acceptodds
Under review as a conference paper at ICLR 2027

HodgeDepth: Spatial Representation Learning through Global Geometry

Abstract

A central challenge in spatial representation learning is to connect local feature patterns with global geometric patterns. For example, understanding layered tissues requires connecting local molecular variation to global spatial organization. We introduce HodgeDepth, a global-geometry-aware framework that learns an interpretable depth coordinate describing progression across layers and corresponding positions along curved tissue contours. The method aligns noisy local directions of molecular change and integrates them into a coherent coordinate, making global geometry an explicit representation-learning target. The learned depth provides a spatial axis for inspecting tissue contours and gene-expression profiles, with a complementary construction for recurrent layers. Our theory characterizes geometric bias and contour recovery, establishes conditions for correct global ordering from local estimates, and bounds recovery of a canonical depth scale based on cumulative spatial mass. We evaluate HodgeDepth on 23 human cortical Visium sections across two cohorts. On the 12-section DLPFC benchmark, mean absolute depth–layer Spearman correlation increases from 0.690 for GASTON to 0.885. Downstream clustering achieves mean ARI 0.618 versus 0.533 for SEDR, the strongest of nine tested baselines, a 16% relative improvement. Controlled synthetic examples further illustrate recovery of layered geometry. Together, these results support global spatial depth as an interpretable representation of tissue organization with measurable benefits for downstream analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spatial representation learning, Global geometric structure, Hodge decomposition, Layered geometry

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_uW4kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author of this paper introduce a simple and interpretable geometric framework for learning depth coordinate of tissue from spatial transcritomics data. The idea of the paper is follow. First, identifies dominant direction of expression change at each spot, second assign the sign to these axis through spectral synchronization so the directions points consistently across the tissue and finally using Hodge projection integrate direction field into a scaler depth. The authors also provide a clean theoretical analysis when the layer structure is identifiable, and the order of layer can be recovered. On both synthetic data and Visium data outperforms the baselines.

Critical strengths
  1. The central idea of the paper is clean and clear, and they break down the problem into distinct steps make it very easy to understand and follow.

  2. I found the theoretical analysis the paper strong and substantial, especially analysis where the method can fail.

  3. Depth ordering results is very convincing ( correlation improves from 0.690 (GASTON) to 0.885, and adjacent layer AUC from 0.834 to 0.901 )

Critical weaknesses
  1. Biological scope and validation is very narrow: The experimental data is used in this study is strongly layered (cortical tissue) and its matches strongly with the assumption of the model. It's unclear how the method works on tissues with branches, niches or other competing spatial patterns.

  2. The paper lack clean and well controlled ablation to attribute the clustering gain to the depth.

  3. Limited sensitivity analysis of the hyper parameters

Questions for authors
  1. Could you please show results on a non layered tissue?

2.Could you please run every baseline through the same downstream decoder to estimate the contribution of the depth?

  1. How the results changes w.r.t the neighborhood size k and other hyper parameters?
3 comments
AuthorsOpenReview ↗

Shared-Decoder Comparisons and Hyperparameter Sensitivity

We greatly appreciate your thoughtful comments and thank you for these suggestions. The original ablations left open how much of the clustering improvement came from the learned depth coordinate under a shared decoder. Following your and other reviewers’ comments, we conducted additional ablation experiments, sensitivity analyses, and applications to non-layered tissue data. To avoid redundancy, we attach the results in the responses to the first reviewer's questions above

2. Shared-decoder comparison and other ablations.

In Point 4 of our response to Reviewer 1, we use a shared GMM to compare expression alone; expression combined with HodgeDepth or GASTON depth (the alternative single-index representation); and HodgeDepth alone, each with and without spatial voting. Adding HodgeDepth improves mean ARI over expression alone by 0.170 before voting and 0.115 after voting. Compared with expression plus GASTON depth, the corresponding gains are 0.106 and 0.117.

