introduces SPATIA, a hierarchical multimodal model for image-based spatial transcriptomics that fuses cell morphology, gene expression, and spatial context at cell, niche, and tissue levels. The main methodological novelty is a spatially conditioned flow-matching module that generates target-state cell morphology from unpaired control and target populations, using entropy-regularized optimal transport in gene-expression PCA space to construct weak control-target pairs, confidence-aware reweighting, a condition-contrastive loss, and a morphology-profile alignment loss. The authors assemble a large multi-scale dataset (MIST) of 25.9 million cell-gene pairs across 17 tissues and evaluate SPATIA on 12 tasks, reporting improved generative fidelity (8%) and predictive accuracy (up to 3%) over 18 baselines.

Significance. If the generative claim is correct, SPATIA would be a notable advance: it would demonstrate that weak, unpaired cross-sectional data can be used to learn biologically meaningful control-to-target morphological transitions in spatial transcriptomics, a setting where paired before/after observations are typically unavailable. The MIST dataset itself is a substantial community resource, and the paper is commendable for releasing code, using donor-disjoint splits, and including robustness analyses. However, the core generative claim rests on distributional evaluation in a feature space that is also used for training, and the paper lacks a critical control-free ablation; these issues must be resolved before the headline claim can be accepted. The predictive benchmarking is extensive and appears sound, but it is not the paper's principal novelty.

major comments (4)
  1. [Section 3.3, Appendix C.2, Table 1] The claim that SPATIA can 'faithfully simulate microenvironment-dependent changes' is not supported by the reported evidence. All generative metrics (FID/KID and CellProfiler Wasserstein/KS statistics) compare the generated population to the target-state population. Because OT pseudo-pairs are formed between cross-sectional control and target cells and no cell is ever observed before and after perturbation, these metrics cannot verify that the per-cell displacement u = ℓ_tgt − ℓ_ctrl (Eq. 7) corresponds to the actual biological transition direction. The paper should either (a) validate on a dataset with true paired observations or a strong proxy (e.g., known marker shifts in generated images, independent pathological annotation, or time-course data), or (b) explicitly limit the claim to target-state distribution modeling rather than faithful per-cell transition simulation. The Impact Statement's caution that outputs are 'predictive simulations rather than experimentally validated outcomes' is in tension with the abstract and conclusion.
  2. [Section 3.4, Table 1] The morphology-profile alignment loss Lmorph = SWD(D_gen, D_real) is computed in the feature space of a frozen encoder φ that regresses CellProfiler features, and Table 1 evaluates 'morphology correctness' with CellProfiler-derived Wasserstein correlation and KS statistics. This makes the generation evaluation directly circular: the training objective optimizes the same empirical feature distributions used as the headline metric. The reported improvements in morphology correctness are therefore partly an artifact of the optimization. Please evaluate with an independent feature set (e.g., different morphological descriptors, marker-specific cell measurements, or expert scoring) and also report an ablation without Lmorph.
  3. [Section 5, Fig. 7] The paper states that a 'cell-only variant' of SPATIA was tested to determine whether the conditional flow exploits dataset co-occurrence rather than genuine spatial/instance conditioning, but no quantitative results for this ablation are reported; Fig. 7 and Table 12 instead show OT-pairing corruption robustness. A control-free or control-shuffled ablation is essential to establish that generated output depends on the specific control cell rather than only on the transition token and target-state statistics. Without such an ablation, the instance-specific 'control-to-target' interpretation is not justified. Please report this ablation quantitatively.
  4. [Table 1, Section 4.2] The KS statistic is a distance measure: a higher value indicates greater dissimilarity between two distributions, so reporting 'KS↑' as better morphology correctness is inverted. As printed, SPATIA's higher KS values (0.65 vs. 0.58 for GeneFlow) would mean the generated CellProfiler feature distributions are less similar to the target distributions than those of baselines, contradicting the text's interpretation. Please clarify whether the reported statistic is actually 1−KS or a different quantity, and correct the table and discussion accordingly. If the standard KS distance is used, the 'morphology correctness' conclusions may reverse.
minor comments (6)
  1. [Abstract, Section 4.2] The abstract states 'improving generative fidelity by 8%,' but the reported FID improvements in Table 1 are approximately 4.6% relative to GeneFlow (59.5 vs. 62.4); please specify the baseline and calculation used for the 8% figure.
  2. [Eq. (4) and Eq. (14)] The confidence exponent is denoted γ in Eq. (4) and α in Eq. (14); please unify the notation.
  3. [Section 3.3, Eq. (5)] The text says Δm is computed once per transition from training pairs and 'does not require per-sample target morphology features as input at inference, avoiding target leakage,' but the transition descriptor is derived from target-state CellProfiler features in the training set; please clarify how this is not a form of target-state supervision for the transition conditioning.
  4. [Table 1] The metric 'Wass. Corr.' is undefined; please specify whether it is the Pearson correlation between per-feature Wasserstein distances, a correlation of feature means, or another quantity.
  5. [Section 4.1] The sentence 'Each MIST-C example consists of this uint8 image patch paired with the per-cell transcript vector of a single gene' appears to be a typo; it should read 'the per-cell transcript vector' or 'a single cell's transcript vector.'
  6. [Appendix E, Table 9] The table header and entries use 'SPATCH' (e.g., 'SPATCH Stereo-seq'), while the dataset is called MIST; please correct the label.

