acceptodds
Under review as a conference paper at ICLR 2027

Sudoku: Understanding Spatial Organization of Tissue by Learning to Generate

Abstract

Tissue function emerges from the spatial organization of cells across multiple scales. Local interactions shape the state of each cell, and their collective effects give rise to multicellular structures. Spatial transcriptomics (ST) now profiles millions of cells at known coordinates, making it possible to learn this organization from data. Most existing pre-training methods, however, focus on masked reconstruction at fixed masking rates to learn representations. Others generate cells autoregressively along an order that tissue does not have. Here we present Sudoku, a spatial foundation model that learns tissue organization through spatial discrete denoising. Sudoku represents cell states with a hierarchical vocabulary, and learns the distribution of hidden cell states given spatial context across a full masking range. This conditional formulation captures dependencies from local cellular interactions to coherent multicellular configurations without imposing a predefined order, and naturally unifies representation and generation as two inference regimes under full and partial observation. Pre-trained on a rigorously quality-controlled corpus of 49 million cells from 12 organs across human and mouse, Sudoku outperforms existing pre-trained models and specialized methods on niche retrieval, multi-scale clustering, spatial generation and immune-recruitment prediction, taking a step toward a virtual tissue model built from ST data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, computational biology, single-cell foundation models, generative pre-training, representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8X6EOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a spatial transcriptomics foundation model using discrete diffusion to learn cellular representations and inpaint spatial cellular configurations.

Critical strengths

The paper addresses biologically relevant tasks, including niche retrieval, spatial clustering, and immune recruitment prediction.

Critical weaknesses
  1. Insufficient motivation for discrete diffusion. The authors do not convincingly explain why discrete diffusion is necessary for these tasks, particularly niche retrieval and clustering, which are fundamentally representation-learning problems. The advantages over conventional masked modeling with variable masking rates remain unclear.
  2. Unclear practical advantages of generation. Although the downstream tasks are biologically meaningful, the paper does not demonstrate why generative modeling is preferable to simpler discriminative approaches. For example, immune recruitment could be predicted directly without generating surrounding cells. Furthermore, differences in baseline adaptation protocols complicate fair comparisons.
  3. SUDOKU can be considered as a pre-trained spatial transcriptomics model, but claiming it as a general-purpose tissue foundation model is not convincing.
Questions for authors

Key question on the motivation for discrete diffusion: What specific advantages does discrete diffusion offer for spatial transcriptomics compared with conventional masked modeling? In particular, why is a generative formulation necessary for downstream tasks such as niche retrieval and clustering?

10 comments
AuthorsOpenReview ↗

Response to Reviewer 8X6E (1/3)

Thank you for your review. We believe there are some misunderstandings about our motivation, the practical value of generation and the scope of our foundation model term. We apologize for not stating these points more clearly and clarify them below. We sincerely look forward to further discussion.

Weakness 1: motivation for discrete diffusion

We first clarify the motivation for discrete diffusion. Since no specific method is named in the review, and the masked modeling baselines (CellPLM, Nicheformer, scGPT-spatial, SToFM, CellWorld and TERRA) all use one or a few fixed masking rates (Table 1), we answer for both forms that masked modeling with variable masking rates can take:

  • Predicting discrete tokens. Masked prediction of a distribution over discrete tokens at variable masking rates is masked discrete diffusion itself (Section 2). This is the formulation SUDOKU adapts to spatial transcriptomics, where no existing model uses it (Table 1).
  • Regressing continuous embeddings. Most existing spatial models regress the expression or embedding of masked genes or cells (Table 1), so their output is a point estimate. Table 5 contains exactly this form: a latent regression control trained at the same variable masking rates as SUDOKU, with backbone, corpus and schedule also fixed. It is worse than SUDOKU on both representation and generation (excerpt below).

