acceptodds
Under review as a conference paper at ICLR 2027

Laminar: Querying Biology at Any Scale with Tissue Embedding Fields

Abstract

Biology is organized at nested spatial scales, from molecules to cells to multicellular neighborhoods, whose functions emerge from multi-scale organizations rather than any single cell. Imaging-based spatial transcriptomics (iST) captures all of these scales in a single experiment, measuring up to billions of individual transcripts, each with its gene identity and sub-micrometer position. To handle this volume, existing foundation models pool transcripts into predefined units such as cells, patches, or bins before learning, which fixes the representation to a single spatial scale and discards the molecular geometry within each unit. We introduce *tissue embedding fields*, a representation that assigns contextual embeddings on a shared unit hypersphere to every measured transcript. A region of any size or shape, from a single cell to a whole section, can then be represented by summing and normalizing the embeddings of the transcripts it contains, without re-encoding the tissue. The spatial unit of analysis is therefore chosen at query time rather than fixed during training. To compute such fields at billion-transcript scale, we introduce Laminar, a local-to-global encoder whose cost grows linearly with transcript count. A graph neural network exchanges short-range molecular information over local transcript graphs while a quadtree spatial Transformer propagates information to progressively larger tissue tiles. Each transcript then retrieves this multi-scale context through cross-attention, yielding its final embedding. We train Laminar using a joint-embedding predictive architecture, where a context encoder processes spatially- and gene-masked tissue, while a predictor matches multi-scale regional representations of the complete tissue. A single Laminar encoder pretrained on 632 public human iST sections, evaluated frozen on eight transfer tasks spanning cell, niche, and tissue scales, achieves the best scores on tissue-scale tasks while remaining competitive with fixed-scale foundation models at cell scale, with an overall best mean rank. Laminar shows that biological representations need not be tied to predefined spatial units; a unified representation enables reasoning across the hierarchy of tissue organization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Foundation Models, AI for Science, Spatial Data, Omics data, Biology, Biomedical data, JEPA, transfomers, geometric deep learning, spatial omics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UJNkOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies self-supervised representation learning for imaging-based spatial transcriptomics and proposes a local-to-global encoder whose cost grows linearly with transcript count. It combines a local transcript graph, a quadtree spatial Transformer, and a cross-attention readout back to each transcript. It is trained with a multi-scale JEPA objective under spatial and gene masking. It is pretrained on 632 public iST sections and evaluated frozen against existing spatial and single-cell foundation models on eight transfer tasks spanning cell, niche, and tissue scales.

Critical strengths
  1. The studied problem is meaningful, and the transcript-level perspective is novel. Unlike existing single-cell or spot-based foundation models, which pool transcripts into predefined units before learning, this work represents tissue directly at the level of individual transcripts.
  2. The proposed approach is interesting and intuitive. The tissue embedding field allows representations to be read out at multiple spatial resolutions from a single encoding pass. Such a design choice can potentially inspire future studies.
Critical weaknesses
  1. My main concern is that the performance gains are not significant, and Laminar falls well behind other methods on several tasks. On cell type, Laminar scores 0.596, ranking 12th of 14 methods and below the raw-count baseline and scGPT-spatial. Even on region composition, where Laminar ranks first (0.869), simply predicting the training-fold mean already reaches 0.876.
  2. I am not sure that the evaluated tasks and comparators fit this modeling paradigm, since the settings are not fully aligned. It is unclear whether this benchmark effectively reflects the model’s capability.
  3. The comparators differ substantially in pretraining data, platforms, and gene vocabularies, which limits the comparability of the results. Performance differences may therefore reflect differences in training data rather than in modeling choices.
Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer uJNk

We thank the reviewer for the constructive review and are glad the reviewer found the problem meaningful, the transcript-level perspective novel, and the tissue embedding field interesting and intuitive. We ran new experiments for each of the three concerns. Changes in the paper are in blue.

W1. "My main concern is that the performance gains are not significant, and Laminar falls well behind other methods on several tasks …"

Laminar is now in the top three on all seven tasks, with the best mean rank (1.86 against 5.43 for the next method) and the highest Elo (1674 against 1631) (Table 1, lines 379–397). Two changes brought this about. We added two pretraining losses that predict the relative abundance of each gene's transcripts in local regions and their surrounding tissue (lines 323–334). And for each task Laminar now sums the transcript embeddings of the region being scored, instead of averaging cell embeddings.