Circularity Check

1 steps flagged · score 6.0 of 10

Morphology-correctness evaluation is partly circular: the morphology-profile alignment loss is trained in the same CellProfiler feature space used for the reported Wasserstein/KS metrics.

  1. fitted input called prediction [Section 3.4 (Morphology-Profile Alignment) and Section 4.2 / Table 1 (Morphology correctness evaluation)]
    "We propose a morphology-profile alignment loss in the same feature space used for evaluation. Let phi(.) be a differentiable morphology encoder pretrained to regress CellProfiler features; we freeze phi during training to avoid representation drift. For each mini-batch, we define two empirical distributions D in morphology feature space Dgen={phi(xhat_tgt)} and Dreal={phi(x_tgt)}, and align them using sliced Wasserstein distance. ... Morphological correctness is assessed using CellProfiler-derived features ..."

    The reported morphology-correctness columns (Wass. Corr., KS) in Table 1 are computed on CellProfiler features of generated versus real target-state cells. The training objective Lmorph is exactly a distributional alignment (sliced Wasserstein) between phi(generated) and phi(real) in that same CellProfiler feature space. Therefore the improvement attributed to '+ Morph. Loss' in Table 1 is, by construction, a re-measurement of the thing being optimized: the model is trained to shrink the distance between generated and real CellProfiler feature distributions and then scored by distances between those same feature distributions.

full rationale

The central generation pipeline is otherwise self-contained: OT pairing, confidence reweighting, and flow matching are defined from the paper's own equations, and the FID/KID image-fidelity metrics use Inception features that are not optimized by any training loss. The predictive benchmarks (annotation, clustering, receptor status, gene-expression prediction) are evaluated against external or standard datasets and do not reduce to the training objectives. The one significant circularity is explicitly admitted in Section 3.4: the morphology-profile alignment loss is placed 'in the same feature space used for evaluation,' and Table 1's morphology-correctness statistics are computed on CellProfiler features of the same target-state distributions. Thus the '+ Morph. Loss' improvement is partially guaranteed by construction. The paper's unpaired-control/target assumption (Appendix C.1) is a genuine scientific limitation but not a circularity; the Impact Statement appropriately cautions that outputs are 'predictive simulations rather than experimentally validated outcomes.' No load-bearing self-citations or imported uniqueness theorems appear. Score 6 reflects one evaluation axis that reduces partly to its own training objective while the central method retains independent content via image fidelity and predictive tasks.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

No new physical entity is introduced; the transition descriptor zpert and confidence weights are model-internal constructs, and the MIST atlas is an aggregate of public datasets. The central generation claim rests on the unverified assumption that cross-sectional OT pseudo-pairs in gene expression space are faithful proxies for longitudinal perturbation, and on an evaluation metric that is partially optimized during training through the morphology-profile alignment loss.

