Pith. sign in

REVIEW 3 major objections 4 minor 17 references

Global Context-aware Representation Learning for Spatially Resolved Transcriptomics

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Spotscape: global pairwise similarity, not local attention, is the learning target for spatial transcriptomics.

desk verdict The Similarity Telescope as written enforces cross-view symmetry, not augmentation-invariant global consistency; the experiments are thorough and the multi-slice results are strong, but the central mechanism is overclaimed. read the letter →

arxiv 2506.15698 v2 pith:NI7H3ZCC submitted 2025-05-28 cs.LG cs.CV

Yunhak Oh , Junseok Lee , Yeongmin Kim , Sangwoo Seo , Namkyeong Lee , Chanyoung Park This is my paper
classification cs.LGcs.CV
keywords spatiallyresolvedtranscriptomicsself-supervisedrepresentationlearninggraphneuralnetworksglobalsimilarityconsistencyTelescopespatialdomainidentificationmulti-sliceintegrationbatcheffectcorrection
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

Spotscape is a self-supervised method for spatial transcriptomics built on the claim that local-neighborhood graph message passing cannot separate spots near domain boundaries, because gene expression varies smoothly across space and nearby spots look alike. The paper proposes to train a graph encoder so that the full matrix of pairwise cosine similarities between all spots is stable under random masking of genes and edges, a module it calls the Similarity Telescope, and adds a similarity-scaling loss that matches top-k similarity levels within and across slices to counter batch effects. The central claim, backed by experiments on DLPFC, MTG, mouse embryo, NSCLC, and breast-cancer data, is that this global similarity-geometry objective gives spot representations that outperform prior SRT methods for spatial domain identification, trajectory inference, imputation, multi-slice integration, and alignment. A sympathetic reader cares because, if correct, it offers a single self-supervised objective that improves downstream biological analyses without labels or bespoke alignment machinery.

What carries the argument

The carrying object is the Similarity Telescope, the loss $\mathcal{L}_{\mathrm{SC}} = \mathrm{MSE}\big(\tilde{Z}_{\mathrm{norm}}(\tilde{Z}'_{\mathrm{norm}})^\top,\, \tilde{Z}'_{\mathrm{norm}}(\tilde{Z}_{\mathrm{norm}})^\top\big)$, which forces the matrix of cosine similarities between every pair of spots to be identical under two randomly masked views of the same spatial graph. Around it, Spotscape stacks a prototypical contrastive loss $\mathcal{L}_{\mathrm{PCL}}$ that groups spots assigned to the same K-means prototype and a similarity-scaling loss $\mathcal{L}_{\mathrm{SS}}$ that matches the average top-$k$ similarity a spot has within its own slice to the average top-$k$ similarity it has in each other slice. These losses sit on top of a graph-autoencoder reconstruction loss; the paper's appendix shows that the consistency loss, unlike reconstruction alone, couples the gradient updates of different spots, which is the mechanism by which information flows globally.

What would settle it

Retrain Spotscape on a labeled slice with the reconstruction loss removed and monitor the pairwise cosine-similarity matrix during training: if the entries converge to a near-constant value across all spot pairs, or cluster only by slice identity, then the augmentation-invariance objective alone does not encode biological relatedness and the claimed mechanism fails. The batch-effect half could be tested separately by removing $\mathcal{L}_{\mathrm{SS}}$ on the MTG AD/control integration and directly measuring whether the top-$k$ inter-slice similarity means drift apart from the intra-slice means.

Watch

Extended reading notes

Core claim

The paper's central claim is that the right learning target for spatial transcriptomics is the global map of pairwise similarities between spots, not the local edge-weight structure of a spatial graph. Because gene expression varies continuously across tissue, neighboring spots carry nearly identical signals, and even perfectly learned attention cannot disambiguate spots at domain boundaries. Spotscape therefore optimizes a relation-consistency loss that keeps the full cosine-similarity matrix between all spot pairs invariant under two stochastic augmentations of the graph, adds prototype-based contrastive learning and a similarity-scaling loss for multi-slice data, and reports that this combination outperforms prior methods in all 16 single-slice evaluations across four datasets as well as in homogeneous and heterogeneous integration and alignment tasks. The reconstruction loss is retained as the mechanism that prevents the similarity matrix from collapsing to a degenerate solution.