SubmissionRevision
Cell type.596 (12th of 14).701 (2nd of 15)
WHO grade.278.454 (1st)
Region composition.869.897 (1st of 15, now above the training-fold mean of .876)
Gene imputation.660.684 (3rd)

On significance: Elo is fitted to wins and losses on paired folds, and we added paired tests corrected for multiple testing (lines 466–467, Appendices J and P). Over the seven tasks, Laminar is significantly above 12 of the 13 comparators, and none is significantly above it (Table 18). At the multicellular scale it is significantly above scGPT-spatial (adjusted p < .01). On a given task its median is higher than another method's in 91 of 97 cases, and its Elo is the highest in 99.9% of 2,000 bootstrap refits (Fig. 11).

W2. "I am not sure that the evaluated tasks and comparators fit this modeling paradigm …"

We read this concern as follows. The submission evaluated Laminar like a cell model, by embedding each cell and averaging cells over larger regions, although the field is meant to be read at the region itself. We now evaluate it that way. Laminar sums the transcript embeddings of the region being scored and normalizes once, with the same frozen encoder (Eq. 3, lines 341–344). On both tissue-core tasks this scores higher than averaging cells: WHO grade .454 against .421 and methylation subclass .541 against .501 (lines 2360–2362). Read this way, Laminar is second at the cell scale and first at the multicellular and tissue-core scales, while seven of the eight models pretrained on cells or cell neighborhoods do best at their own scale (Fig. 3b, lines 460–464). And when Laminar is read exactly like the comparators, by averaging cells, it still has the best mean rank of the 15 methods (2.86 against 5.43, Appendix I).

W3. "The comparators differ substantially … Performance differences may therefore reflect differences in training data rather than in modeling choices."

We agree. Models that were pretrained independently differ in data, platforms and gene vocabularies, so comparing them cannot isolate the model. We therefore ran two sets of experiments in which all three are the same.

Amount of pretraining. We trained four more runs of Laminar and ranked their checkpoints among the comparators (Appendix N, Table 16, lines 2572–2580). From 10k steps on, every checkpoint is first.

Pretraining stepsPasses over the corpusRank of Laminar
10k, four runs.111st in all four
18k, 30k, 60k.20 to .671st
90k, reported model1.001st

At 10k steps Laminar has seen about a ninth of its corpus. A run on a random quarter of the sections (31M cells, about the size of scGPT-spatial's corpus) is also first at 10k steps (mean rank 3.86, Elo 1652). The five published models with the largest corpora (88M to 112M cells) all rank below scGPT-spatial, which was pretrained on 30M profiles (Table 11, lines 2582–2586). The lead does not come from a larger pretraining corpus.

Model. Each of the ten ablations removes or replaces one part of the full 95.1M-parameter encoder and is compared with a control that has the same corpus, platforms, gene vocabulary, seed and number of steps. All ten score lower than their control over the seven tasks combined, and each is significantly lower on at least one task (Fig. 4). We also replaced Laminar's neighborhood attention with PTv3 (Wu et al., 2024), OctFormer (Wang, 2023) or Erwin (Zhdanov et al., 2025), again with everything else the same (30k steps). All three rank above every published method by Elo, and Laminar's own encoder has the highest Elo (Fig. 5, Table 15).

Elo at 30k stepsLaminarPTv3OctFormerErwinHighest published
One fit of 18 methods16801650164816451630

We thank the reviewer for raising these points and would appreciate a reassessment in light of the new experiments.

Reviewer_d2qfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a tissue embedding field assigning a contextual unit vector to each measured transcript. Regional states are additive sums, with normalization deferred to query time. Laminar computes the field through a local molecular GNN, quadtree pooling with neighborhood attention, and transcript-level context readout. Training combines multi-scale JEPA prediction, a field-composition objective, and anti-collapse regularization. A 95.1M-parameter encoder is pretrained on 632 sections and evaluated frozen on eight tasks.