free parameters (7)
  • gamma (confidence reweighting exponent) = not reported
    Eq 4 in Section 3.3 (written as alpha in Appendix C.3 Eq 14); controls downweighting of uncertain OT pairs; no sensitivity analysis given.
  • rho (condition-contrastive loss weight) = not reported
    Section 3.4 total loss L = L_FM + rho L_cond + lambda_morph L_morph; no value or ablation reported.
  • lambda_morph (morphology-profile alignment weight) = not reported
    Section 3.4; controls strength of the sliced Wasserstein morphology alignment; no ablation reported.
  • OT entropy regularization epsilon = not reported
    Appendix C.2 Eq 11; sets softness of the coupling; no value given in the paper.
  • PCA dimension d for OT preprocessing = not reported
    Appendix C.2; gene expression vectors are projected to top d principal components; d is not specified.
  • theta_min (minimum cells per state) = 50
    Appendix C.1 quality control criterion; cell states with fewer than 50 cells are excluded from pairing.
  • niche and tissue tile sizes = 256x256 px niche, 1024x1024 px tissue
    Section 3.1 and Appendix E; defines the multi-scale hierarchy; niche size is asserted to contain roughly 10 to 30 cells.
assumptions (6)
  • domain assumption Cross-sectional OT pseudo-pairs in PCA gene-expression space are a valid proxy for unobserved longitudinal control-to-target transitions.
    Section 3.3 and Appendix C.1; the entire generation module is trained on these pseudo-pairs.
  • domain assumption Entropy-regularized OT coupling strength (max P*) reflects pairing uncertainty.
    Section 3.3 Eq 3; used to reweight flow-matching supervision.
  • domain assumption A frozen encoder phi trained to regress CellProfiler features provides an appropriate biological morphology space for both training and evaluation.
    Section 3.4 and Section 4.2; used for the alignment loss and the reported morphology-correctness metrics.
  • domain assumption The same cell is never observed before and after perturbation, so distribution-level target-state generation is the correct objective.
    Section 3.3 Weak Pair Construction and Appendix C.1.
  • domain assumption Donor-disjoint splits prevent cross-modal leakage between training and test.
    Section 4.1; relies on donor identity being the dominant source of morphological variation.
  • domain assumption Pretrained scPRINT and ImageNet ViT-MAE provide transferable priors; random initialization degrades performance.
    Appendix A and Table 11; used as pretrained initialization and retained during training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPATIA: Multimodal Generation and Prediction of Spatial Cell Phenotypes." pith.science (2026). https://pith.science/paper/WXJM7D3D

@misc{pith2026250704704,
  author       = {Pith},
  title        = {Pith review of: SPATIA: Multimodal Generation and Prediction of Spatial Cell Phenotypes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXJM7D3D}},
  note         = {Machine review of arXiv:2507.04704}
}
read the original abstract

Understanding how cellular morphology, gene expression, and spatial context jointly shape tissue function is a central challenge in biology. Image-based spatial transcriptomics technologies now provide high-resolution measurements of cell images and gene expression profiles, but existing methods typically analyze these modalities in isolation or at limited resolution. We address the problem by introducing SPATIA, a multi-level generative and predictive model that learns unified, spatially aware representations by fusing morphology, gene expression, and spatial context from the cell to the tissue level. SPATIA also incorporates a spatially conditioned generative framework with confidence-aware OT reweighting and morphology-profile alignment for modeling target-state morphology distributions. Specifically, we propose a confidence-aware flow matching objective that reweights weak optimal-transport pairs based on uncertainty. We further apply morphology-profile alignment to encourage biologically meaningful image generation, enabling the modeling of microenvironment-dependent phenotypic transitions. We assembled a multi-scale dataset consisting of 25.9 million cell-gene pairs across 17 tissues. We benchmark SPATIA against 18 models across 12 tasks, spanning categories such as phenotype generation, annotation, clustering, gene imputation, and cross-modal prediction. SPATIA achieves improved performance over state-of-the-art models, improving generative fidelity by 8% and predictive accuracy by up to 3%.

Figures

Figures reproduced from arXiv: 2507.04704 by the authors.