Load-bearing premise

The load-bearing premise is that the pairwise cosine-similarity structure among all spots, forced to be invariant under random feature and edge masking, is a faithful, non-degenerate proxy for biological relatedness; nothing in the proof rules out the trivial solution in which all similarities become equal, so the reconstruction loss is doing unseen security work. A second assumption, stated as a heuristic without theoretical justification, is that equalizing the mean of top-$k$ intra-slice and inter-slice similarities is enough to correct batch effects.

Editorial extensions

If this is right

  • Spatial domain identification can be framed as clustering a globally consistent similarity geometry, so spots at domain borders are assigned by their relation to the whole tissue rather than by a noisy local neighborhood.
  • The decoder trained with the reconstruction loss doubles as an imputer, so denoising, marker-gene detection, and trajectory inference can be served by the same self-supervised model.
  • Multi-slice integration no longer requires bespoke optimal-transport alignment: matching the scale of top-$k$ similarities across slices is shown to mitigate batch effects in both homogeneous and heterogeneous settings.
  • The same representation space supports slice alignment, with the paper reporting higher label-transfer ARI on mouse embryo than SLAT, a method designed specifically for alignment.
  • The objective scales to large datasets, with reported training times on synthesized slices up to 100,000 spots and on a 1.79-million-spot collection, making it usable as data volumes grow.

Reading between the lines

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

  • The paper leaves implicit that the similarity-scaling loss is a general batch-effect recipe: any multi-batch embedding task in which same-batch top-$k$ similarity levels systematically exceed cross-batch ones could adopt the same scale-matching objective, not just transcriptomics.
  • A testable extension the paper does not pursue is making the mask rates or the top-$k$ value adaptive per spot; the paper fixes those hyperparameters, so it does not reveal how sensitive the similarity geometry is to boundary-localized corruption.
  • The layer-patterned pseudotime reported on DLPFC suggests the learned geometry encodes developmental ordering; testing the same representation on a single-cell atlas with known differentiation trajectories would show whether the global similarity map carries ordering beyond tissue layers, but that test is outside the paper.
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

3 major / 4 minor

Summary. The paper proposes Spotscape, a self-supervised graph neural network for spatially resolved transcriptomics. The method combines a graph autoencoder with a 'Similarity Telescope' relation-consistency loss that is intended to make global pairwise spot similarities invariant to graph augmentation, plus a prototypical contrastive learning loss and a similarity-scaling loss aimed at batch-effect correction for multi-slice integration. The authors evaluate on DLPFC, MTG, Mouse Embryo, NSCLC, and Breast Cancer data across spatial domain identification, trajectory inference, imputation, homogeneous and heterogeneous integration, and multi-slice alignment, reporting that Spotscape outperforms a wide range of baselines on most tasks.

Significance. If the claims hold, the paper makes a useful empirical contribution: it demonstrates that a global pairwise-similarity objective can add signal beyond local message passing in SRT representation learning, and it proposes a simple, computationally cheap similarity-scaling heuristic for multi-slice integration. The experimental effort is substantial: five datasets, multiple downstream tasks, ablations of every loss component, sensitivity analyses over all balancing parameters, statistical testing over 10 runs, an unsupervised hyperparameter-selection protocol, and a runtime study on a dataset with nearly 1.8 million spots. The code is released. However, two load-bearing points need attention before the contribution can be accepted as stated: the equation for the Similarity Telescope loss does not implement the augmentation-invariance described in the text, and the paper's headline claim of outperforming on 'all 16 slices' in every metric is contradicted by its own Table 1(d).

