Spatially Gene Expression Prediction using Dual-Scale Contrastive Learning

Donglin Di; Lei Fan; Mingcheng Qu; Tonghua Su; Yang Song; Yue Gao; Yuncong Wu

REVIEW 3 major objections 4 minor 36 references

Spatially Gene Expression Prediction using Dual-Scale Contrastive Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read NH2ST claims that coupling a patch-to-gene translator with a hypergraph-based neighbor branch and dual-scale contrastive learning predicts spatial gene expression from histology images more accurately than existing methods, with Pearson…

desk verdict A plausible neighbor-branch architecture, but the central claim is contradicted by the paper's own Table 1 and the evaluation has enough protocol looseness to make the reported gains unverifiable. read the letter →

arxiv 2506.23827 v1 pith:AGYCCZNZ submitted 2025-06-30 cs.CV

Mingcheng Qu , Yuncong Wu , Donglin Di , Yue Gao , Tonghua Su , Yang Song , Lei Fan This is my paper · ORCID

classification cs.CV

keywords geneexpressionpredictionspatialtranscriptomicswholeslideimagescontrastivelearningcross-attentionhypergraphmultimodalalignmentPearsoncorrelation

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 tries to establish that spatial gene expression can be predicted from routine pathology images more accurately when the model sees not only the target image patch but also its neighboring patches and their gene profiles. To that end it proposes NH2ST, a dual-branch network in which a query branch aligns a patch's pathology and gene features with cross-attention and contrastive learning, while an auxiliary neighbor branch uses hypergraphs to capture higher-order relationships among nearby patches and spots. The two branches share encoders, and at inference only the predictor translator on the pathology branch is used. On six public datasets the paper reports lower MSE and MAE and higher Pearson correlation than prior methods, with the largest PCC gains on the STNet, PCW, and Mouse benchmarks. If those results hold, the method would make spatially resolved expression estimates available from ordinary H&E slides, reducing reliance on costly spatial transcriptomics assays.

What carries the argument

The load-bearing pieces are (i) a cross-attention module $\phi_{ca}$ that lets each modality's features attend to the other's, (ii) InfoNCE contrastive losses $L_s$ and $L_n$ that pull pathology and gene features from the same spot together in a shared space, and (iii) a hypergraph branch in which each node is a neighbor patch and each hyperedge connects a node to its $\tau-1$ most similar nodes, processed by an L-layer HGNN with a pooling layer. A predictor translator, two fully connected layers opposite to the ST encoder, turns pathology features into the 250-gene expression vector and is trained with MSE loss. The neighbor branch is auxiliary: it shares the encoders with the query branch and is dropped at inference, so its role is to shape the pathology encoder during training.

What would settle it

Run NH2ST and its main baselines (ST-Net, BLEEP, TRIPLEX) on the six benchmark datasets under identical k-fold splits, the same 250-gene selection, the same log normalization, and matching seeds, then compare Pearson correlation on the same held-out spots; the central claim fails if the reported PCC margins (for example, 0.572 vs 0.352 on STNet) are not reproduced.

Watch

Extended reading notes

Core claim

NH2ST's central claim is that gene co-expression and tissue morphology are spatially continuous, so ignoring a patch's neighborhood discards signal that is visible in the slide and in nearby transcriptomes. The model couples the direct patch-to-gene translator with a neighbor branch that builds two hypergraphs—one over the K nearest pathology patches, one over their gene expression profiles—where each hyperedge joins a node to its τ−1 most similar nodes. Both graph representations are refined by cross-attention and aligned by InfoNCE contrastive losses at the patch scale and the neighborhood scale. The paper reports that this design outperforms ST-Net, EGN, BLEEP, HistoGene, His2ST, and TRIPLEX across six datasets, achieving a PCC of 0.572 on STNet (vs 0.352 for TRIPLEX), 0.560 on PCW, and 0.593 on Mouse, and it attributes the gains to the neighbor branch plus cross-modal alignment rather than to the encoders alone, which are the standard ResNet18 and two-layer MLP used by prior work.