Figure 1
Figure 1. SPATIA is a multimodal generative and predictive model of spatial cell phenotypes. dependent morphological changes under perturbations. Un￾like generic image synthesis, modeling these effects requires generative approaches that respect both the intrinsic cell state and the extrinsic spatial niche. Present Work. We introduce SPATIA1 , a multi-level model for generative and predictive modeling of spatial cell phe￾noty… view at source ↗
Figure 2
Figure 2. Example of the three levels of MIST dataset. define the niche as the local spatial neighborhood of cell i. SPATIA addresses two coupled objectives: Unified Representation Learning. We aim to learn a fusion encoder F that maps a cell and its spatial context to a unified embedding zi = F(xi , gi , si). This embedding integrates intrinsic features (cell image and gene tokens) with extrinsic context (niche and tissue-le… view at source ↗
Figure 3
Figure 3. A) Tissue distribution of our MIST dataset. B) A landscape showcasing the variety of disease states in MIST. C) MIST contains four platforms containing different tissue and organ types. D) Overview of SPATIA. E) Processing controltarget pairs with optimal transport. F) Our conditional contrastive flow matching approach for predicting cell morphology. G) Downstream task performance gain compared to existing models. w… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative image analysis of generated cell morphology change images with the target image. Target denotes real cells sampled from the target state distribution [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Batch-effect mitigation. Different colors represent differ￾ent datasets or sample sources. Tab. 1 highlights both the image fidelity and biological va￾lidity of our generative framework. Compared to GeneFlow, CellFlux, and MorphDiff, our method achieves lower FID and K…
Figure 6
Figure 6. Figure 6: Gene expression prediction [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Robustness evaluation of SPATIA in OT-based control￾perturbed matching. ∼ 40% within-slide corruption and 20% cross-slide corrup￾tion in Tab. 12. SPATIA degrades gracefully and remains competitive with baselines under substantial pairing noise. 6. Conclusion We present…
Figure 8
Figure 8. Figure 8: Embeddings before batch-effect mitigation. UMAP visualization of PCA-level embeddings prior to SPATIA normalization, multimodal pretraining, and batch-invariant representation learning. Cells cluster primarily by dataset/source, indicating dataset-driven variation befo…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 54 canonical work pages

  1. [1]

    scmulan: A multitask generative pre-trained language model for single-cell analysis

    Haiyang Bian, Yixin Chen, Xiaomin Dong, Chen Li, Minsheng Hao, Sijie Chen, Jinyi Hu, Maosong Sun, Lei Wei, and Xuegong Zhang. scmulan: A multitask generative pre-trained language model for single-cell analysis. In Research in Computational Molecular Biology (RECOMB) 2024 , volume 14758 of Lecture Notes in Computer Science , pages 479--482. Springer, 2024

  2. [2]

    S. Birk, I. Bonafonte-Pard \`a s, A. M. Feriz, and et al. Quantitative characterization of cell niches in spatially resolved omics data. Nature Genetics , 2025

  3. [3]

    Token merging: Your vit but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 , 2022

  4. [4]

    Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations

    Srinivas Niranj Chandrasekaran, Beth A Cimini, Amy Goodale, Lisa Miller, Maria Kost-Alimova, Nasim Jamali, John G Doench, Briana Fritchman, Adam Skepner, Michelle Melanson, et al. Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations. Nature Methods , 21(6):1114--1121, 2024

  5. [5]

    Combining spatial transcriptomics with tissue morphology

    Eduard Chelebian, Christophe Avenel, and Carolina Wählby. Combining spatial transcriptomics with tissue morphology. Nature Communications , 2025

  6. [6]

    Stimage-1k4m: A histopathology image-gene expression dataset for spatial transcriptomics

    Jiawen Chen, Muqing Zhou, Wenrong Wu, Jinwei Zhang, Yun Li, and Didong Li. Stimage-1k4m: A histopathology image-gene expression dataset for spatial transcriptomics. arXiv preprint arXiv:2406.06393 , 2024

  7. [7]

    Spatially resolved, highly multiplexed rna profiling in single cells

    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

  8. [8]

    Chen, Chengkuan Chen, Yicong Li, Tiffany Y

    Richard J. Chen, Chengkuan Chen, Yicong Li, Tiffany Y. Chen, Andrew D. Trister, Rahul G. Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 16144--16155, June 2022

Show all 72 references
  1. [9]

    Chen, Tong Ding, Ming Y

    Richard J. Chen, Tong Ding, Ming Y. Lu, Drew F.K. Williamson, Guillaume Jaume, Bowen Chen, Andrew Zhang, Daniel Shao, Andrew H. Song, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine , 2024