major comments (3)
  1. [Section 4.2, Eq. (1), Appendix L] As written, L_SC = MSE(Z_norm (Z'_norm)^T, Z'_norm (Z_norm)^T) = MSE(A, A^T), which enforces symmetry of the cross-view similarity matrix A, not consistency between the two augmented views' within-view similarity structures. The two are not equivalent: there exist normalized representations with L_SC = 0 whose within-view similarity matrices differ (e.g., take Z'_norm = Z_norm(I+B) with any symmetric B that changes row directions). Appendix L confirms this reading: the proof shows only that gradients of L_SC couple different spots, and the statement 'the consistency loss equals zero ideally' refers to A = A^T, not to augmentation invariance. The paper's claim that L_SC learns relationships 'invariant under augmentation' is therefore not supported by the presented equation. Please either change the objective to directly compare within-view similarity matrices (e.g., MSE(Z_norm Z_norm^T, Z'_norm Z'_norm^T)) or revise the conceptual claims and present L_SC as a symmetry regularizer whose empirical benefits are demonstrated by the ablation.
  2. [Section 5.1, Table 1(d)] The claim that 'Spotscape consistently outperforms in all 16 slices across four datasets in terms of ARI, NMI, and CA' is contradicted by Table 1(d): on the NSCLC dataset, SpaceFlow achieves CA 0.75 (0.02) while Spotscape achieves 0.74 (0.01), and this particular comparison is not marked as statistically significant. The claim should be corrected to an accurate statement (e.g., 'on 15 of 16 slices for CA and on all 16 for ARI and NMI'), and the discussion in Section 5.1 should be adjusted accordingly.
  3. [Section 4.4, Eq. (6)] The similarity-scaling loss operates on H = Z_norm (Z'_norm)^T, a cross-view similarity matrix, although the text and Figure 2 describe 'intra-slice' and 'inter-slice' similarity matrices. Because H is not guaranteed to be symmetric, 'top-k within slice c' and 'top-k within slice j' in Eq. (6) may not correspond to a well-defined pairwise similarity between spot i and the selected spots. Please clarify the exact quantity being matched and state what invariant L_SS is designed to enforce; the current heuristic is presented without analysis, and given that the ablation shows L_SS is important for the multi-slice results, a precise formulation would make the claim more robust.
minor comments (4)
  1. [Section 4.2] The sentence 'we aim to minimize the cosine similarity between the spot representations' appears to be a typo; the intended meaning is likely 'minimize the discrepancy between the two similarity matrices'.
  2. [Figure 2 and Section 4.4] The notation H is used both for the cross-view similarity matrix and for the 'Intra-slice Similarity Matrix' in the figure; please disambiguate these uses.
  3. [Appendix E.1] The appendix states that baseline hyperparameters were tuned over specified ranges but does not report the selected values per dataset; providing this table would strengthen the fairness claim.
  4. [Section 5.3] Figure 9 illustrates a single anchor spot from DLPFC; the text should acknowledge that this is an anecdotal visualization rather than a quantitative evaluation of learned similarity quality.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Spotscape's objectives are training losses evaluated on independent benchmarks, not fitted constants read back as predictions.

full rationale

The paper does not derive a downstream prediction from fitted constants. The Similarity Telescope loss (Eq. 1), reconstruction loss (Eq. 2), PCL loss (Eq. 4-5), and similarity scaling loss (Eq. 6-7) are training objectives, and the central empirical claims are supported by ablations and comparisons against external baselines. The PCL module cites the authors' own prior work (Lee et al., 2023), but it is also attributed to external methods (Li et al., 2020a; De Donno et al., 2023), so the self-citation is not load-bearing. Hyperparameters are selected using NMI on the first seed, which is a selection-on-labels concern rather than a circularity: the paper explicitly provides an unsupervised silhouette-based alternative in Appendix E.2, and the reported numbers are means over 10 seeds. I note a correctness issue outside the circularity definition: Eq. (1) as written is MSE of a matrix with its transpose, i.e., a symmetrization of the cross-view similarity matrix, and it would be zero for any symmetric cross-view matrix; this weakens the claim that the loss enforces augmentation-invariant global consistency, but it is a mis-specification of the objective, not a case where the result is equivalent to the input by construction.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

All of the model's losses and hyperparameters are listed. The method relies on the smoothness of gene expression across space, the stability of cosine similarities under augmentation, and the heuristic that equalizing top-k similarity scales across slices mitigates batch effects. No new physical entities are introduced. The free parameters are mostly fixed by hand (loss weights, temperature, top-k) or selected per dataset (learning rate, PCL cluster count), and the PCL cluster count is aligned with the number of annotated domains.