Load-bearing premise

The performance comparison assumes that the quoted baseline numbers were produced under the same data splits, 250-gene selection, normalization, and evaluation protocol as the authors' own runs; if not, the reported advantage could shrink or disappear.

Editorial extensions

If this is right

Histology-only prediction of spatial gene expression becomes a practical screening tool for tissue microenvironments when spatial transcriptomics assays are too expensive.
Including neighborhood context, not just the target patch, is a necessary ingredient for high correlation between image and expression.
Contrastive alignment at both patch and neighborhood scales improves the pathology encoder's features enough to lift downstream translation.
A trained NH2ST model can be deployed on new WSI patches without paired gene data, since inference uses only the predictor translator.
The achievable accuracy on a given tissue is bounded by how much morphology actually encodes transcription, which the reported PCC values (roughly 0.2–0.6 across benchmarks) suggest varies by tissue.

Reading between the lines

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

A test the paper leaves implicit: because the ablation on PCW shows HGNN depth L=3 and L=4 degrade performance, the neighbor branch may be acting mostly as a local feature smoother; comparing it against a simple K-NN feature-averaging baseline would isolate what hypergraphs specifically add.
The same dual-scale alignment recipe could be applied to other paired tissue readouts—for example, protein expression or methylation arrays registered to H&E—since none of the components are gene-specific.
A fair replication study that runs every baseline under identical splits and gene-selection rules would settle how much of the reported margin is architectural, since results for two of the six datasets are quoted from a prior paper rather than re-run.
The ablation also suggests that contrastive alignment alone, even without the graph branch, already improves over single-patch baselines; a per-dataset decomposition of the neighbor branch's contribution would show where the hypergraph matters most.

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

Punchline: the architecture is plausible, but the paper's central claim is contradicted by its own Table 1. On ZEN, NH2ST is worse than TRIPLEX on all three metrics; on Skin, PCC (0.478) trails TRIPLEX (0.490). So the abstract's "consistently outperforms" is unsupported on the reported numbers.

What's new: a dual-branch trainer with a query branch (cross-attention + InfoNCE) and an auxiliary neighbor branch that uses feature-similarity hypergraphs on patches and expression vectors. That combination is absent from the cited TRIPLEX, BLEEP, and Hist2ST. The ablations on PCW show the neighbor branch buys about 0.1 PCC over the query branch alone, which is a real, if modest, effect. The writing is clear and the method section is reproducible in principle.

Where it wobbles. First, the central claim is false on two of six datasets by the paper's own table. Section 3.1 even says "best performance across three metrics on STNet, Skin, and two STimage-1K4M," but Skin's PCC is below TRIPLEX. Second, all hyperparameter choices (K, L, batch size, lambda1/2, tau) are tuned on PCW, and the final numbers and ablations are also on PCW; without a separate validation split, the PCW gain is selection, not prediction. Third, baselines are uneven: STNet and Skin numbers are quoted from TRIPLEX, and only three baselines were rerun on the other datasets. The code link is unverified here, so the comparison protocol is a black box.

One more point: the limitation note is honest that inference uses only the query branch; the neighbor branch is training-time only. That tempers the "spatial context" claim but is not a flaw by itself.

Who should read it: researchers in histology-to-ST prediction will get the neighbor-branch idea, and the paper is a good case study in how evaluation looseness sinks an otherwise sensible method. I would not cite or build on it as-is, but I would assign a referee rather than desk reject, because the method is coherent and a major revision could repair the evaluation.

Referee Report

3 major / 4 minor

Summary. The paper proposes NH2ST, a dual-branch framework for predicting spatial gene expression from histology patches. A query branch processes a target patch and its paired gene expression vector with cross-attention and contrastive learning; a neighbor branch builds hypergraphs over spatially neighboring patches and their gene profiles and aligns the two modalities at the neighborhood level. The two branches share encoders, and at inference only the query branch's pathology encoder and translator are used. The method is evaluated on six datasets (STNet, Skin, two HEST-1k subsets, two STimage-1K4M subsets) against ST-Net, BLEEP, TRIPLEX, and other baselines, with MSE, MAE, and PCC metrics. The abstract claims that the model 'consistently outperforms existing methods, achieving over 20% in PCC metrics.'