  2. [10]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages 785--794, 2016

  3. [11]

    scgpt: Towards building a foundation model for single-cell multi-omics using generative ai

    Haotian Cui, Cheng Wang, Han Maan, Kai Pang, Fei Luo, and Bo Wang. scgpt: Towards building a foundation model for single-cell multi-omics using generative ai. bioRxiv , page 2023.04.30.538439, 2023

  4. [12]

    Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

  5. [13]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. Advances in Neural Information Processing Systems , 34:8780--8794, 2021

  6. [14]

    Wagner, Andrew H

    Tong Ding, Sophia J. Wagner, Andrew H. Song, Richard J. Chen, Ming Y. Lu, Andrew Zhang, Anurag J. Vaidya, Guillaume Jaume, Muhammad Shaban, et al. Multimodal whole slide foundation model for pathology. In arXiv preprint arXiv:2411.19666 , 2024

  7. [15]

    Cross-tissue immune cell analysis reveals tissue-specific features in humans

    C Dom \' nguez Conde, Chao Xu, Louie B Jarvis, Daniel B Rainbow, Sara B Wells, Tamir Gomes, SK Howlett, O Suchanek, K Polanski, HW King, et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science , 376(6594):eabl5197, 2022

  8. [16]

    Laurent, Anqi Shao, Mar \' a del Mar Alvarez-Torres, Tianji Yu, Jimin Tan, Jiayu Su, Romella Sagatelian, Adolfo A

    Xi Fu, Shentong Mo, Alejandro Buend \' a, Anouchka P. Laurent, Anqi Shao, Mar \' a del Mar Alvarez-Torres, Tianji Yu, Jimin Tan, Jiayu Su, Romella Sagatelian, Adolfo A. Ferrando, Alberto Ciccia, Yanyan Lan, David M. Owens, Teresa Palomero, Eric P. Xing, and Raul Rabadan. A fou...

  9. [17]

    Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential

    Michael Gadermayr and Maximilian Tschuchnig. Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations & future potential. Computerized Medical Imaging and Graphics , 112:102337, 2024

  10. [18]

    xtrimogene: An efficient and scalable representation learner for single-cell rna-seq data

    Jing Gong, Minsheng Hao, Xingyi Cheng, Xin Zeng, Chiming Liu, Jianzhu Ma, Xuegong Zhang, Taifeng Wang, and Le Song. xtrimogene: An efficient and scalable representation learner for single-cell rna-seq data. arXiv preprint arXiv:2311.15156, 2023

  11. [19]

    Large-scale foundation model on single-cell transcriptomics

    Minsheng Hao, Jing Gong, Xin Zeng, Chiming Liu, Yucheng Guo, Xingyi Cheng, Taifeng Wang, Jianzhu Ma, Xuegong Zhang, and Le Song. Large-scale foundation model on single-cell transcriptomics. Nature Methods , 21(8):1481--1491, 2024

  12. [20]

    Masked autoencoders are scalable vision learners, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021

  13. [21]

    DePianto, Omar Salem, Tobias Heigl, Nathaniel Diamant, Gabriele Scalia, Tommaso Biancalani, Shannon J

    Graham Heimberg, Tony Kuo, Daryle J. DePianto, Omar Salem, Tobias Heigl, Nathaniel Diamant, Gabriele Scalia, Tommaso Biancalani, Shannon J. Turley, Jason R. Rock, H \'e ctor Corrada Bravo, Josh Kaminker, Jason A. Vander Heiden, and Aviv Regev. A cell atlas foundation model for...

  14. [22]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems , 33:6840--6851, 2020

  15. [23]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 , 2022

  16. [24]

    Stpath: A generative foundation model for integrating spatial transcriptomics and whole slide images

    Tinglin Huang, Tianyu Liu, Mehrtash Babadi, Rex Ying, and Wengong Jin. Stpath: A generative foundation model for integrating spatial transcriptomics and whole slide images. bioRxiv , pages 2025--04, 2025

  17. [25]

    A visual--language foundation model for pathology image analysis using medical twitter

    Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual--language foundation model for pathology image analysis using medical twitter. Nature Medicine , pages 1--10, 2023

  18. [26]

    High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis

    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, et al. High resolution mapping of the tumor microenvironment using integrated single-cell, ...