In our benchmark, GraphST and SpaGCN already use the same clustering decoder and spatial voting as HodgeDepth. The remaining baselines retain their native clustering pipelines. Most already incorporate gene expression and neighborhood-based smoothing or voting at intermediate stages, so applying our joint expression–depth decoder is not straightforward for those approaches.

We also ablate projector smoothing, neighborhood-asymmetry correction, eigen-contrast, and confidence weighting, including stage-specific confidence removals (Point 5). Ablating projector smoothing, neighborhood-asymmetry correction, eigen-contrast, or confidence weighting, including one stage at a time, has little effect on layer ordering on the original counts.

3. Sensitivity to neighborhood size and other hyperparameters

Point 6 reports sweeps over local regression neighborhood size kk, the number of expression PCs, smoothing strength, depth-feature weight, voting neighborhood size, and the number of voting passes. Mean absolute Spearman correlation stays between 0.887 and 0.892 for the tested settings with k≥48k \geq 48 and at least eight PCs. At 14 PCs, it falls to 0.843 at k=24k=24 and 0.467 at k=12k=12. It is stable for the tested smoothing coefficients from 0 to 1 and falls to 0.849 at λP=10\lambda_P=10. In the clustering sweeps, mean ARI ranges from 0.618 to 0.619 for depth-feature weights from 0.5 to 8, and from 0.614 to 0.619 for voting neighborhoods of 25 to 150 spots. These results indicate that HodgeDepth's strong performance is robust to hyperparameter choices.

AuthorsOpenReview ↗

Evaluation on non-layered cancer tissues

1. Evaluation on non-layered cancer tissues

To assess whether HodgeDepth is useful beyond layered cortex, we evaluated it on spatial transcriptomics data from two cancers whose pathology annotations lack a clear laminar organization.

  • Colorectal cancer. We analyzed two sections, SN123_A798015_Rep1 and SN124_A798015_Rep2, each with 12 valid pathology categories, from Valdeolivas et al. (2024).
  • Breast cancer. We analyzed spatial transcriptomics data with HE images with seven annotated pathology categories in Wu et al. (2021).

ARI values are relatively low across all methods on both datasets, consistent with the difficulty of recovering fragmented pathology regions in which spots from different categories are spatially intermingled. HodgeDepth achieves the highest ARI on both datasets. Its ARI is 0.2635 in colorectal cancer, compared with 0.1651 for the strongest baseline, SpatialPCA, and 0.3913 in breast cancer, compared with 0.2136 for the strongest baseline, SpaGCN-noHE. These results suggest that the learned depth representation can aid pathology-domain recovery even in tissues without a clear layered structure.

In addition to the colorectal and breast cancer studies, we evaluated HodgeDepth on a lung adenocarcinoma section and reported the results in our response to Point 3.2 above. HodgeDepth achieved the highest ARI in all three cancer studies, providing evidence that its utility for spatial domain recovery extends beyond layered tissues.

1.1. Colorectal cancer

MethodARI rankARI ↑NMI ↑AMI ↑Matched accuracy ↑Completeness ↑V-measure ↑FMI ↑
HodgeDepth10.26350.41850.40640.42280.46690.41850.4458
GraphST90.12930.33440.32270.29350.32070.33440.2667
GASTON80.13730.33830.32590.33090.31850.33830.2663
STAGATE60.14490.37470.36300.31500.35050.37470.2705
SEDR70.14450.37970.36820.31310.35030.37970.2668
SpaceFlow100.12890.34730.33540.27390.31680.34730.2489
SpaGCN-noHE40.15160.37600.36450.31500.34630.37600.2721
SpatialPCA20.16510.35340.34120.33090.33190.35340.2930
BayesSpace50.15020.37270.36110.34190.34720.37270.2740
BASS30.15440.37080.35890.33460.34890.37080.2825