Significance. If the claimed consistent gains were reproducible under a uniform protocol, the method would be practically valuable because it would provide a cheaper, image-only route to spatial gene expression estimates and would give a concrete use of hypergraph and contrastive learning in this setting. The paper also releases code and reports ablation studies, which are useful. However, the headline empirical claim is contradicted by the paper's own Table 1 on two datasets, and the evaluation protocol is not sufficiently controlled for the remaining comparisons. The significance of the contribution therefore cannot be assessed from the reported evidence.

major comments (3)

[§3.1, Table 1] The paper's own Table 1 contradicts the abstract's claim that 'our model consistently outperforms existing methods.' On the ZEN (HEST-1k) dataset, NH2ST reports MSE 0.108±0.01, MAE 0.193±0.01, and PCC 0.216±0.02, while TRIPLEX reports 0.099±0.02, 0.150±0.01, and 0.242±0.03; NH2ST is worse on all three metrics. On Skin, NH2ST's PCC of 0.478±0.11 is below TRIPLEX's 0.490±0.07. Section 3.1's statement that the model achieves 'best performance across three metrics on STNet, Skin, and two STimage-1K4M datasets' is therefore inaccurate for Skin PCC, and the statement that 'PCC results showed significant improvements over previous methods across all datasets' is false for ZEN. Since the central claim of the paper is this consistent superiority, the reported results do not support the conclusion as stated.
[§3.1, 'Results and Discussion'] The comparison protocol is not sufficiently controlled to support the claim of outperformance. For the STNet and Skin datasets, results are quoted from the TRIPLEX paper, while for the other four datasets 'results were reproduced using the released codes.' There is no description of how the data split, the 250-gene selection, the log-normalization, or the cross-validation seeds were matched between NH2ST and the quoted or reproduced baselines. If any of these preprocessing or evaluation choices differ, the reported PCC differences could reflect evaluation differences rather than model quality. The paper should either run all baselines under the exact same code path and protocol or substantially weaken the comparative claims.
[§3.2, Tables 2 and 3] The hyperparameters (K, L, B, λ1, λ2, τ) are selected by sweeping on the PCW dataset, and the final NH2ST row in Tables 2 and 3 reports the best configuration evaluated on the same PCW data. This is a form of test-set selection: the reported PCW numbers are obtained after tuning on the test set itself. The authors should use a held-out validation split or nested cross-validation for hyperparameter selection and then report test-set performance, or explicitly disclose that the reported PCW result is the result of test-set tuning.

minor comments (4)

[§2.2, Eq. (2)] The notation in Eq. (2) is unclear: the denominator uses P−(z_{s_i}, z_{s_j}) with j running to B−1, but the subtraction of the positive pair from the set of negatives inside the log is not shown explicitly. Please rewrite the InfoNCE formula with an explicit positive term and negative set.
[§2.2, Eq. (1)] The symbol d_k is used for the key projection dimension but the denominator is written as √d_k; this is standard unless the key vectors have dimension d_k, so please state that d_k is the dimension of W_k^s h_g^s.
[Table 2] The ablation rows appear to be inconsistent with the final model: in the right part, K=8 and L=2 give MSE 0.033/0.033 and PCC 0.530, while NH2ST is reported as 0.029/0.129/0.560. Please clarify which exact configuration (including all contrastive hyperparameters) corresponds to the NH2ST row so that the ablation and final result are comparable.
[§3.1, Table 1 and Figure 2] The paper uses both 'NH2ST' and 'NH²2ST' in the text and tables; please unify the notation. Also, the phrase 'over 20% in PCC metrics' in the abstract is ambiguous—it is unclear whether this means absolute PCC values above 0.2 or a 20% relative improvement over baselines.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-chain circularity; one mild evaluation-selection loop on the PCW benchmark prevents a fully clean verdict.

fitted input called prediction [Section 3.2 (Ablation Studies), Tables 2 and 3; PCW row of Table 1]
"We explored the impact of different contrastive learning on the PCW dataset (see Table 3). We observed that all three metrics achieved optimal results with a batch size of 8, particularly PCC, which reached 0.550."