Critical strengths
  1. Deferring spatial support to query time is a useful representation contribution. It enables segmentation-independent encoding and subsequent region queries from one cache. The paper correctly distinguishes additive unnormalized states from normalized directions, including cancellation cases in C.2.
  2. The architecture and objective are designed coherently around that interface. C.7 provides a linear neural-interaction count for fixed architecture after graph construction, and the appendices describe inference, corpus accessions, checkpoint selection, and measured encoding cost.
  3. The evaluation includes diverse endpoints and substantial cohorts. Laminar leads the reported methods on pathology region (0.445) and methylation subclass (0.580). The paper is transparent that cell-type performance is weaker, some folds share donors, and region composition is mainly diagnostic. These qualifications improve the credibility of the contribution.
Critical weaknesses
  1. The main evaluation averages normalized cell vectors for larger regions (G.3), rather than directly testing the advertised transcript-sum readout. Comparing both on existing region/core tasks would better establish the native interface’s utility.
  2. For region composition, G.3 describes full-section encoding before ring readout. Please clarify whether the center is removed before encoding, since ring embeddings may otherwise contain target context. The reported 0.869 cosine also trails the training-mean and ring-copy controls (0.876/0.963).
  3. Transfer and architectural attribution remain uncertain: performance is mixed across tasks, some folds share donors, and ablations use a small model trained on one section. A donor-grouped sensitivity check and one targeted full-model ablation would strengthen the positive results.
Questions for authors
  1. How do native transcript-sum and cell-averaged readouts compare on existing region/core tasks?
  2. Is the center removed before encoding in region composition, and what does a masked-center control show?
  3. Do the main gains persist under donor-grouped evaluation and a targeted full-model ablation?
7 comments
AuthorsOpenReview ↗

Response to Reviewer d2qf

We thank the reviewer for the constructive review and are glad the reviewer found tissue embedding fields a useful representation contribution and the evaluation diverse. Following the reviews we made two changes. We added two pretraining losses that predict the relative abundance of each gene's transcripts in local regions and their surrounding tissue (lines 323–334), and we now train for one full pass over the corpus (Table 9). And for each task Laminar now sums the transcript embeddings of the region being scored and normalizes once (Eq. 3, lines 341–344). Averaging cell embeddings, as in the submission, is kept as a control (cell-pooled Laminar).

W1 / Q1. "How do native transcript-sum and cell-averaged readouts compare on existing region/core tasks?"

The transcript sum scores higher on all four region and core tasks, and it is now how Laminar is evaluated (Appendix I, lines 2360–2362).

Median score (rank among the 15 methods)Cell-pooled Laminar (averaged cells)Laminar (transcript sum)
Region composition.891 (1st).897 (1st)
Breast niche.706 (3rd).708 (3rd)
WHO grade.421 (3rd).454 (1st)
Methylation subclass.501 (6th).541 (1st)

Over all seven tasks Laminar has the best mean rank of the 15 methods in both cases (1.86 with the transcript sum and 2.86 with averaged cells, against 5.43 for the highest-ranked published model).

W2 / Q2. "Is the center removed before encoding in region composition, and what does a masked-center control show?"

Removing the center before encoding does not change the result. In the submission the whole section was encoded and the center was left out only when the ring was read. As the reviewer suggested, we ran a control that removes the target tile and a 15 µm margin before encoding. Over 20 paired folds the score changes by −.0010, and the ring embeddings have a median cosine of .9994 with the standard encoding (Appendix G.1, lines 2020–2022).

Laminar now scores .897 on this task, first of the 15 methods and above the training-fold mean (.876). Since copying the ring is a strong baseline here, we keep region composition as a diagnostic of spatial smoothness. Laminar is also first without it (mean rank 2.00 against 5.67, lines 467–469, 2407–2409).

W3 / Q3. "Do the main gains persist under donor-grouped evaluation and a targeted full-model ablation?"

Yes, both.

… performance is mixed across tasks …

Laminar is now in the top three on all seven tasks (lines 458–460). Cell type rose from .596 (12th) to .701 (2nd). In paired tests over the seven tasks, Laminar is significantly above 12 of the 13 comparators and none is above it (Table 18).

… some folds share donors … a donor-grouped sensitivity check …

For cell type and pathology region, every method is also scored on folds that hold out whole donors (Appendix L, lines 2511–2521). With these folds Laminar still has the best mean rank (2.86 against 4.86 for scGPT-spatial) and the highest Elo, and is still significantly above 12 of the 13 comparators. On pathology region it stays in the top three, and no method is significantly better. The two meningioma tasks already hold out donors: their folds hold out whole slides, and 263 of 265 donors have all their cores on one slide (Appendix G.5, lines 2149–2150).