  19. [27]

    Song, Ming Y

    Guillaume Jaume, Paul Doucet, Andrew H. Song, Ming Y. Lu, Cristina Almagro-P \'e rez, Sophia J. Wagner, Anurag J. Vaidya, Richard J. Chen, Drew F.K. Williamson, Ahrong Kim, and Faisal Mahmood. Hest-1k: A dataset for spatial transcriptomics and histology image analysis. arXiv p...

  20. [28]

    scprint: Pre-training on 50 million cells allows robust gene network predictions

    Jason Kalfon, Jon Samaran, Gabriel Peyr \'e , and Laura Cantini. scprint: Pre-training on 50 million cells allows robust gene network predictions. Nature Communications , 16:3607, 2025

  21. [29]

    Token fusion: Bridging the gap between token pruning and token merging, 2023

    Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging, 2023

  22. [30]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

  23. [31]

    scinterpreter: Training large language models to interpret scrna-seq data for cell type annotation

    Cong Li, Meng Xiao, Pengfei Wang, Guihai Feng, Xin Li, and Yuanchun Zhou. scinterpreter: Training large language models to interpret scrna-seq data for cell type annotation. arXiv preprint arXiv:2402.12405 , 2024

  24. [32]

    Single-cell transcriptome analysis reveals dynamic cell populations and differential gene expression patterns in control and aneurysmal human aortic tissue

    Yanming Li, Pingping Ren, Ashley Dawson, Hernan G Vasquez, Waleed Ageedi, Chen Zhang, Wei Luo, Rui Chen, Yumei Li, Sangbae Kim, et al. Single-cell transcriptome analysis reveals dynamic cell populations and differential gene expression patterns in control and aneurysmal human ...

  25. [33]

    SnapFusion: Text-to-image diffusion model on mobile devices within two seconds

    Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. SnapFusion: Text-to-image diffusion model on mobile devices within two seconds. arXiv preprint arXiv:2306.00980 , 2023

  26. [34]

    Revolutionizing drug discovery: Integrating spatial transcriptomics with advanced computer vision techniques

    Zichao Li, Shiqing Qiu, and Zong Ke. Revolutionizing drug discovery: Integrating spatial transcriptomics with advanced computer vision techniques. In 1st CVPR Workshop on Computer Vision For Drug Discovery (CVDD): Where are we and What is Beyond? , 2025

  27. [35]

    A contrastive deep learning approach to cryptocurrency portfolio with us treasuries

    Zichao Li, Bingyang Wang, and Ying Chen. A contrastive deep learning approach to cryptocurrency portfolio with us treasuries. Journal of Computer Technology and Applied Mathematics , 1(3):1--10, 2024

  28. [36]

    Knowledge graph embedding and few-shot relational learning methods for digital assets in usa

    Zichao Li, Bingyang Wang, and Ying Chen. Knowledge graph embedding and few-shot relational learning methods for digital assets in usa. Journal of Industrial Engineering and Applied Science , 2(5):10--18, 2024

  29. [37]

    St-align: A multimodal foundation model for image-gene alignment in spatial transcriptomics

    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 preprint arXiv:2411.16793 , 2024

  30. [38]

    Lu, Bowen Chen, Drew F.K

    Ming Y. Lu, Bowen Chen, Drew F.K. Williamson, Richard J. Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, et al. A visual\-language foundation model for computational pathology. Nature Medicine , 30(3):863--874, 2024

  31. [39]

    Lu, Bowen Chen, Andrew Zhang, Drew F

    Ming Y. Lu, Bowen Chen, Andrew Zhang, Drew F. K. Williamson, Richard J. Chen, Tong Ding, Long Phi Le, Yung-Sung Chuang, and Faisal Mahmood. Visual language pretrained multiple instance zero-shot transfer for histopathology images. In Proceedings of the IEEE/CVF Conference on C...