1.2. Breast cancer

MethodARI rankARI ↑NMI ↑AMI ↑Matched accuracy ↑Completeness ↑V-measure ↑FMI ↑
HodgeDepth10.39130.33960.33760.63930.29990.33960.6479
GraphST80.13030.30800.30610.35760.24500.30800.4169
GASTON30.19680.26740.26530.40880.20910.26740.4619
STAGATE90.12650.30020.29820.37000.23530.30020.4103
SEDR100.10080.26670.26470.26310.20530.26670.3669
SpaceFlow60.15470.30790.30600.34420.23760.30790.4140
SpaGCN-noHE20.21360.32390.32200.41270.25380.32390.4712
SpatialPCA70.14350.27080.26880.38460.20950.27080.4098
BayesSpace40.19030.33480.34300.38900.26590.34480.4483
BASS50.17900.30160.29970.32520.23130.30160.4345
Reviewer_uW4kOpenReview ↗

Response to Authors’ Rebuttal and Updated Assessment

Thanks to authors for addressing my all three concerns and conducting experiments for my question 1 about non layered tissue and 3 about neighbourhood size k/hyperparameters. I suggest including these results in the main paper/ supplementary material. For question 2 of downstream decoder, the authors provided an explanation of not applying the joint expression depth decoder of HodgeDepth to all baselines but I still recommend to mention this in benchmark tables and annotate that which methods were evaluated using exactly the same downstream decoder and spatial voting as HodgeDepth and which methods were evaluated using their native clustering pipelines. This will help the readers to clearly understand the implementation details of all benchmarks including the HodgeDepth. Based on the author response I am happy to raise my rating to weak acceptance.

Reviewer_58X3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed HodgeDepth, a method for learning a global depth coordinate from expression patterns in layered tissues. Instead of learning depth directly through a neural network like GASTON, it estimates local directions of molecular change and combines them into a globally consistent coordinate using orientation synchronization and Hodge integration. The paper also introduces a source-eikonal variant for recurrent layer structures and provides theoretical analysis of depth and contour recovery. The method is evaluated on synthetic data and 23 human cortical tissue sections. The authors report improvements over GASTON in anatomical depth ordering and better downstream clustering performance compared with several ST baselines.

Critical strengths
  1. Interesting and reasonably well motivated idea. They approach the problem through local molecular directions and reconstruct a globally consistent representation rather than directly learning a scalar depth coordinate. This is a meaningful alternative to approach taken by GASTON, and the geometric interpretation is useful.

  2. Encouraging experimental results. The authors evaluate the continuous depth coordinate separately from downstream clustering. The improvement in depth-layer correlation over GASTON is substantial. The additional cortical cohort and theoretical analysis also strengthen the work.

Critical weaknesses
  1. My main concern is that it is not entirely clear which parts of the proposed method are responsible for the improvements. HodgeDepth combines several processing steps while the main downstream comparisons use different pipelines. Although the paper includes ablations, I would have preferred a more direct comparison of Hodgedepth and GASTON under the same downstream decoder. The relatively small difference between unit-normal and magnitude-weighted integration also makes me question how much of the improvement comes from the proposed depth parameterization itself.

  2. The paper emphasizes recovering global tissue geometry, but the real-data experiments mainly demonstrate improved layer ordering and clustering rather than directly validating geometric fidelity. While the synthetic experiments provide additional evidence, both real-data cohorts focus on layered cortical tissue, where a dominant depth direction is expected. This limitation is not unique to HodgeDepth, as GASTON has a similar evaluation scope. It remains unclear whether the learned coordinate would be biologically meaningful in more complex tissues with multiple spatial expression gradients.

Questions for authors

Please refer to weaknesses.

4 comments
AuthorsOpenReview ↗

Depth Representation Ablations and Scope of Geometric Validation

We greatly appreciate your insightful comments and helpful suggestions. Following your and other reviewers' suggestions, we conducted additional experiments to isolate the contribution of the learned depth representation. Specifically, under a shared downstream decoder, we compare expression alone, expression combined with HodgeDepth or GASTON depth, and HodgeDepth-only inputs, including comparisons with and without spatial voting (point 4 above). We also compare fixed and gate-selected depth coordinates (point 3 above), test unit-normal against magnitude-weighted integration (point 2 above), and ablate projector smoothing, neighborhood-asymmetry correction, eigen-contrast, and confidence weighting (point 5 above). These experiments include stage-specific confidence ablations. In addition, we conduct sensitivity analysis on those hyperparameters (point 6 above), which vary the local regression neighborhood size, number of expression PCs, smoothing strength, depth-feature weight, voting neighborhood size, and number of voting passes. To avoid redundancy, we attach the results in the responses to the first reviewer's questions above.