Section 3.2 selects the model's hyperparameters (K, L, batch size, lambda1, lambda2, tau) by maximizing performance on the PCW dataset, and the same PCW benchmark is then reported as the final NH2ST result in Table 1 (PCC 0.560). Thus the PCW 'prediction' is not an independent out-of-sample result; it is the outcome of a model-selection procedure run on the test set itself. The PCW component of the 'consistently outperforms' claim is therefore partly fitted by construction. This loop is confined to PCW; the other five datasets are not tuned in this way, so the central comparison retains independent content.

full rationale

This is an empirical systems paper, not a derivation: no theoretical claim is shown to reduce to an earlier equation or to a self-citation. The load-bearing claim is a benchmark comparison against ST-Net, BLEEP, TRIPLEX, and others on six datasets, with metrics reported in Table 1. The only circularity-adjacent step is in Section 3.2, where hyperparameters are chosen by optimizing performance on PCW and then the PCW result is reused as final evidence; this is a mild evaluation-selection loop rather than a constructional equivalence, and it does not affect the other five datasets. The self-citations in the paper (e.g., refs. [13], [20], [21]) are used for method motivation and background, not as load-bearing evidence for the empirical claims. A separate correctness concern, independent of circularity, is that the abstract's 'consistently outperforms' is not supported by the paper's own Table 1 on ZEN and Skin: on ZEN, NH2ST is worse than TRIPLEX on MSE, MAE, and PCC, and on Skin its PCC trails TRIPLEX. That is a factual/validity issue, not a circularity issue.

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

All model hyperparameters were selected on the PCW dataset and then reported on PCW, which makes PCW results partially fitted. The method also assumes that contrastive alignment trained with gene data transfers to gene prediction from images alone, and that feature-similarity hypergraphs capture meaningful spatial biology rather than visual artifacts.

free parameters (5)

K (number of neighbor patches) = 8
Tuned on PCW dataset (Table 2/3); chosen as best among 4, 8, 16, 25.
L (number of HGNN layers) = 2
Tuned on PCW; L=2 best, L>2 degrades.
batch size B = 8
Tuned on PCW; B=8 best in Table 3.
temperature tau = 0.05
Tuned on PCW; best among 0.025, 0.05, 0.1, 0.15, 0.2.
lambda1, lambda2 = 1, 0.5
Tuned on PCW; selected over (0,1), (1,0), (0.5,1), (1,1).

assumptions (4)

domain assumption Cross-modal contrastive alignment learned with paired pathology and gene features transfers to pathology-only inference
The query branch removes the gene encoder and the contrastive loss at inference; the paper relies on this transfer for the translator to work.
domain assumption Hypergraphs whose hyperedges connect each patch to its most feature-similar neighboring patches capture biologically relevant spatial and co-expression structure
No biological validation is provided; the neighbor branch uses this construction and its gains are the paper's main contribution.
domain assumption A truncated, pre-trained ResNet18 provides sufficiently rich features for gene expression prediction
The pathology encoder is not trained from scratch and is only fine-tuned indirectly through contrastive and MSE losses.
domain assumption Reported baseline numbers quoted from TRIPLEX and self-reproduced are directly comparable to NH2ST's numbers
The central comparison uses numbers from other papers or pipelines without evidence that splits, preprocessing, and gene selection are identical.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spatially Gene Expression Prediction using Dual-Scale Contrastive Learning." pith.science (2026). https://pith.science/paper/AGYCCZNZ