… ablations use a small model trained on one section …

We repeated the ablations with the full 95.1M-parameter encoder and multi-section pretraining. There are ten, each trained for 10k steps (1,500 in the submission), compared with a control that has the same seed and number of steps, and scored on all seven tasks (two in the submission). All ten score lower than their control over the seven tasks combined, and each is significantly lower on at least one task (Fig. 4, lines 473–482, Appendix H). Five of them:

AblationTasks with a significant drop
Frozen random gene tablecell type −.046, pathology region −.028, region composition −.022
Flat ViTWHO grade −.180, cell type −.037
Graph only (no quadtree)pathology region −.022, region composition −.022
Tile tokensWHO grade −.116, methylation −.116, cell type −.109
No abundance headsWHO grade −.183, cell type −.067

We also tested whether the result depends on Laminar's encoder. We replaced its neighborhood attention with PTv3 (Wu et al., 2024), OctFormer (Wang, 2023) or Erwin (Zhdanov et al., 2025) and kept everything else the same (30k steps). All three rank above every published method by Elo, and Laminar's own encoder has the highest Elo (Fig. 5, Appendix M).

We thank the reviewer for these suggestions, which led to the masked-center control, the donor-held-out evaluation and the full-model ablations, and would appreciate a reassessment.

Thank you for the authors’ response. It has largely addressed my concerns. I maintain my score of 3.

Reviewer_d2qfOpenReview ↗

Another Small Question

Thanks for you response. I have another question on W2: you call Laminar (.897) first among the 15 methods, but ring-copy was .963. Under the new transcript-sum setup, is ring-copy still higher? And does “first of 15” exclude that control? If so, please state in the text or caption that region composition is only a smoothness diagnostic, so .897 isn’t read as the best score.

Reviewer_aKTgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Laminar learns an embedding for every transcript, so you can get a representation for any region by pooling over a cached field. The encoder has a local molecular graph part, a quadtree attention hierarchy, and a cross-attention readout. Training is multi-scale JEPA plus a field-composition loss and an anti-collapse term. The model is pretrained on 632 sections, then frozen and tested on eight tasks (cell identity, expression recovery, tissue regions, tumour classification).

Critical strengths
  • Encoding once and then pooling over whatever region you want is a nice idea, and it's practical at this data size. Splitting local and larger-scale context also makes sense.
  • The paper covers a decent range of tasks, keeps pretraining and eval cohorts separate, includes raw-count baselines, and uses the same probe throughout. The pathology region and methylation subclass results look promising.
Critical weaknesses
  • The benchmark doesn't really show an overall win. Laminar has three first-place results in Table 1, but region composition shouldn't count for much. Laminar gets 0.869, the training-fold mean gets 0.876, and just copying the surrounding ring gets 0.963 (p. 8). The authors admit this, but then the task can't be evidence for the representation. On other tasks it's often clearly worse: Cell type: 0.596, vs 0.698 for raw counts and 0.733 for scGPT-spatial WHO grade: 0.278, vs 0.450 for raw counts Malignancy: 0.253, vs 0.467 for raw counts So the problem isn't limited to cell-scale tasks, and best mean rank isn't enough here. Paired fold-level comparisons on the tasks where it does win would help.
  • The ablations don't support the architecture story. In Table 2, the graph-only model ties the full model on the pathology region (0.387 for both), so the quadtree doesn't seem to matter for that task. Shuffling gene identities gives 0.400, which is actually higher. That doesn't mean identities don't matter, but it doesn't show that keeping their spatial layout helps either. Removing the field-composition loss does hurt, which is good. All of this is also a 3.2M model trained on a single section, so it says little about the full pretrained model. A matched local-only model with proper multi-section pretraining would be much more convincing.
Questions for authors
  • What does the comparison look like without region composition? Do the pathology region and methylation gains hold fold by fold against the best baselines?
  • How do you explain graph-only tying the full model and gene-shuffled beating it? Does the local-only result still hold with multi-section pretraining?
  • The benchmark pools transcripts into cells and then averages cells for bigger regions. That isn't the direct regional pooling the paper is selling. Can you test that directly, especially for regions that cross cell boundaries?
3 comments
AuthorsOpenReview ↗

Response to Reviewer aKTg

We thank the reviewer for the careful reading. We ran every analysis the reviewer asked for and made two changes:

  1. Updated training. We added two losses that predict the relative abundance of each gene's transcripts in local regions and their surrounding tissue (lines 323–334).
  2. Task-scale pooling. Laminar now sums the transcript embeddings of the region being scored (a cell, a niche or a tissue core) and normalizes once (Eq. 3, lines 341–344). Averaging cells, as in the submission, is kept as a control.