Our motivation for the generative formulation over the second form is to unify generation and representation in one model, which gives the following advantages:

  1. Generative model: a distribution can be sampled and a point estimate cannot. Trained over the full range of masking rates, our model fills in missing regions from any amount of visible context (Sections 3.3 and 4.2), which no conventional masked spatial model can do (Table 1). We clarify the practical value of this generative feature in our reply to Weakness 2(a).
  2. Generative pre-training facilitates understanding, as shown in vision (iGPT, MAGE, Vision Banana). The reason is that regression is optimized by the average cell state given the context, so the visible embeddings only need to tell apart contexts with different averages. The generative formulation is optimized by the full distribution, so they must tell apart contexts that admit different cell states, even when the averages coincide. The distribution determines the average but not the reverse: generation asks the embeddings to encode strictly more about the context. Table 5 tests this directly: with masking, backbone, corpus and schedule fixed, replacing the generative formulation by regression lowers both niche retrieval and generation (excerpt below).
Pre-training objective (Table 5)Niche retrieval (HESTA AUPRC)Generation (type accuracy)
SUDOKU0.5030.405
Conventional masked modeling0.4290.368

We would like to note that we did not claim the generative training objective to be a must for representation learning. We do not deny that conventional masked modeling is useful for representation learning, but our experiments have shown that generative pretraining yields better representations.

AuthorsOpenReview ↗

Response to Reviewer 8X6E (2/3)

Weakness 2(a): practical value of generation

The purpose of generation is not limited to predicting immune recruitment. It is to recover the cell states of a missing region and their spatial arrangement, which SUDOKU does more accurately than every baseline across organs, species and missing geometries (Section 4.2, Table 4); immune recruitment (Section 4.3) is one way to validate the generated tissue against measurement. A simple discriminative model outputs one quantity chosen before training and must be retrained for each new quantity, whereas any quantity of the region can be read from the generated cells afterwards. Even on immune recruitment, the one quantity it is trained for, the direct discriminative approach is clearly worse than SUDOKU. To show this, we add a direct prediction baseline: a ridge regression model from the mean expression of a clone to its immune recruitment, fitted on the clones of the training sections.

MethodSpearman, mean of three sectionsMedian rank error
SUDOKU0.770.12
Direct prediction0.620.24

Beyond recruitment, the same generated band shows where the immune cells are and which cells sit next to the clone, as Figure 2c illustrates for one clone. In the revision we score three such quantities on all 125 clones, from the same generated cells as Figure 2 and with no retraining.

Question about the band around a cloneAgreement with measurementSUDOKUGeSTInterpolation
Which side holds more immune cellsPearson over four sectors0.610.520.22
Which cell types lie within 25 µm of the cloneSpearman over 12 cell types0.810.460.32
Which cell types sit next to each otherPearson over 67 type pairs, against shuffling0.740.390.29

SUDOKU agrees best with the measurement on all three questions. A simple discriminative approach cannot answer these questions: the model built for recruitment returns only recruitment, and each row would need its own target and its own model, whereas SUDOKU answers all three from the generation it has already made.

Generation also enables uses that a predictor cannot serve. In future work, we will alter the cells of one region by hand and generate the tissue around them, which gives an in silico spatial perturbation. We see this as a step toward the virtual tissue model.

Weakness 2(b): baseline adaptation protocols

Every method is adapted in the same way wherever its released code allows. SUDOKU is fine-tuned on the benchmark sections with its own self-supervised objective and no labels, and no benchmark section is in its pre-training corpus (Appendix E). GeST, CellPLM, Nicheformer and Novae, the baselines whose training code is released, are fine-tuned on the same sections in the same way (Appendix F); GeST is also pre-trained on our corpus, so that comparison differs only in the model. Four baselines are used as released: scGPT-spatial, SToFM and CellWorld release inference code only, and the self-supervised adaptation of TERRA exceeds our compute budget. For a fair comparison with these four, we added zero-shot evaluation with SUDOKU. It is far above the best of the four on every representation benchmark, and on HESTA and multi-scale clustering it already outperforms every baseline, including the fine-tuned ones (0.486 vs 0.447 for GeST on HESTA, 0.408 vs 0.295 for STAGATE on clustering).