@misc{pith2026250623827,
  author       = {Pith},
  title        = {Pith review of: Spatially Gene Expression Prediction using Dual-Scale Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AGYCCZNZ}},
  note         = {Machine review of arXiv:2506.23827}
}

read the original abstract

Spatial transcriptomics (ST) provides crucial insights into tissue micro-environments, but is limited to its high cost and complexity. As an alternative, predicting gene expression from pathology whole slide images (WSI) is gaining increasing attention. However, existing methods typically rely on single patches or a single pathology modality, neglecting the complex spatial and molecular interactions between target and neighboring information (e.g., gene co-expression). This leads to a failure in establishing connections among adjacent regions and capturing intricate cross-modal relationships. To address these issues, we propose NH2ST, a framework that integrates spatial context and both pathology and gene modalities for gene expression prediction. Our model comprises a query branch and a neighbor branch to process paired target patch and gene data and their neighboring regions, where cross-attention and contrastive learning are employed to capture intrinsic associations and ensure alignments between pathology and gene expression. Extensive experiments on six datasets demonstrate that our model consistently outperforms existing methods, achieving over 20% in PCC metrics. Codes are available at https://github.com/MCPathology/NH2ST

Figures

Figures reproduced from arXiv: 2506.23827 by the authors.

**Figure 1.** The architecture of NH²2ST. Our model consists of two key components: a query branch and a neighbor branch. It employs dual-level contrastive losses to strengthen the intrinsic connections between the two modalities, facilitating the final gene expression prediction. modality (e.g., pathology or gene) to attend to relevant patterns in the other, establishing semantic correspondences and aligning histological structu… view at source ↗

**Figure 2.** Visualization of expression values on the PCW Dataset [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

[1]

Nature Reviews Genetics22(2), 71–88 (2021)

Armingol, E., Officer, A., Harismendy, O., et al.: Deciphering cell–cell interactions and communication from gene expression. Nature Reviews Genetics22(2), 71–88 (2021)

work page 2021
[2]

arXiv preprint arXiv:2406.06393 (2024)

Chen, J., Zhou, M., Wu, W., et al.: Stimage-1k4m: A histopathology image-gene expression dataset for spatial transcriptomics. arXiv preprint arXiv:2406.06393 (2024)

arXiv 2024
[3]

In: CVPR

Chung, Y., Ha, J.H., Im, K.C., et al.: Accurate spatial gene expression prediction by integrating multi-resolution features. In: CVPR. pp. 11591–11600 (2024)

work page 2024
[4]

Nature568(7751), 235–239 (2019)

Eng, C.H.L., Lawson, M., Zhu, Q., et al.: Transcriptome-scale super-resolved imag- ing in tissues by rna seqfish+. Nature568(7751), 235–239 (2019)

work page 2019
[5]

In: MICCAI

Fan, L., Sowmya, A., Meijering, E., Song, Y.: Learning visual features by coloriza- tion for slide-consistent survival prediction from whole slide images. In: MICCAI. pp. 592–601. Springer (2021)

work page 2021
[6]

IEEE Transactions on Medical Imaging42(5), 1401–1412 (2022)

Fan, L., Sowmya, A., Meijering, E., et al.: Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE Transactions on Medical Imaging42(5), 1401–1412 (2022)

work page 2022
[7]

In: MICCAI

Fan, L., Sowmya, A., Meijering, E., et al.: Fast ff-to-ffpe whole slide image trans- lation via laplacian pyramid and contrastive learning. In: MICCAI. pp. 409–419. Springer (2022)

work page 2022
[8]

IEEE Transactions on Pattern Analysis and Machine Intelligence44(5), 2548–2566 (2020)

Gao, Y., Zhang, Z., Lin, H., et al.: Hypergraph learning: Methods and practices. IEEE Transactions on Pattern Analysis and Machine Intelligence44(5), 2548–2566 (2020)

work page 2020

Show all 36 references

[9]

Nature biomedical engineering 4(8), 827–834 (2020)

He, B., Bergenstråhle, L., Stenbeck, L., et al.: Integrating spatial gene expression and breast tumour morphology via deep learning. Nature biomedical engineering 4(8), 827–834 (2020)

2020
[10]

In: CVPR

He, K., Zhang, X., Ren, S., et al.: Deep residual learning for image recognition. In: CVPR. pp. 770–778 (2016)

2016
[11]

NeurIPS37, 53798–53833 (2025)