W1. "The benchmark doesn't really show an overall win … Paired fold-level comparisons … would help."

We agree that mean rank alone is not enough and followed the reviewer's suggestion of paired fold-level comparisons. Laminar is now the only one of the 15 methods in the top three on every task, and it has the highest Elo (1674 against 1631 for scGPT-spatial, the highest-ranked published model). Elo is fitted to wins and losses on paired folds (Table 1, Fig. 3). In paired tests over the seven tasks, corrected for multiple testing, it is significantly above 12 of the 13 comparators and none is above it (Table 18, lines 466–467).

The tasks the reviewer flagged have all improved:

SubmissionRevisionFor comparison
Cell type.596.701 (2nd)raw counts .698
WHO grade.278.454 (1st)raw counts .450
Region composition.869.897 (1st)training-fold mean .876

Malignancy class also rose, from .253 to .468 (raw counts .467). It is no longer a separate task, since it simply regrouped the methylation subclasses into three families, two thirds of the cores benign, and so repeated that task with coarser, less balanced labels. With it as an eighth task Laminar keeps the best mean rank (2.38).

Q1. "… without region composition? Do the pathology region and methylation gains hold fold by fold against the best baselines?"

Without region composition, Laminar is still first (mean rank 2.00 against 5.67), and it is first in every subset of four or more tasks (lines 467–469, 2407–2409). Fold by fold, no method is significantly better than Laminar on pathology region or on methylation subclass (lines 2411–2418). On pathology region Laminar is significantly better than six comparators, including raw counts (+.045, adjusted p = .020). Only scVI, which is fitted on the evaluation sections, has a higher median there; fold by fold Laminar is .003 below it. On methylation subclass Laminar has the highest median.

W2 / Q2. "The ablations don't support the architecture story … How do you explain graph-only tying the full model and gene-shuffled beating it?"

Neither result holds in the full model, and the reviewer was right that a small model trained on one section says little about it. We repeated the ablations with the full 95.1M-parameter encoder and multi-section pretraining: 10k steps instead of 1,500, a control with the same seed and steps for each, and all seven tasks instead of two (lines 473–482, Fig. 4, Appendix H). The ten ablations cover the encoder, the input and the objective, and include the local-only model the reviewer asked for (graph only: the quadtree removed). All ten score lower than their control over the seven tasks combined, and each is significantly lower on at least one task.

Ablation, largest overall drop firstTasks with a significant drop
Frozen random gene tablecell type −.046, path. region −.028, region comp. −.022
Flat ViTWHO grade −.180, cell type −.037
No field-composition lossregion comp. −.010
No hierarchy readoutregion comp. −.019
No isotropy termregion comp. −.011
Graph onlypath. region −.022, region comp. −.022
Tile tokensWHO grade −.116, methylation −.116, cell type −.109
No abundance headsWHO grade −.183, cell type −.067
Genes shuffledWHO grade −.110, cell type −.041
Gene identity onlypath. region −.047

Graph only no longer ties the full model on pathology region: it is .022 lower over three seed pairs (16 of 20 folds), and lower in each pair. Training on shuffled genes no longer scores higher (−.007). When we shuffle gene positions inside 256 µm squares at test time, which keeps each square's gene counts, pathology region drops by .111 (19 of 20 folds, Appendix O), so the model uses where the genes are.

Q3. "… Can you test that [direct regional pooling] directly, especially for regions that cross cell boundaries?"

Yes. This is now how Laminar is evaluated, and it scores higher than averaging cells on both tissue-core tasks: WHO grade .454 against .421 and methylation subclass .541 against .501 (lines 2360–2362). For regions that cross cell boundaries, we also summed every transcript inside the region, including those outside any cell, and the scores change by at most .006 (lines 2397–2400).

We thank the reviewer for these requests and would appreciate a reassessment.

I thank the authors for their detailed response and the substantial effort invested in revising the manuscript. I particularly appreciate the additional fold-level statistical comparisons, the comprehensive full-scale ablations, and the direct regional pooling experiments. These additions address several of my original concerns and considerably strengthen the technical evaluation. I acknowledge these improvements and have accordingly revised my assessment of the paper's technical quality.