TaskMetricSUDOKU, zero-shotBest of the four, as released
Niche retrieval, HESTAAUPRC, mean of 8 organs0.4860.303 (scGPT-spatial)
Niche retrieval, DLPFCPCC, mean of 7 classes0.4900.223 (TERRA)
Multi-scale clusteringARI, mean of 3 levels0.4080.255 (scGPT-spatial)

We add this table to Appendix J. The revision also adds a table to Appendix F (Table A.8) that lists, for every method, its starting weights and its adaptation; all models are read out at the same neighborhood radius as SUDOKU.

AuthorsOpenReview ↗

Response to Reviewer 8X6E (3/3)

Weakness 3: foundation model

We present SUDOKU as a spatial foundation model, that is, a foundation model for spatial transcriptomics. It is not meant to cover every aspect of tissue, for example other modalities such as histology images. Within this scope, SUDOKU has the two properties that define a foundation model (Bommasani et al., 2021): pre-training on broad data, and adaptation to a wide range of downstream tasks.

  1. Broad data. SUDOKU is pre-trained on 396 whole-transcriptome sections at single-cell resolution, from 12 organs, two species, and normal and tumor tissue. We favored quality over quantity: the 49.1 million cells are what remains of about 80 million after strict quality control (Appendix A). This corpus is already larger than those of Novae (Nature Methods, 2025) and scGPT-spatial, two spatial foundation models pre-trained on about 30 million cells or spots.
  2. A wide range of downstream tasks. SUDOKU serves both paradigms of downstream use including representation and generation. It is the only model that both generates cells and gives representations at multiple spatial scales (Table 1). The experiments span four spatial transcriptomics technologies, two species and ten organs, from embryonic heart and liver to cortex, kidney and lung tumor tissue (Appendix E), and SUDOKU outperforms existing methods in all of these evaluation scenarios.

We hope these clarifications and results answer your concerns. If any concern remains, please let us know, and we will do our best to address it within the discussion period.

Reviewer_b5GTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SUDOKU, a pre-trained model for single-cell resolution spatial transcriptomics. Each cell's whole-transcriptome profile is encoded into a two-level residual-quantized code, and a bidirectional transformer with RoPE encoding is trained to recover the codes of masked cells within a neighborhood of certain radius, with a masking rate drawn from a uniform distribution, following masked discrete diffusion. The same model can give cell and niche representations, or generates missing cells by confidence-ordered parallel unmasking when a neighborhood is partially observed.

Critical strengths
  • Learning tissue organization as a conditional distribution over cell states is an important problem. I particularly like that cell coordinates enter the attention directly through a relative positional encoding, so that both the representation and the generation are aware of the spatial arrangement of cells.
  • The results are encouraging and many baselines methods are included in the comparison including some of the very recent models like TERRA
Critical weaknesses

I have several concerns about whether the main claims are sufficiently supported.