1.1. Isolate the contribution of learned depth. We greatly appreciate your insightful comments and helpful suggestions. Following your points, we provide additional experiments to isolate the contribution of the learned depth representation. Specifically, under a shared downstream decoder, we compare expression alone, expression combined with HodgeDepth or GASTON depth, and HodgeDepth-only inputs, including comparisons with and without spatial voting (point 4 above). We also compare fixed and gate-selected depth coordinates (point 3 above), test unit-normal against magnitude-weighted integration (point 2 above), and ablate projector smoothing, neighborhood-asymmetry correction, eigen-contrast, and confidence weighting (point 5 above). These experiments include stage-specific confidence ablations. In addition, we conduct sensitivity analysis on those hyperparameters (point 6 above), which vary the local regression neighborhood size, number of expression PCs, smoothing strength, depth-feature weight, voting neighborhood size, and number of voting passes. To avoid redundancy, we attach the results in the responses to the first reviewer's questions above.

All experimental results indicate that HodgeDepth’s advantage over GASTON depth and expression-only features under the shared decoder persists with and without voting, and depth ordering and clustering performance are stable across broad ranges around the default settings. Specifically, as reported in our shared-decoder ablation (response 4 above), expression plus HodgeDepth achieves mean ARI of 0.584 without voting and 0.618 with voting. The corresponding values for expression plus GASTON depth are 0.478 and 0.501. With expression features, scaling, and GMM settings matched, the HodgeDepth advantage is 0.106 before voting and 0.117 after the same spatial vote.

1.2. Clarification on magnitude-weighted integration ablation.

Thank you for raising this point. We also apologize for the ambiguity in our discussion of the unit-normal and magnitude-weighted integration variants. To clarify, Unit-normal: at each spot, the encoder keeps only the direction of the local expression gradient, so in Hodge integration, every spot contributes a step of the same size; Magnitude-weighted alternative: it scales each step by the local gradient strength, so regions where expression changes rapidly contribute larger steps. This small difference between the two variants indicates that the strong performance of the proposed depth representation is robust to scaling the increments by the local gradient strength, even though the coordinate values shift.

AuthorsOpenReview ↗

Beyond layered cortex

2. Geometric fidelity and evaluation beyond layered cortex

Thank you for raising this distinction. We agree with your point that the cortical benchmarks assess layer ordering and clustering, which provide indirect evidence of geometric fidelity. Direct validation would require an independent reference depth field, which is in general unavailable. However, HodgeDepth provides an interpretable way to visualize and screen high-dimensional expression patterns across a spatial domain, with many potential applications. For example, it can also support an initial spatial segmentation to flag regions for closer review by pathologists.

We agree that a single depth coordinate may capture only a dominant spatial pattern. This representation could be extended to multiple dimensions by estimating spatial gradients for different feature sets and learning a separate depth coordinate for each, allowing multiple spatial patterns to be represented jointly.

We also apologize for any lack of clarity. By “layered,” we mean that a continuous coordinate on the tissue manifold can organize spatial expression into coherent regions. This includes tissues whose level sets curve or fold and whose regions lack a visibly laminar arrangement in the observed Euclidean space. To assess downstream utility beyond layered cortex, we evaluated HodgeDepth on colorectal and breast cancer tissues (Reviewer 3, Point 1). ARIs were modest across all methods, consistent with the difficulty of recovering fragmented pathology regions where spots from different categories are intermingled. Nevertheless, HodgeDepth achieved the highest ARI on both datasets: 0.2635 in colorectal cancer versus 0.1651 for SpatialPCA, and 0.3913 in breast cancer versus 0.2136 for SpaGCN-noHE, the strongest baseline in each case. These results support the use of the learned coordinate for exploratory segmentation beyond the layered cortex.