My remaining reservation primarily concerns the magnitude and practical significance of the demonstrated performance improvements.

While Laminar achieves impressive consistency across the seven benchmark tasks and significantly outperforms 12 of 13 comparators in the combined analysis, the comparison with scGPT-spatial, the strongest published baseline overall, is not statistically significant (adjusted p = 0.48). Moreover, the improvements on several individual tasks remain relatively modest.

For instance, on WHO-grade classification, Laminar achieves 0.454 compared with 0.450 for raw counts, while on methylation-subclass classification the scores are 0.541 and 0.540, respectively. These near-identical results raise questions about the practical benefit of the learned representation over considerably simpler approaches on these endpoints. At the cell level, scGPT-spatial also remains stronger for both cell-type classification (0.733 vs. 0.701) and gene imputation (0.721 vs. 0.684).

I acknowledge that Laminar achieves stronger improvements over several pretrained models, particularly at larger spatial scales, and that the revised experiments provide considerably better statistical support for these findings. Nevertheless, I am not fully convinced that the magnitude of the improvements on key biological endpoints is sufficiently substantial.

Regarding the proposed representation, I appreciate the flexibility of encoding tissue once and querying regions at different spatial scales. The direct regional pooling experiments demonstrate additional benefits, particularly for tissue-core classification. However, cell-pooled Laminar still achieves the best overall ranking, suggesting that much of the benchmark advantage is not specific to the proposed flexible regional readout. Given that subcellular molecular organization has also been explored in previous work, stronger evidence of distinctive biological capabilities would further support the significance of this contribution.

Overall, I appreciate the authors' responsiveness and recognize that the revised manuscript is substantially stronger. My remaining concern is not primarily about technical soundness, but whether the performance gains and demonstrated biological utility are sufficiently significant to justify acceptance. While the revision has improved my assessment of the empirical evidence, it has not changed my overall recommendation, which I therefore maintain.

AuthorsOpenReview ↗

2nd Response to Reviewer aKTg

We thank the reviewer for the follow-up and for acknowledging that the new analyses "considerably strengthen the technical evaluation".

"… the comparison with scGPT-spatial … is not statistically significant (adjusted p = 0.48)."

This p-value is from the test that combines all seven tasks. Two of these are cell-level tasks. scGPT-spatial is pretrained on cell-level and at least 45 of the 59 sections of these tasks, while Laminar's corpus contains none of the evaluation sections or tasks (Table 17). On all five tasks at the multicellular and tissue-core scales, Laminar scores higher than scGPT-spatial.

Task (rank of 15)LaminarscGPT-spatial
Pathology region.405 (2nd).369 (8th)
Region composition.897 (1st).824 (4th)
Breast niche.708 (3rd).696 (5th)
WHO grade.454 (1st).314 (6th)
Methylation subclass.541 (1st).431 (13th)

The mean rank of scGPT-spatial drops monotonically, from 1.00 on the cell-level tasks to 5.67 on the multicellular tasks and 9.50 on the tissue cores. Over these five tasks, Laminar is significantly better than scGPT-spatial (adjusted p < .001, test of Appendix P on these tasks, lines 466 to 467 for the multicellular scale). Laminar also keeps the best mean rank (2.00 against 5.67, lines 2407 to 2409).

"… cell-pooled Laminar still achieves the best overall ranking …"

Cell-pooled Laminar does rank first among the other 14 methods (Table 14). But compared with Laminar, it ranks lower:

Against the same 14 methodsMean rankEloWHO gradeMethylation subclass
Laminar1.861674.454 (1st).541 (1st)
Cell-pooled Laminar2.861647.421 (3rd).501 (6th)

Particularly, the gap start to increase on at the tissue core level.

"… the practical benefit of the learned representation over considerably simpler approaches …"

Raw counts do well on the two tumor tasks because the baseline is summed over a whole core. Tumor grade and methylation class are properties of the whole tumor, at the bulk count level. One scale below, on the multicellular tasks, raw counts fall to a mean rank of 9.50, against 2.00 for Laminar. Laminar scores higher than raw counts on all seven tasks, the two cell-level tasks included, and significantly so across them (adjusted p = .008, Table 18). The practical benefit is our field that ranks first at both scales, while raw counts profile only perform at the tissue core level.