  1. Discrete denoising versus regression. A main selling point of the paper is that SUDOKU predicts a distribution over discrete codes rather than regressing a latent (e.g. Table 1). I can see the merit of this for sampling, but it needs more support.
  • (a) How diverse are the learnt level-1 and level-2 codes? The paper does not report codebook usage, so it is unclear whether the codes are well spread or partly collapsed. A reasonable check would be to report how many of the 256 level-1 codes, the 256 level-2 codes and the 65,536 pairs are actually used, how different are they comparing to each other measured by similarities (cosine, Euclidean, …) to make sure that they don’t collapse and are actually properly used.
  • (b) Table 5 compares latent regression with the generative objective, but the latent-regression variant is not explained anywhere in the paper: which latent is regressed, with which loss, and how generation/sampling is done there. Could a regression with a different objective and proper tuning, or a continuous distributional head on z (like NB used in the paper) close the gap? This is not enough to support that discrete denoising here is for sure a better choice compared to regression.
  • (c) The two-level vocabulary covers 256 x 256 combinations, but generated cells are still restricted to the prototypes learnt in pre-training. Since the model is presented as a foundation model, how would it handle cell states that are absent from the pre-training corpus, for example from new tissues or diseases? Relatedly, are the tokenizer and codebooks updated during fine-tuning on the benchmark sections?
  1. Independent prediction of the two code levels. I have a major concern about one design choice. The two code levels are predicted by two separate heads and are independent given the context (lines 271 to 273). Although the training targets come from the tokenizer, where c2 is computed from the residual z - e1[c1], the two heads are trained with separate cross-entropies and the second head never receives c1. The predicted distribution is therefore factorized both in training and in generation, and c2 is drawn without any knowledge of c1, even though it naturally depends on the coarse code. What is the motivation for this choice, and why are the generated results still reasonable under it? Proper motivation and justification is needed here.

  2. Niche retrieval task (Table 2). The AUPRC values are hard to compare across organs, because the chance level of AUPRC equals the share of positive candidates, which depends on the size of each organ on the target section. Please report the chance level per organ, or a normalized (calibrated) score. It is also unclear how the scale is chosen: the NOI size is fixed without justification, and the readout radius used for retrieval is not reported. One interesting follow up is, If SUDOKU captures spatial dependencies at different scales, I would expect its advantage to grow for larger niches, where the spatial arrangement matters more. It would be nice to see if that is the case by repeating this experiment with larger and larger niche windows. Also, how robust are these numbers? Several margins are small (for example 0.765 vs 0.762 on L5L6), and there are no repeated fine-tuning seeds or confidence intervals. These are necessary to draw a convincing conclusion.

  3. Spatial generation task (Table 4). It is nice to include the 1-NN and 5-NN baselines, and not surprising that they are already strong compared with the trained models. However, the scoring is not the same across methods: k-NN uses a hard majority vote, while SUDOKU and GeST are scored by the expected probability of the true type. Why is this the case? How were k = 1 and k = 5 chosen? SUDOKU sees all visible cells of the 500-cell window, whereas k-NN only uses the k nearest ones, which can be fairly noisy. For a fair comparison, please evaluate, on the same windows and visible cells and with the same expected-accuracy score: (i) k-NN over a wider range of k, (ii) distance-weighted k-NN (can try, but I think the performance would be close to the (iii) following), and (iii) a composition baseline that assigns every masked cell the type composition of all visible cells in the window. I wonder whether these would match SUDOKU, especially in the more mixed windows. On robustness, again, single values without error bars are weak support.

  4. What does the positional encoding contribute? I do like the idea of building coordinates directly into the model, but it is not clear how the RoPE encoding benefits it. The only related analysis is the rotation test in Appendix, which checks robustness rather than usefulness. An ablation in Table 5 without RoPE, or without any spatial encoding (treating the neighborhood as a set, as in a deep set), would directly test the claim that the model is spatially aware. One can imagine that, without RoPE, all masked cells therefore receive the same input and hence the same predicted distribution at each decoding step, so this variant would show how much of the generation accuracy comes from knowing where a cell is, rather than only knowing the composition of the window. Without this ablation it is hard to draw a conclusion that the model is spatially aware.

  5. Fine-tuning protocol and zero-shot performance. SUDOKU is fine-tuned with its self-supervised objective on the benchmark sections before evaluation. However, the fine-tuning is not described: the number of steps, the learning rate, which parameters are updated (including the tokenizer and codebooks), the masking schedule and the computing cost is not reported. Please report these details. More importantly, since the model is presented as a spatial foundation model, please report zero-shot results as well for Tables 2 to 4, using the pre-trained weights without any fine-tuning. The Discussion acknowledges that new sections "still require self-supervised fine-tuning", but without zero-shot numbers it is hard to tell how much of the performance comes from pre-training itself and also make the “foundation model” claim weak.