  32. [40]

    Multimodal contrastive learning for spatial gene expression prediction using histology images

    Wenwen Min, Zhiceng Shi, Jun Zhang, Jun Wan, and Changmiao Wang. Multimodal contrastive learning for spatial gene expression prediction using histology images. Briefings in Bioinformatics , 25(6):bbae551, 2024

  33. [41]

    Morphodiff: Cellular morphology painting with diffusion models

    Zeinab Navidi, Jun Ma, Esteban Miglietta, Le Liu, Anne E Carpenter, Beth A Cimini, Benjamin Haibe-Kains, and BO WANG. Morphodiff: Cellular morphology painting with diffusion models. In The Thirteenth International Conference on Learning Representations , 2025

  34. [42]

    Hibou: A family of foundational vision transformers for pathology, 2024

    Dmitry Nechaev, Alexey Pchelnikov, and Ekaterina Ivanova. Hibou: A family of foundational vision transformers for pathology, 2024

  35. [43]

    Gest: Towards building a generative pretrained transformer for learning cellular spatial context

    Nima Nouri and [et al.]. Gest: Towards building a generative pretrained transformer for learning cellular spatial context. bioRxiv , page 2025.04.09.648072, 2025

  36. [44]

    schyena: Foundation model for full-length single-cell rna-seq analysis in brain

    Gyutaek Oh, Baekgyu Choi, Inkyung Jung, and Jong Chul Ye. schyena: Foundation model for full-length single-cell rna-seq analysis in brain. arXiv preprint arXiv:2310.02713, 2023

  37. [45]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings , pages 1--10, 2022

  38. [46]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

  39. [47]

    Image super-resolution via iterative refinement

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence , 45(4):4713--4726, 2022

  40. [48]

    H-optimus-0, 2024

    Charlie Saillard, Rodolphe Jenatton, Felipe Llinares-López, Zelda Mariet, David Cahané, Eric Durand, and Jean-Philippe Vert. H-optimus-0, 2024

  41. [49]

    Schaar, Alejandro Tejada-Lapuerta, Giovanni Palla, Robert Gutgesell, Lennard Halle, Mariia Minaeva, Larsen Vornholz, Leander Dony, Francesca Drummer, Mojtaba Bahrami, and Fabian J

    Anna C. Schaar, Alejandro Tejada-Lapuerta, Giovanni Palla, Robert Gutgesell, Lennard Halle, Mariia Minaeva, Larsen Vornholz, Leander Dony, Francesca Drummer, Mojtaba Bahrami, and Fabian J. Theis. Nicheformer: a foundation model for single-cell and spatial omics. bioRxiv , page...

  42. [50]

    Neuronal vulnerability and multilineage diversity in multiple sclerosis

    Lucas Schirmer, Dmitry Velmeshev, Staffan Holmqvist, Max Kaufmann, Sebastian Werneburg, Diane Jung, Stephanie Vistnes, John H Stockley, Adam Young, Maike Steindel, et al. Neuronal vulnerability and multilineage diversity in multiple sclerosis. Nature , 573(7772):75--82, 2019

  43. [51]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 , 2020

  44. [52]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469 , 2023

  45. [53]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 , 2020

  46. [54]

    Visualization and analysis of gene expression in tissue sections by spatial transcriptomics

    Patrik L St hl, Fredrik Salm \'e n, Sanja Vickovic, Anna Lundmark, Jos \'e Fern \'a ndez Navarro, Jens Magnusson, Stefania Giacomello, Michaela Asp, Jakub O Westholm, Mikael Huss, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomic...

  47. [55]

    Cellprofiler 4: improvements in speed, utility and usability

    David R Stirling, Madison J Swain-Bowden, Alice M Lucas, Anne E Carpenter, Beth A Cimini, and Allen Goodman. Cellprofiler 4: improvements in speed, utility and usability. BMC bioinformatics , 22:1--11, 2021

  48. [56]

    Transformers in single-cell omics: a review and new perspectives

    Artur Sza ata, Karin Hrovatin, S \"o ren Becker, Alejandro Tejada-Lapuerta, Haotian Cui, Bo Wang, and Fabian J Theis. Transformers in single-cell omics: a review and new perspectives. Nature methods , 21(8):1430--1443, 2024

  49. [57]

    Transfer learning enables predictions in network biology

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. Transfer learning enables predictions in network biology. Nature , 618(7965):616--624, 2023

  50. [58]

    Dependency-aware deep generative models for multitasking analysis of spatial omics data