I appreciate the authors' efforts to address my concerns, particularly through the shared-decoder comparisons and evaluations on additional tissue types. These results help clarify the contribution of HodgeDepth. However, I still have concerns regarding the direct validation of geometric fidelity and the generalizability of the learned representation. Therefore, I will maintain my original score of 2.

3. Validation and generalizability

We would like to thank you again for all your careful reading and thoughtful comments throughout the discussion. We appreciate your concerns about geometric validation and generalizability and would like to clarify both points further.

3.1 Validation through downstream analysis, visualization and interpretability

For representation learning, usually we would not expect there to be a truth. Therefore, downstream tasks such as clustering and spatial annotation provide a way to assess the utility of latent representations, including those learned by HodgeDepth and all other benchmarks such as GASTON, GraphST, SpaGCN, and SpatialPCA. We compare HodgeDepth with these baselines by measuring how well their predicted domains agree with reference annotations. We complement these comparisons with continuous-depth metrics that assess how well the learned coordinate preserves annotated cortical layer order.

In addition, a central contribution of HodgeDepth is that the learned single depth coordinate can also be mapped directly onto the tissue to visualize spatial organization. Furthermore, it can provide an interpretable axis for downstream biomarker analysis, for example by testing whether candidate gene expression is significantly associated with depth.

3.2 Additional lung cancer analysis

To assess generalizability beyond layered cortex, in addition to the colorectal and breast cancer experiments reported in our response to Reviewer 3 (Point 1), we extended to an invasive lung adenocarcinoma section containing 531,165 cells from 10x Genomics Xenium Human Lung Preview Data.

HodgeDepth achieved the highest value on every reported metric, including an ARI of 0.8631, compared with 0.5192 for the next-best method. The three cancer studies support the use of HodgeDepth for spatial domain recovery in those various complex tissues tested.

MethodARI rankARI ↑NMI ↑AMI ↑Matched accuracy ↑Completeness ↑V-measure ↑FMI ↑
HodgeDepth10.86310.71570.71430.87020.69250.71570.9078
GraphST70.24910.39470.39220.41050.34620.39470.4378
GASTON90.13520.16640.16300.26670.14300.16640.3352
STAGATE60.26260.36460.36200.38590.31100.36460.4454
SEDR40.33080.43110.42870.44610.37840.43110.5077
SpaceFlow80.17910.31140.30870.28850.26140.31140.3647
SpaGCN-noHE50.32020.50110.49900.50930.43030.50110.4970
SpatialPCA100.11980.18870.18540.27910.16260.18870.3198
BayesSpace20.51920.59070.58900.64090.51800.59070.6607
BASS30.48790.43280.43030.53370.39210.43280.6361
Reviewer_WtvHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HodgeDepth, an interpretable global depth coordinate for layered tissues learned from spatial transcriptomics data. The method combines local Jacobian estimation, sign-invariant projector fusion, orientation synchronization, and Hodge integration, with a Source–Eikonal extension for recurrent layers. The theoretical analysis characterizes geometric bias and provides recovery guarantees for depth ordering and cumulative spatial mass coordinates. On 12 DLPFC sections, HodgeDepth achieves a mean ARI of 0.618 versus 0.533 for SEDR and a mean absolute depth–layer Spearman correlation of 0.885 versus 0.690 for GASTON. Further evaluations cover 11 Batiuk sections and controlled synthetic datasets.

Critical strengths
  1. The paper explicitly distinguishes four aspects of depth reconstruction: local normal-axis estimation, orientation consistency, global integrability, and depth parameterization.

  2. The analysis goes beyond conditional recovery guarantees by characterizing the limitations and geometric bias of unit-normal integration.

  3. The paper clearly distinguish the evaluated local affine estimator from the untested neural-field extension and acknowledge that each synthetic setting uses only one data realization.