  6. Finally, paper is poorly written. Several notation symbols are introduced without definition. There is a lack of explanation and intuition for different design choices.

Minor: Fig. 2c: "Surrounding band" reads as an independent legend item without a marker for the first glimpse. I suggest placing it as a heading above the Immune, Tumor and Other markers, so that it is clear these colors refer to the band cells. The caption also says the clone is shown in grey, while it appears pink in the figure. That’s a typo. Fig. 2b: the dots next to the boxes are not explained in the caption.

Questions for authors

See weaknesses above.

10 comments
AuthorsOpenReview ↗

Response to Reviewer b5GT (1/5)

Thank you for this careful review, which we found very helpful for improving the soundness and readability of our paper. We have added all the experiments you mentioned and we believe we have adequately addressed your concerns. In brief, we summarize your concerns in the following seven aspects and address each of them:

  1. Discrete codes versus regression. Codebooks are not collapsed. Both regression baselines, including a variant with a distributional head, fall below the distribution over codes.
  2. Independent prediction of the two code levels. Conditioning the second level on the first gives only a small gain, so we keep the simpler design.
  3. Niche retrieval. SUDOKU remains the best after normalizing AUPRC by the chance level and at every window size from 250 to 2000 cells.
  4. Spatial generation. All methods are scored in the same way, and SUDOKU outperforms every k-NN variant and the composition baseline.
  5. Positional encoding. Removing it greatly lowers both retrieval and generation.
  6. Fine-tuning and zero-shot results. Without fine-tuning, SUDOKU already achieves remarkable performance compared with zero-shot baselines.
  7. Writing. We have added a notation table, the intuition behind each design choice and the missing definitions.

We answer each point in detail below.

Weakness 1(a): codebook usage

We inspected the utilization and diversity of the codebook and confirmed that it is not collapsed. Specifically, we calculated the codebook utilization of all 49.1 million cells in the pre-training corpus. Every code at both levels is in use, the least used one by 2075 cells at level 1 and 2263 cells at level 2. Figure A.1 in Appendix B plots the number of cells on each code. Of the 65536 two-level code pairs, 39718 are used.

UsedTotalUtilization
Level-1 codes256256100.0%
Level-2 codes256256100.0%
Code pairs397186553660.6%

We then evaluated the code diversity in the codebook by calculating the code pair cosine similarity. The results show that no two codes are near duplicates. The most similar pair has a cosine similarity of 0.794 at level 1 and 0.809 at level 2. The median pairwise cosine similarity is close to zero at both levels, so the codes point in diverse directions instead of crowding together.

Among the 32640 pairs of codes in a codebookLevel 1Level 2
Median cosine similarity0.0190.003
Largest cosine similarity0.7940.809

We have added the figure and both tables to Appendix B.

Weakness 1(b): latent regression and continuous heads

The latent-regression variant is not explained anywhere in the paper: which latent is regressed, with which loss, and how generation/sampling is done there.

Here we clarify the settings of the latent-regression variant, which has been added to Appendix G:

  • Training recipe: compared to the standard SUDOKU model, the latent-regression baseline only replaces the two classification heads with one linear head that predicts, for each masked cell, its continuous, 128-dimensional tokenizer latent zz. The loss of latent regression is MSE weighted by 1/t, where t is the masking rate drawn for each neighborhood, as in the loss of SUDOKU (Section 3.2).
  • Generation: since this baseline is not sampleable, for generation it predicts the latents of all masked cells in one forward pass, quantizes each prediction into a code pair, and reads cell types through the same code-to-type table as SUDOKU.

Could a regression with a different objective and proper tuning, or a continuous distributional head on z (like NB used in the paper) close the gap?