free parameters (8)
  • lambda_Recon = 0.1
    Hand-chosen loss weight, fixed across all datasets (Table 7).
  • lambda_SC = 1.0
    Hand-chosen weight for the similarity consistency loss (Table 7).
  • lambda_PCL = 0.01
    Hand-chosen PCL loss weight for multi-slice tasks (Table 7).
  • lambda_SS = 1.0
    Hand-chosen similarity scaling loss weight (Table 7).
  • learning_rate = varies per dataset, 1e-5 to 1e-3
    Selected per dataset by NMI on the first seed, or by silhouette score in the unsupervised protocol (Appendix E.1, E.2).
  • PCL cluster count K = number of annotated spatial domains per dataset, e.g., 5-7 for DLPFC
    K-means granularity in PCL is set to the ground-truth domain count, a mild label leak not disclosed as supervision (Appendix E.3, Figure 20).
  • top-k for similarity scaling = 5
    Fixed number of top similarities used in L_SS (Table 7).
  • tau = 0.75
    Temperature in the PCL loss, fixed (Table 7).
assumptions (4)
  • domain assumption Gene expression varies smoothly along spatial coordinates, so local neighborhoods contribute redundant information and global similarities are necessary for discriminative representations.
    Central motivation in Sections 1 and 4.2, supported by citations but not demonstrated within this paper.
  • ad hoc to paper Cosine similarity between spot pairs, enforced to be invariant under augmentation, is a meaningful biological distance and the MSE loss over the two similarity matrices does not collapse to a degenerate solution.
    The relation consistency loss is a bespoke objective; Appendix L shows gradient coupling but does not prove absence of trivial solutions.
  • ad hoc to paper Matching the mean of top-k intra-slice and inter-slice similarities corrects batch effects in multi-slice integration.
    The similarity scaling strategy is a heuristic introduced in Section 4.4 with no theoretical justification.
  • domain assumption The SNN graph constructed from spatial coordinates (radius/top-k) adequately encodes tissue geometry for message passing.
    Graph construction in Section 3 follows prior work (Dong & Zhang, 2022), assumed valid across technologies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Global Context-aware Representation Learning for Spatially Resolved Transcriptomics." pith.science (2026). https://pith.science/paper/NI7H3ZCC

@misc{pith2026250615698,
  author       = {Pith},
  title        = {Pith review of: Global Context-aware Representation Learning for Spatially Resolved Transcriptomics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI7H3ZCC}},
  note         = {Machine review of arXiv:2506.15698}
}
read the original abstract

Spatially Resolved Transcriptomics (SRT) is a cutting-edge technique that captures the spatial context of cells within tissues, enabling the study of complex biological networks. Recent graph-based methods leverage both gene expression and spatial information to identify relevant spatial domains. However, these approaches fall short in obtaining meaningful spot representations, especially for spots near spatial domain boundaries, as they heavily emphasize adjacent spots that have minimal feature differences from an anchor node. To address this, we propose Spotscape, a novel framework that introduces the Similarity Telescope module to capture global relationships between multiple spots. Additionally, we propose a similarity scaling strategy to regulate the distances between intra- and inter-slice spots, facilitating effective multi-slice integration. Extensive experiments demonstrate the superiority of Spotscape in various downstream tasks, including single-slice and multi-slice scenarios. Our code is available at the following link: https: //github.com/yunhak0/Spotscape.

Figures

Figures reproduced from arXiv: 2506.15698 by the authors.