Critical weaknesses
  1. Table 1 compares complete analysis pipelines rather than encoders under a shared decoder. HodgeDepth’s clustering results depend jointly on expression features, depth coordinates, a GMM, and spatial voting. Consequently, the contribution of the proposed depth encoder cannot yet be separated from that of downstream decoding and post-processing.

  2. Table 17 primarily examines orientation strategies and increment normalization. Independent ablations of projector fusion, confidence weighting, neighborhood-asymmetry correction, and Source–Eikonal selection are missing. Controlled experiments that remove or replace individual components while holding the remaining configuration fixed would better establish their contributions.

  3. Section 5.2 reports mean ARIs of 0.5663, 0.4651, and 0.4499 for HodgeDepth, GraphST, and GASTON, respectively, whereas Tables 20–21 report 0.5686, 0.4560, and 0.4407. These discrepancies cannot be explained by rounding and require clarification regarding the evaluation samples, configurations, or run versions.

  4. The value of 0.885 in Table 2 reflects coordinate selection between signed Hodge and Source–Eikonal depth, whereas parts of the ablation discussion attribute the ordering results to signed Hodge depth alone. It is unclear whether Table 17 also includes coordinate selection. Separate results for signed Hodge depth alone and the complete selection pipeline are needed to distinguish their contributions.

  5. The evaluation includes only two cohorts, both comprising cortical tissue. Larger cohorts and additional layered tissue types would strengthen the evidence for generalizability.

Questions for authors
  1. Why do the mean ARIs in Section 5.2 differ from those in Tables 20–21: 0.5663 versus 0.5686 for HodgeDepth, 0.4651 versus 0.4560 for GraphST, and 0.4499 versus 0.4407 for GASTON? Which values are final, and do the discrepancies arise from different scored spot sets, run versions, or decoder configurations?

  2. Table 1 reports a mean ARI of 0.618 for HodgeDepth, whereas the increment-normalization ablation in Appendix C.5 reports 0.546 for unit-direction increments. Both descriptions involve a joint GMM using expression PCs and depth, followed by 50-NN spatial voting. What differences in coordinate selection, feature scaling, GMM configuration, or post-processing explain this discrepancy?

  3. Tables 2 and 17 both report a mean absolute Spearman correlation of 0.885. However, parts of the text attribute the ordering results to E2, while the main text describes gate-selected coordinates, with Source–Eikonal selected for section 151509. Does each variant in Table 17 include the gate, and are gate decisions fixed or recomputed during ablation? Please report continuous-depth metrics and ARI separately for signed Hodge depth alone (E2) and the complete gated pipeline, together with the gate decision for each section.

  4. Could the authors compare expression-only, expression plus GASTON depth, and expression plus HodgeDepth under identical expression features, feature scaling, GMM, and spatial voting settings? Results without spatial voting would also help separate the contributions of the depth representation, decoder, and post-processing.

  5. Could the authors separately remove projector smoothing, confidence weighting, and neighborhood-asymmetry correction while holding all other settings fixed, and report continuous-depth metrics? Since asymmetry correction contributes to confidence construction, please specify what each ablation removes and which stages have their weights modified when confidence weighting is disabled.

  6. Hyperparameter sensitivity and configuration transfer. How sensitive are continuous-depth metrics to the local regression neighborhood size, number of PCs, and smoothing coefficient λP\lambda_{P}? How sensitive are clustering metrics to the depth-feature weight and spatial voting neighborhood size? Are these parameters and gate thresholds fixed across sections and cohorts? Does the Batiuk evaluation directly reuse the configuration established on DLPFC? If adjustments were made, please explain the selection criteria, whether reference labels were used, and whether baselines received comparable tuning effort.

4 comments
AuthorsOpenReview ↗

Clarification on Discrepancy and Depth Encoder Ablation

1. Discrepancy in Batiuk experiment results. Thank you for catching this discrepancy. We apologize for the mismatched numbers here. Section 5.2 inadvertently retained the unupdated mean ARIs from an earlier 12-section summary that included MB7. Tables 20–21 report the final results on the 11 retained sections.