We added a baseline with a Gaussian distributional head on zz, which outputs a mean and a per-dimension variance for each masked cell. Its loss is the Gaussian negative log-likelihood, weighted by 1/t in the same way. For generation, we quantize its predicted mean in the same way as the latent-regression baseline. Both continuous heads perform worse than SUDOKU on both tasks.

Pre-training objectiveHESTA Niche retrieval (mean AUPRC)Generation (mean cell-type accuracy)
SUDOKU0.5030.405
Latent regression (MSE)0.4290.368
Latent regression (distributional head)0.3780.370
AuthorsOpenReview ↗

Response to Reviewer b5GT (2/5)

Weakness 1(c): cell states outside the pre-training corpus

Indeed generated cells are restricted to the learned codebooks. This holds for any generative model built on a fixed VQ tokenizer: in vision, VQGAN, MaskGIT and LlamaGen freeze the tokenizer after training, and how well they cover new data depends mainly on the breadth of the data the tokenizer was trained on. The same applies to SUDOKU, and a larger and more diverse pre-training corpus, which the Discussion names as future work, is the main remedy. Pre-training at larger scale is under way, on a spatial transcriptomics corpus of nearly 800 million cells.

The current model already shows generalization beyond the pre-training corpus. On spatial generation, where the tokenizer is kept frozen, two of the six tissue groups are mouse kidney, an organ absent from the pre-training corpus, one of them injured kidney, and SUDOKU is the most accurate method on both groups in Table 4. For data that differ more from the pre-training corpus, such as SPAC-seq, where CRISPR perturbations create tumor clones and surrounding tissue that do not occur in normal samples, the codebook may not cover all cell states. In this case, the tokenizer can be fine-tuned on the new sections with its own self-supervised objective, without labels, which moves the codebook toward their cell states.

Weakness 2: independent prediction of the two code levels

During the development of SUDOKU, we did try a variant whose second-level head also takes the first-level code as input, trained in the teacher-forcing fashion. However, we observed little gain over the independent-head design, so following Occam's razor we kept the simpler design. A similar design is used in MusicGen (Copet et al., 2023), which predicts multiple residual levels of its audio tokens in one forward pass, with one head per level.

Second-level headHESTA Niche retrieval (mean AUPRC)Generation (mean cell-type accuracy)
SUDOKU0.5030.405
SUDOKU (conditioned code prediction)0.5010.410

However, we do think that conditioning may become useful with more data, larger models and particularly, more quantization levels. We are currently scaling SUDOKU's pre-training to a spatial corpus of nearly 800 million cells. We appreciate this suggestion and will revisit this choice in future versions of SUDOKU.

Weakness 3: niche retrieval

The chance level varies from 0.038 to 0.152 across organs, and after normalizing the AUPRC score with the chance level, SUDOKU remains the best method. For each pair of an NOI and a target section, we define the chance level as the fraction of cells on the target section that carry the organ of the NOI, which is the AUPRC of a random ranking. The table below gives the chance level averaged over its pairs and the normalized score (AUPRC − chance) / (1 − chance), computed from the AUPRC of Table 2 and the chance level of the organ. We have added this table to Appendix H.

OrganChance levelSUDOKU (normalized)QueST (normalized)GeST (normalized)
Brain0.1520.4300.3990.347
Liver0.0760.9260.7800.834
Spinal cord0.1350.4250.3410.301
Cartilage0.1510.2450.2370.193
Ganglion0.0430.2310.1090.217
Skeletal muscle0.0700.3840.2230.366
Heart0.0380.7480.6990.643
Jaw and teeth0.0610.1550.0690.204

The NOI size is fixed without justification, and the readout radius used for retrieval is not reported.