Jaume, G., Doucet, P., Song, A., et al.: Hest-1k: A dataset for spatial transcrip- tomics and histology image analysis. NeurIPS37, 53798–53833 (2025)

2025
[12]

Cell182(2), 497–514 (2020)

Ji, A.L., Rubin, A.J., Thrane, K., et al.: Multimodal analysis of composition and spatial architecture in human squamous cell carcinoma. Cell182(2), 497–514 (2020)

2020
[13]

Pattern Recognition165, 111544 (2025)

Jing, W., Wang, J., Di, D., et al.: Multi-modal hypergraph contrastive learning for medical image segmentation. Pattern Recognition165, 111544 (2025)

2025
[14]

Bioengineering 10(8), 897 (2023)

Lee, M.: Recent advancements in deep learning using whole slide imaging for cancer prognosis. Bioengineering 10(8), 897 (2023)

2023
[15]

Neuropsychopharmacology45(1), 232 (2019)

Maynard, K., Jaffe, A., Martinowich, K.: Spatial transcriptomics: putting genome- wide expression on the map. Neuropsychopharmacology45(1), 232 (2019)

2019
[16]

Briefings in Bioinformatics 25(6), bbae551 (2024) 10 Qu M

Min, W., Shi, Z., Zhang, J., et al.: Multimodal contrastive learning for spatial gene expression prediction using histology images. Briefings in Bioinformatics 25(6), bbae551 (2024) 10 Qu M. et al

2024
[17]

arXiv preprint arXiv:1807.03748 (2018)

Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

2018 arXiv
[18]

BioRxiv pp

Pang, M., Su, K., Li, M.: Leveraging information in spatial transcriptomics to predict super-resolution gene expression from histology images in tumors. BioRxiv pp. 2021–11 (2021)

2021
[19]

Clinical Epigenetics14(1), 31 (2022)

Qi, L., Teschendorff, A.E.: Cell-type heterogeneity: Why we should adjust for it in epigenome and biomarker studies. Clinical Epigenetics14(1), 31 (2022)

2022
[20]

In: BIBM

Qu, M., Wu, Y., Di, D., et al.: Boundary-guided learning for gene expression pre- diction in spatial transcriptomics. In: BIBM. pp. 445–450. IEEE (2024)

2024
[21]

IJCAI (2025)

Qu,M.,Yang,G.,Di,D.,etal.:Multimodalcancersurvivalanalysisviahypergraph learning with cross-modality rebalance. IJCAI (2025)

2025
[22]

In: ICML

Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PMLR (2021)

2021
[23]

Nature 596(7871), 211–220 (2021)

Rao,A.,Barkley,D.,França,G.S.,etal.:Exploringtissuearchitectureusingspatial transcriptomics. Nature 596(7871), 211–220 (2021)

2021
[24]

Lab on a Chip22(24), 4774–4791 (2022)

Shen, X., Zhao, Y., Wang, Z., et al.: Recent advances in high-throughput single- cell transcriptomics and spatial transcriptomics. Lab on a Chip22(24), 4774–4791 (2022)

2022
[25]

PLoS computational biology11(11), e1004574 (2015)

Song, W.M., Zhang, B.: Multiscale embedded gene co-expression network analysis. PLoS computational biology11(11), e1004574 (2015)

2015
[26]

medRxiv pp

Srinivas, A.A., Jaroensri, R., Wulczyn, E., et al.: Estrogen receptor gene expression prediction from h&e whole slide images. medRxiv pp. 2024–04 (2024)

2024
[27]

Science353(6294), 78–82 (2016)

Ståhl, P.L., Salmén, F., Vickovic, S., et al.: Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science353(6294), 78–82 (2016)

2016
[28]

CVPR (2025)

Tang, Q., Fan, L., Pagnucco, M., Song, Y.: Prototype-based image prompting for weakly supervised histopathological image segmentation. CVPR (2025)

2025
[29]

Nature Reviews Cancer 16(5), 275–287 (2016)

Topalian, S.L., Taube, J.M., Anders, R.A., et al.: Mechanism-driven biomarkers to guide immune checkpoint blockade in cancer therapy. Nature Reviews Cancer 16(5), 275–287 (2016)