    Tian Tian, Jie Zhang, Xiang Lin, Zhi Wei, and Hakon Hakonarson. Dependency-aware deep generative models for multitasking analysis of spatial omics data. Nature Methods , 21:1501--1513, 2024

  51. [59]

    scgpt-spatial: Continual pretraining of single-cell foundation model for spatial transcriptomics

    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 , page 2025.02.05.636714, 2025

  52. [60]

    Jha, and Yuchen Liu

    Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K. Jha, and Yuchen Liu. Attention-driven training-free efficiency enhancement of diffusion models, 2024

  53. [61]

    Transformer-based unsupervised contrastive learning for histopathological image classification

    Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis , 81:102559, 2022

  54. [62]

    Uni-yolo: Vision-language model-guided yolo for robust and fast universal detection in the open world

    Xudong Wang, Weihong Ren, Xi'ai Chen, Huijie Fan, Yandong Tang, and Zhi Han. Uni-yolo: Vision-language model-guided yolo for robust and fast universal detection in the open world. MM '24, page 1991–2000, New York, NY, USA, 2024. Association for Computing Machinery

  55. [63]

    Fine-grained heartbeat waveform monitoring with rfid: A latent diffusion model

    Yiting Wang, Tianya Zhao, and Xuyu Wang. Fine-grained heartbeat waveform monitoring with rfid: A latent diffusion model. In Proceedings of the 3rd International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems , pages 86--91, 2025

  56. [64]

    Nichetrans: Spatial-aware cross-omics translation

    Zhikang Wang, Senlin Lin, Qi Zou, Yan Cui, Chuangyi Han, Yida Li, Jianmin Li, Yi Zhao, Rui Gao, Jiangning Song, et al. Nichetrans: Spatial-aware cross-omics translation. bioRxiv , pages 2024--12, 2024

  57. [65]

    Cellplm: Pre-training of cell language model beyond single cells

    Hongzhi Wen, Wenzhuo Tang, Xinnan Dai, Jiayuan Ding, Wei Jin, and Yuying Xie. Cellplm: Pre-training of cell language model beyond single cells. bioRxiv , page 2023.10.03.560734, 2023

  58. [66]

    Wright, Ari Robicsek, Brian Piening, Carlo Bifulco, Sheng Wang, and Hoifung Poon

    Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, Yanbo Xu, Mu Wei, Wenhui Wang, Shuming Ma, Furu Wei, Jianwei Yang, Chunyuan Li, Jianfeng Gao, Jaylen Rosemon, Tucker Bower, Soohee Lee, Roshant...

  59. [67]

    scbert: large-scale pretrained deep language model for cell type annotation of single-cell rna-seq data

    Fan Yang, Yaoyao Mu, Wen Zhu, Zidong Wang, Xia Guo, Huaqing Yu, and Lei Ni. scbert: large-scale pretrained deep language model for cell type annotation of single-cell rna-seq data. Nature Machine Intelligence , 4(11):852--866, 2022

  60. [68]

    Agp-net: A universal network for gene expression prediction of spatial transcriptomics

    Yan Yang, Xuesong Li, Liyuan Pan, Guoxun Zhang, Liu Liu, and Eric Stone. Agp-net: A universal network for gene expression prediction of spatial transcriptomics. bioRxiv , pages 2025--03, 2025

  61. [69]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023

  62. [70]

    Efficient generation of protein pockets with pocketgen

    Zaixi Zhang, Wan Xiang Shen, Qi Liu, and Marinka Zitnik. Efficient generation of protein pockets with pocketgen. Nature Machine Intelligence , pages 1--14, 2024

  63. [71]

    Hist2cell: Deciphering fine-grained cellular architectures from histology images

    Weiqin Zhao, Zhuo Liang, Xianjie Huang, Yuanhua Huang, and Lequan Yu. Hist2cell: Deciphering fine-grained cellular architectures from histology images. bioRxiv , page 2024.02.17.580852, 2024

  64. [72]

    Diffusion generative modeling for spatially resolved gene expression inference from histology images

    Sichen Zhu, Yuchen Zhu, Molei Tao, and Peng Qiu. Diffusion generative modeling for spatially resolved gene expression inference from histology images. In International Conference on Learning Representations (ICLR) , 2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.