On HESTA, the readout radius is the neighborhood radius of pre-training, 150 µm, as listed in Table A.1. Each NOI is a region of about 1000 cells, cut from one organ by k-means on the coordinates (Appendix E). This is about 0.5 mm across, a size that balances two needs of the benchmark. HESTA is annotated at the organ level, so an NOI should be a piece of tissue large enough to reflect the organization of an organ region, rather than a few cells. It should also fit inside one organ, so that small organs still give several NOIs per section: at this size, ganglion and heart give about four NOIs per section. Additionally, this scale is also close to that of the DLPFC NOIs released with QueST, which are 100 spots that each hold several cells.

As for DLPFC, this niche query benchmark is released by QueST, whose NOIs are fixed sets of 100 spots. We follow QueST for both the NOIs and the readout radius, 318 µm, which covers about 37 spots.

We have added both readout radii and the reason for the HESTA NOI size to Appendix E. The experiment below also shows that our conclusion does not depend on the NOI size.

AuthorsOpenReview ↗

Response to Reviewer b5GT (3/5)

Weakness 3: niche retrieval (continued)

If SUDOKU captures spatial dependencies at different scales, I would expect its advantage to grow for larger niches.

We ran this experiment and observed that SUDOKU performs best at every window size from 250 to 2000 cells. The performance margin over the best baseline widens gradually as the difficulty grows with NOI size. For each NOI, we grew the query window from its center through the tissue of the same organ in steps of 250 cells, and kept the NOIs whose organ holds at least 2000 cells, 84064 pairs in all. We compared SUDOKU with the two strongest baselines in Table 2, GeST and QueST.

Query window (cells)25050075010001250150017502000
SUDOKU0.5010.5010.5040.5090.5100.5120.5130.512
GeST0.4780.4790.4730.4790.4740.4710.4730.470
QueST0.4470.4500.4490.4510.4520.4520.4490.451
Gain over the best baseline+4.8%+4.6%+6.6%+6.3%+7.6%+8.7%+8.5%+8.9%

Note that the score of SUDOKU at 1000 cells slightly differs from the 0.503 in Table 5, since each window here holds exactly 1000 cells, while the original NOIs had a median of around 1000 cells. We have added this experiment to Appendix H.

Several margins are small (for example 0.765 vs 0.762 on L5L6), and there are no repeated fine-tuning seeds or confidence intervals.

We would like to clarify that in Table 2, the values of SUDOKU and of the two strongest baselines, QueST and GeST, are already means over three runs with seeds 42, 43 and 44. The other baselines were run once due to limited computational resources. Here we give the standard deviations below for SUDOKU and the best baseline of each column. On HESTA, SUDOKU leads on seven of the eight organs, and on six of them the intervals of one standard deviation around the two means do not overlap. On DLPFC, SUDOKU leads on five of the seven classes and ties QueST on L3L4, but only the margins on L4L5 and L5L6WM exceed the variation across seeds, and small margins such as L5L6 lie within it. Although the lead on DLPFC is less pronounced than on HESTA, SUDOKU leads on 12 of the 15 columns of Table 2 taken together, and on 8 of them by more than the variation across seeds.

HESTABrainLiverSpinal cordCartilageGanglionMuscleHeartJaw
SUDOKU, mean ± std over 3 seeds0.517 ± 0.0070.932 ± 0.0050.503 ± 0.0010.359 ± 0.0090.264 ± 0.0050.427 ± 0.0060.758 ± 0.0120.207 ± 0.009
Best baseline in Table 20.490 ± 0.002 (QueST)0.847 ± 0.017 (GeST)0.430 ± 0.010 (QueST)0.352 ± 0.004 (QueST)0.251 ± 0.006 (GeST)0.410 ± 0.002 (GeST)0.710 ± 0.007 (QueST)0.253 ± 0.012 (GeST)
DLPFCL3L4L4L5L5L6L6WML3L4L5L4L5L6L5L6WM
SUDOKU, mean ± std over 3 seeds0.566 ± 0.0020.698 ± 0.0090.765 ± 0.0050.751 ± 0.0090.604 ± 0.0060.723 ± 0.0020.705 ± 0.005
Best baseline (QueST) in Table 20.566 ± 0.0040.673 ± 0.0100.762 ± 0.0030.746 ± 0.0110.628 ± 0.0010.720 ± 0.0040.696 ± 0.002

We have added these results to Appendix H.