2016
[30]

arXiv preprint arXiv:2402.13723 (2024)

Vaessen, N., van Leeuwen, D.A.: The effect of batch size on contrastive self- supervised speech representation learning. arXiv preprint arXiv:2402.13723 (2024)

2024 arXiv
[31]

In: CVPR

Wei, X., Zhang, T., Li, Y., et al.: Multi-modality cross attention network for image and sentence matching. In: CVPR. pp. 10941–10950 (2020)

2020
[32]

NeurIPS36 (2024)

Xie, R., Pang, K., Chung, S., et al.: Spatially resolved gene expression prediction from histology images via bi-modal contrastive learning. NeurIPS36 (2024)

2024
[33]

In: WACV

Yang, Y., Hossain, M.Z., Stone, E.A., et al.: Exemplar guided deep neural network for spatial transcriptomics analysis of gene expression prediction. In: WACV. pp. 5039–5048 (2023)

2023
[34]

In: Topological, Algebraic and Geometric Learning Workshops 2022

Yi, K., Chen, J., Wang, Y.G., et al.: Approximate equivariance so (3) needlet convolution. In: Topological, Algebraic and Geometric Learning Workshops 2022. pp. 189–198. PMLR (2022)

2022
[35]

arXiv preprint arXiv:2205.01917 (2022)

Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)

2022 arXiv
[36]

Briefings in Bioinformatics 23(5), bbac297 (2022)

Zeng, Y., Wei, Z., Yu, W., et al.: Spatial transcriptomics prediction from histology jointly through transformer and graph neural networks. Briefings in Bioinformatics 23(5), bbac297 (2022)

2022

Pith tools

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

[1] [1]

Nature Reviews Genetics22(2), 71–88 (2021)

Armingol, E., Officer, A., Harismendy, O., et al.: Deciphering cell–cell interactions and communication from gene expression. Nature Reviews Genetics22(2), 71–88 (2021)

work page 2021

[2] [2]

arXiv preprint arXiv:2406.06393 (2024)

Chen, J., Zhou, M., Wu, W., et al.: Stimage-1k4m: A histopathology image-gene expression dataset for spatial transcriptomics. arXiv preprint arXiv:2406.06393 (2024)

arXiv 2024

[3] [3]

In: CVPR

Chung, Y., Ha, J.H., Im, K.C., et al.: Accurate spatial gene expression prediction by integrating multi-resolution features. In: CVPR. pp. 11591–11600 (2024)

work page 2024

[4] [4]

Nature568(7751), 235–239 (2019)

Eng, C.H.L., Lawson, M., Zhu, Q., et al.: Transcriptome-scale super-resolved imag- ing in tissues by rna seqfish+. Nature568(7751), 235–239 (2019)

work page 2019

[5] [5]

In: MICCAI

Fan, L., Sowmya, A., Meijering, E., Song, Y.: Learning visual features by coloriza- tion for slide-consistent survival prediction from whole slide images. In: MICCAI. pp. 592–601. Springer (2021)

work page 2021

[6] [6]

IEEE Transactions on Medical Imaging42(5), 1401–1412 (2022)

Fan, L., Sowmya, A., Meijering, E., et al.: Cancer survival prediction from whole slide images with self-supervised learning and slide consistency. IEEE Transactions on Medical Imaging42(5), 1401–1412 (2022)

work page 2022

[7] [7]

In: MICCAI

Fan, L., Sowmya, A., Meijering, E., et al.: Fast ff-to-ffpe whole slide image trans- lation via laplacian pyramid and contrastive learning. In: MICCAI. pp. 409–419. Springer (2022)

work page 2022

[8] [8]

IEEE Transactions on Pattern Analysis and Machine Intelligence44(5), 2548–2566 (2020)

Gao, Y., Zhang, Z., Lin, H., et al.: Hypergraph learning: Methods and practices. IEEE Transactions on Pattern Analysis and Machine Intelligence44(5), 2548–2566 (2020)

work page 2020