Figure 1
Figure 1. (a) Feature similarity comparison from global and local perspectives. In global view, the similarity between the anchor (i.e., red dot) and other spots gradually changes with their spatial coordinates. In contrast, in the local view, neighboring spots exhibit minimal feature discrepancy compared to the anchor, irrespective of the true spatial domain. (b) Clustering performance comparison in terms of clustering accur… view at source ↗
Figure 2
Figure 2. Overall framework of Spotscape, which is trained with the SNN graph using (a) similarity telescope and reconstruction loss, while additionally utilizing (b) PCL and similarity scale matching loss in multi-slice SRT. resent the entire structure as G = (X, A). The adjacency matrix A ∈ R Ns×Ns is defined such that Aij = 1 if there is an edge connecting nodes i and j, and Aij = 0 otherwise. For simplicity, we do not def… view at source ↗
Figure 3
Figure 3. Trajectory con￾servation between pseu￾dotimes and Layers in DLPFC data. NR4A2 (Layer 6 marker gene) Gold Standard Layer_1 Layer_2 Layer_3 Layer_4 Layer_5 Layer_6 WM unknown NR4A2 (Layer 6 marker gene) 0 1 2 3 Expression Level 0.1 0.2 0.3 0.4 0.5 Expression Level [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (25 more)
Figure 4
Figure 4. Figure 4: Spatial ex￾pression of raw and Spotscape imputed data for marker genes of Layer 6 in DLPFC data. experimental results, Bold indicates the best performance, underlining denotes the second-best, and an asterisk (*) marks statistically significant improvements of Spotscap…
Figure 5
Figure 5. Figure 5: Alignment results of Mouse embryo data [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: UMAP of Raw, GraphST, STAligner, CAST, Spotscape (w/o LSS), Spotscape (w/o LPCL), Spotscape by slice, ground truth, and K-means clustering results. w/o 𝓛𝐑𝐞𝐜𝐨𝐧 w/o 𝓛𝐒𝐂 Spotscape w/o 𝓛𝐑𝐞𝐜𝐨𝐧 w/o 𝓛𝐒𝐂 w/o Spotscape 𝓛𝐏𝐂𝐋 w/o 𝓛𝐒𝐒 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Sensitivity analysis of loss balancing parameters. Raw SpaceFlow GraphST SEDR STAGATE SpaCAE Spotscape (w/o 𝓛𝐒𝐂) Spotscape Cosine Similarity with anchor -1 1 Layer 1 Layer 2 Layer 3 Layer 4 Layer 5 Layer 6 White Matter unknown Anchor Node [PITH_FULL_IMAGE:figures/full…
Figure 9
Figure 9. Figure 9: Similarity comparison based on anchor node 489 in Layer 5. pathology and showcasing Spotscape’s utility in uncover￾ing meaningful biological variation. The detailed results are provided in Appendix J due to space limitations. Multi-slice alignment. We conduct experimen…
Figure 11
Figure 11. Figure 11: Spatial coordinates of DLPFC dataset. Middle Temporal Gyrus (MTG). The MTG (middle temporal gyrus) dataset includes samples from both control and Alzheimer’s disease (AD) groups. The MTG is a brain region particularly vulnerable to early AD pathology. In the original …
Figure 12
Figure 12. Figure 12: Spatial coordinates of MTG dataset [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 14
Figure 14. Figure 14: Spatial coordinates of Mouse Development dataset. B. Baseline Methods In [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Comparison with self￾supervised learning. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Sensitivity analysis for reconstruction loss balancing parameter (λRecon) of single DLPFC. 0.001 0.01 0.1 1.0 10.0 0.0 0.2 0.4 0.6 0.8 Patient 1 - Slice 151673 0.001 0.01 0.1 1.0 10.0 0.0 0.2 0.4 0.6 0.8 Patient 1 - Slice 151674 0.001 0.01 0.1 1.0 10.0 0.0 0.2 0.4 0.6…
Figure 17
Figure 17. Figure 17: Sensitivity analysis for similarity telescope loss balancing parameter (λSC) of single DLPFC. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]
Figure 18
Figure 18. Figure 18: Sensitivity analysis for learning rate of single DLPFC. 0.001 0.01 0.1 1.0 10.0 0.0 0.2 0.4 0.6 0.8 PCL Loss Balancing Parameter ( PCL) 0.001 0.01 0.1 1.0 10.0 0.0 0.2 0.4 0.6 0.8 Simliarity Scaling Loss Balancing Parameter ( SS) 0.1 0.25 0.5 0.75 1.0 0.0 0.2 0.4 0.6 …
Figure 19
Figure 19. Figure 19: Sensitivity analysis for multi-slice parameters. K-2 K-1 K K+1 K+2 0.0 0.2 0.4 0.6 0.8 Patient 1 - Slice 151673 K-2 K-1 K K+1 K+2 0.0 0.2 0.4 0.6 0.8 Patient 1 - Slice 151674 K-2 K-1 K K+1 K+2 0.0 0.2 0.4 0.6 0.8 Patient 1 - Slice 151675 K-2 K-1 K K+1 K+2 0.0 0.2 0.4 …
Figure 20
Figure 20. Figure 20: Sensitivity analysis for number of cluster (K) of single DLPFC. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_20.png]
Figure 21
Figure 21. Figure 21: Unsupervised hyperparameter searching strategy using silhouette scores. Single DLPFC Slice 151673 Single DLPFC Slice 151674 Single DLPFC Slice 151675 Single DLPFC Slice 151676 Single DLPFC Slice 151507 Single DLPFC Slice 151508 Single DLPFC Slice 151509 Single DLPFC S…
Figure 22
Figure 22. Figure 22: Performance comparison between optimized and reported hyperparameters. E.3. Implementation Details Model architecture and training. The model employs a 2-layer GCN (Kipf & Welling, 2016) as the GNN-based encoder and a 2-layer MLP as the decoder, both utilizing batch n…
Figure 23
Figure 23. Figure 23: Trade-off between single-slice SDI performance and running time for LPCL GNN vs. MLP in Spatially Resolved Transcriptomics. In this section, we aim to validate the architecture of Spotscapeby investigating whether Graph Neural Networks (GNNs) are indeed more suitable …
Figure 24
Figure 24. Figure 24: The running time of Spotscape and baseline methods over the various number of slices on the large dataset (ST Omics DB - Dataset ID: STDS0000142) 20 [PITH_FULL_IMAGE:figures/full_fig_p020_24.png]
Figure 25
Figure 25. Figure 25: Trajectory conservation between pseudotimes and Layers in DLPFC data. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_25.png]
Figure 26
Figure 26. Figure 26: Trajectory inference results of Spotscape. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_26.png]
Figure 27
Figure 27. Figure 27: Imputation error comparison across various drop rates in the DLPFC. RORB (Layer 4 marker gene) ETV1 (Layer 5 marker gene) NTNG2 (Layer 6 marker gene) NR4A2 (Layer 6 marker gene) OLIG2 (WM marker gene) Gold Standard Layer_1 Layer_2 Layer_3 Layer_4 Layer_5 Layer_6 WM un…
Figure 28
Figure 28. Figure 28: Specifically, RORB serves as a canonical marker for layer 4 neurons (Clark et al., 2020); ETV1 is associated with layer 5 neurons (Goralski et al., 2024); NTNG2 and NR4A2 are well-recognized markers for layer 6 neurons (Maynard et al., 2021; Darbandi et al., 2018); an…
Figure 29
Figure 29. Figure 29: Top 25 biological process that DEGs between AD and Control enriched in a cluster assigned to layer 2. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_29.png]
Figure 30
Figure 30. Figure 30: Top 25 biological process that DEGs between AD and Control enriched in a cluster assigned to layer 5. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_30.png]
Figure 31
Figure 31. Figure 31: Alignment results of triple positive cells. We conduct cross-technology alignment between data obtained from Xenium and Visium. Since Xenium offers higher resolution than Visium, while Visium provides a more comprehensive transcriptome view, aligning Xenium with Visiu…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages

  1. [3]

    Kipf, T. N. and Welling, M. Semi-supervised classifica- tion with graph convolutional networks. arXiv preprint arXiv:1609.02907,

    arXiv

  2. [5]

    exhibit more pronounced synaptic and proteostatic perturbations in later stages (Romito-DiGiacomo et al., 2007). By delineating clusters that map onto these distinct laminar features,Spotscapedemonstrates utility in uncovering meaningful biological variation from spatial transcriptomics data and in corroborating the different influences of AD across corti...

    work page 2007

  3. [7]

    G., Azabou, M., Dyer, E

    Thakoor, S., Tallec, C., Azar, M. G., Azabou, M., Dyer, E. L., Munos, R., Veliˇckovi´c, P., and Valko, M. Large- scale representation learning on graphs via bootstrapping. arXiv preprint arXiv:2102.06514,

    arXiv

  4. [8]

    Graph attention networks.arXiv preprint arXiv:1710.10903,

    Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . Graph attention networks.arXiv preprint arXiv:1710.10903,

    arXiv

  5. [10]

    URL https:// www.jneurosci.org/content/27/32/8496

    1523/JNEUROSCI.1008-07.2007. URL https:// www.jneurosci.org/content/27/32/8496. Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck, W. M., Hao, Y ., Stoeckius, M., Smibert, P., and Satija, R. Comprehensive integration of single-cell data. cell, 177(7):1888–1902,

    work page 2007

  6. [11]

    Deep graph contrastive representation learning

    Zhu, Y ., Xu, Y ., Yu, F., Liu, Q., Wu, S., and Wang, L. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131,

    arXiv 2006

  7. [12]

    It was profiled by Stereo-seq technology, which allows spatial transcriptomics at the cellular level by integrating DNA nanoball-patterned arrays with in situ RNA capture

    Mouse Embryo.It is mouse whole embryo datasets by development stages. It was profiled by Stereo-seq technology, which allows spatial transcriptomics at the cellular level by integrating DNA nanoball-patterned arrays with in situ RNA capture. It offers a detailed spatiotemporal transcriptomic atlas (MOSTA) of mouse embryonic development (see Figure 14). No...