Reviewer_fRBvOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

SUDOKU learns a distribution over cell states at known tissue coordinates. A residual quantizer turns whole cell expression profiles into two discrete codes. A spatial transformer then predicts the codes of masked cells from visible cell representations and coordinates while varying the mask rate across nearly the full range. With all cells visible, its hidden states serve as representations; with cells masked, an iterative sampler fills them in. The model is pretrained on 49.1 million cells from 396 human and mouse tissue sections. The paper evaluates niche retrieval, clustering at three tissue scales, missing cell generation, and immune recruitment around perturbed tumor clones. Its central claim is that one conditional modeling objective supports both representation and generation of tissue organization.

Critical strengths
  1. The method follows a concrete limitation of existing spatial models. Fixed masking rates examine only selected amounts of context, while an autoregressive ordering changes which neighbors can condition a prediction. SUDOKU instead predicts hidden cell states from spatial context over variable masking rates. The design is technically coherent, and Table 5 gives useful evidence for its parts. Generation accuracy falls from 0.405 to 0.379 with a fixed mask rate, to 0.368 with latent regression, and to 0.323 when visible cells enter only as codes. These comparisons make the proposed objective more than a change in terminology.
  2. The evaluation tests the same backbone on several distinct uses of spatial data. Niche retrieval spans independent sections and an unseen Visium platform, while the clustering and generation tasks probe different spatial scales. GeST is pretrained on the same corpus and adapted to the benchmarks, making it a particularly informative comparator. SUDOKU leads on seven of eight HESTA organs in Table 2 and on all three clustering ARI measures in Table 3. This breadth supports the claim that the learned representation transfers across tasks, even though the generative claims need a narrower reading.
Critical weaknesses

The generation evaluation focuses mainly on cell type accuracy for masked cells. A measure of neighborhood structure would give a fuller picture of how well generated cells preserve tissue organization. This is a useful direction for future work, but it does not detract from the paper's strong results on cell state recovery and spatial representation.

Questions for authors

Did the authors examine any neighborhood-level measures on the existing generation samples, in addition to cell-type accuracy? Such results would help characterize the generated tissue patterns.

1 comment
AuthorsOpenReview ↗

Response to Reviewer fRBv

Thank you for the careful reading and the positive assessment of our work. We did not include neighborhood-level measures in the spatial generation benchmark (Table 4), but we did examine this property partly in the immune-recruitment prediction experiment (Section 4.3). There the model generates the whole band around each tumor clone, and the immune fraction of the generated band is compared with ground truth. We further examine the overall spatial neighborhood pattern in the generated band around a key knock-out clone in Figure 2c.

We agree that such measures would make the generation benchmark more complete, and we appreciate this suggestion. In future work, we plan to add the following measures:

  1. Multi-scale neighborhood composition: for each cell in the generated tissue, the cell-type composition of its neighborhood, at multiple radii or spatial scales.
  2. Cell-type co-occurrence: how often two cell types sit next to each other relative to a random placement of the same cells, at several distances.

We plan to evaluate these measures in two settings:

  1. Disease-associated neighborhoods. The generation benchmark pairs diseased and control tissue: Alzheimer's disease and control human brain, injured and control mouse kidney. We will compare the two measures between generated and measured tissue in both, for example whether glial cells gather around amyloid plaques, or immune cells and fibroblasts surround injured tubules.
  2. In silico spatial perturbation. We will alter some key cells of a region, for example, setting a gene to zero in the expression profiles of a tumor clone to mimic its knockout, re-encode the altered cells, and generate the tissue around them. We will then compare the two measures between the tissue generated around the altered region and around the unaltered one.

We have added this limitation and future direction to the Discussion of the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.