"… distinctive biological capabilities … demonstrated biological utility …"

We thank the reviewer for this suggestion, which is new in our discussion, and we would gladly provide any concrete analysis the reviewer has in mind. In our work, one frozen encoder, pretrained without labels, is benchmarked with 14 other methods on seven unseen tasks at three biological scales, on 70 unseen evaluation sections. The methods include six foundation models pretrained on 30 to 112 million cells and are published in the leading biology journals (Nature 2026, Nature Methods 2025, Nature Genetics 2024). To our knowledge, the benchmark is among the largest of its kind in breadth and depth, with seven tasks at three scales, from micrometers to millimeters, 15 methods, 59 held-out sections from 13 organs, and 300 tissue cores from an in-house cohort of 265 patients, labeled by pathologists and neuropathologists. The follow-up asks for "key biological endpoints". Deeper biological case studies are possible and we gladly respond to concrete requests, yet ICLR is not a biological journal. The reviewer's requests made us go deeper, and the paper is better for it.

The follow-up also mentions earlier subcellular work without naming a study. To our knowledge, Laminar is the first model that carries subcellular resolution to larger scales in one representation, the tissue embedding field.

Furthermore, we did not want to make this paper about competition with other methods, although its results would allow that. We wrote it as a scientific contribution, with a novel concept, a new multi-scale architecture, and a comprehensive evaluation. Our aim is stable performance of one frozen encoder across tasks and scales, which is what we believe a foundation model for tissue should deliver. Of the 15 methods, Laminar is the only one among the top three on every task, while every other method places seventh or lower at one of the three scales (lines 463 to 464).

Finally, we respectfully come back to the rating. The follow-up calls the revised manuscript "substantially stronger" and the remaining concern "not primarily about technical soundness". We carried out every analysis the original review asked for, and the rating is 1 (clear rejection), as it was before the revision. With respect, we cannot reconcile this rating with the reviewer's own assessment and kindly ask the reviewer to reassess the rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers: main changes and new results

We thank the reviewers for their careful assessment and constructive suggestions. The reviewers valued the idea of tissue embedding fields, which give every transcript a contextual embedding that can be pooled over a region of any size after a single encoding, but noted that our evaluation aggregated only at the cell scale and questioned whether the method yields measurable gains over existing models. We agree that pooling each region directly from the field had to be tested. The revision makes two main changes:

  1. Updated training. Training now combines multi-scale JEPA with predicting the relative abundance of each gene's transcripts in local regions and their surrounding tissue (Section 2.3).
  2. Task-scale pooling. Laminar now represents the unit of each task, from a single cell to a tissue core, by summing its transcript embeddings and normalizing once (Eq. 3). The submission's averaging of cell embeddings is kept as a control, cell-pooled Laminar.

The main results are summarized below. Details are in the replies to each reviewer, and changes in the revised manuscript are highlighted in blue.

  • Frozen transfer benchmark (aKTg\color{red}{\textsf{aKTg}}, d2qf\color{orange}{\textsf{d2qf}}, uJNk\color{green}{\textsf{uJNk}}): Laminar is now in the top three on all seven tasks, the only one of the 15 compared methods to do so, and first by mean rank (1.86 against 5.43 for scGPT-spatial, the highest-ranked published model) and by Elo (1674 against 1631), a Bradley–Terry rating estimated from paired fold-level comparisons. On a given task its median is higher than another method's in 91 of 97 cases, and in paired fold-level tests it is significantly above 12 of the 13 comparators, with none above it.
  • Full-model ablations (aKTg\color{red}{\textsf{aKTg}}, d2qf\color{orange}{\textsf{d2qf}}): The submission ablated a 3.2M-parameter model trained on a single section for 1,500 steps and reported two tasks. We now ablate the full 95.1M-parameter Laminar encoder with multi-section pretraining: ten ablations spanning the encoder, the input and the training objective, each trained for 10k steps against a matched control with the same seed and number of steps, and each scored on all seven tasks. All ten lower the combined score, and each significantly lowers at least one task.
  • Tissue embedding fields with other architectures: We replaced Laminar's neighborhood attention with three other spatial Transformers (PTv3, OctFormer and Erwin), keeping the objectives, training data and number of steps fixed. Every variant ranks above every published method by Elo, so the performance of tissue embedding fields extends beyond Laminar's encoder, which has the highest Elo of the four.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.