    work page 2020

  8. [13]

    BGRL (Thakoor et al., 2021; Grill et al.,

    is a instance-wise contrastive learning method that learns representations by pushing negative pairs apart and pulling positive pairs together. BGRL (Thakoor et al., 2021; Grill et al.,

    work page 2021

Show all 17 references
  1. [26]

    (a) DLPFC Patient 1 (b) DLPFC Patient 2 (c) DLPFC Patient 3 Figure 25.Trajectory conservation between pseudotimes and Layers in DLPFC data

    In these results,Spotscapedemonstrates effectiveness in the trajectory inference task, further validating its broad applicability. (a) DLPFC Patient 1 (b) DLPFC Patient 2 (c) DLPFC Patient 3 Figure 25.Trajectory conservation between pseudotimes and Layers in DLPFC data. 21 Glo...

    2024

  2. [28]

    The results show that after imputation usingSpotscape, marker genes are more distinctly expressed, demonstrating the practical applicability ofSpotscape

    Specifically, RORB serves as a canonical marker for layer 4 neurons (Clark et al., 2020); ETV1 is associated with layer 5 neurons (Goralski et al., 2024); NTNG2 and NR4A2 are well-recognized markers for layer 6 neurons (Maynard et al., 2021; Darbandi et al., 2018); and OLIG2 i...

    2020

  3. [2016]

    The encoder’s hidden dimensions are set to [Ng,256,64] , while the decoder’s dimensions are configured as [64,256, Ng]

    as the GNN-based encoder and a 2-layer MLP as the decoder, both utilizing batch normalization and ReLU activation functions. The encoder’s hidden dimensions are set to [Ng,256,64] , while the decoder’s dimensions are configured as [64,256, Ng]. The clustering process in PCL is...

    2022

  4. [2017]

    L., Li`o, P., Bengio, Y ., and Hjelm, R

    Veliˇckovi´c, P., Fedus, W., Hamilton, W. L., Li`o, P., Bengio, Y ., and Hjelm, R. D. Deep graph infomax.arXiv preprint arXiv:1809.10341,

    arXiv

  5. [2020]

    Datasets Table 5.Statistics for datasets used for experiments

    11 Global Context-aware Representation Learning for Spatially Resolved Transcriptomics A. Datasets Table 5.Statistics for datasets used for experiments. Dataset Species Tissue Technology Resolution Cells/Spots Genes # of Spatial Domains Reference DLPFC Human Brain (dorsolatera...

    2021

  6. [2021]

    W., Schulmann, A., Hara, E., Winnubst, J., Liu, C., Valakh, V ., Wang, L., Shields, B

    Phillips, J. W., Schulmann, A., Hara, E., Winnubst, J., Liu, C., Valakh, V ., Wang, L., Shields, B. C., Korff, W., Chan- drashekar, J., et al. A repeated molecular architecture across thalamic pathways. Nature neuroscience, 22(11): 1925–1935,

    1925

  7. [2023]

    doi: 10.1038/ s41392-023-01588-0

    ISSN 2059-3635. doi: 10.1038/ s41392-023-01588-0. URL https://doi.org/10. 1038/s41392-023-01588-0. Goel, P., Chakrabarti, S., Goel, K., Bhutani, K., Chopra, T., and Bali, S. Neuronal cell death mechanisms in alzheimer’s disease: An insight. Frontiers in Molecular Neuroscience, 15,

  8. [2024]

    Li, J., Zhou, P., Xiong, C., and Hoi, S. C. Prototypical con- trastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966, 2020a. Li, X., Wang, K., Lyu, Y ., Pan, H., Zhang, J., Stambolian, D., Susztak, K., Reilly, M. P., Hu, G., and Li, M. Deep learn- i...

    2005 arXiv

  9. [5099]

    URL https://www.frontiersin.org/journals/ molecular-neuroscience/articles/10

    doi: 10.3389/fnmol.2022.937133. URL https://www.frontiersin.org/journals/ molecular-neuroscience/articles/10. 3389/fnmol.2022.937133. Goralski, T. M., Meyerdirk, L., Breton, L., Brasseur, L., Kurgat, K., DeWeerd, D., Turner, L., Becker, K., Adams, M., Newhouse, D. J., et al. S...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.