MethodEarlier mean ARI (12 sections)Final mean ARI (11 sections)Corrected median ARI (11 sections)
HodgeDepth0.56630.56860.5203
GraphST0.46510.45600.4274
GASTON0.44990.44070.4327

As listed in Appendix E.1, we excluded MB7, MB19, and MB22 from the original 14-section Batiuk inventory. These three sections were either controls or had incomplete cortical-layer annotations. For a consistent benchmark, we used the same configuration (number of clusters: K=7K=7) in every section, corresponding to layers L1–L6 and white matter. The final 11-section analysis and the earlier 12-section summary that included MB7 give the same ranking and similar ARI results, with HodgeDepth first in both. We will update the numbers in Section 5.2 to reflect the final 11-section analysis.

2. Increment-normalization ablation in Appendix C.5. Thank you for raising this great point. The ARI difference in C.5 comes from how we select the depth coordinate. In all three ablations in Appendix C.5 that focus on learned depth, we use E2 (asymmetry-corrected signed Hodge depth) across the 12 DLPFC sections. For Table 1, the label-free gate described in Appendix B.6 selects E2 or E5 (source-eikonal depth) separately for each section. Signed Hodge depth (E2) gives a directed coordinate across the layers, while source-eikonal depth (E5) is a downstream refinement built from E2 that measures distance from an inferred source layer, allowing identified layer segments across different locations to share the same depth. The table below compares the three configurations (different depth encoders) using the same decoder as the original main results (joint GMM+voting).

Depth coordinate selectionMean ARIChange from preceding row
E2 for every section (Appendix C.5, unit-direction increments)0.546
Switch to E5 on two folded sections (manually identified 151509, 151510)0.571+0.025
Label-free gate selecting E2 or E5 per section (Table 1)0.618+0.047

We will clarify in Appendix C.5 that the increment-normalization comparison fixes E2 across sections. In our response to the next question, we report the gate’s section-by-section choices and further discuss the continuous-depth metrics. The gate’s benefit mainly lies in the higher downstream clustering ARI.

3. Depth encoder ablation (Tables 2 and 17). Thank you for pointing out the ambiguity and your constructive suggestion. To clarify, the first row of Table 17 copied the results in the main paper (with gate selection) as the reference. The variants compared in Table 17 are based on E2. Essentially, E5 refines E2 so that the identified layer segments at different locations can share a depth value. The gate mainly improves clustering by helping the decoder assign those segments to the same cluster. Following your suggestion, we provide the results for E2 alone and the full gated pipeline across the 12 DLPFC sections with the same decoder.

Depth coordinateSpearman ∣ρ∣\lvert\rho\rvertKendall ∣τb∣\lvert\tau_b\rvertAdjacent-pair AUC (mean / min)Adjacent pairs in orderMonotone R2R^2Mean ARI
E2 on every section0.8870.7790.902 / 0.75462/640.9180.546
Label-free gate (original main results)0.8850.7840.901 / 0.75761/640.8830.618

Both configurations yield similar results on all continuous-depth metrics, while E2 gives a higher monotone R2R^2. The label-free gate (Appendix B.6) selected E2 for sections 151508, 151669–151671, and 151673–151676, and E5 for sections 151507, 151509–151510, and 151672. We will incorporate these results and revise Appendix C.5 accordingly.

AuthorsOpenReview ↗

Ablation Studies of Depth Estimation and Downstream Clustering

4. Ablation study for the contribution of learned depth to clustering. Thank you for the constructive comments. Following your suggestion, under the shared GMM decoder, we compare inputs with different combinations, each with and without spatial voting. We report the mean ARI across 12 DLPFC sections.

GMM inputsWithout spatial votingWith spatial voting
Expression only0.4140.503
Expression + GASTON depth0.4780.501
Expression + HodgeDepth0.5840.618
GASTON depth only0.4360.478
HodgeDepth only0.5230.556

Under the shared decoder, adding HodgeDepth to expression raises mean ARI by 0.170 without spatial voting and by 0.115 with voting. The corresponding changes from adding GASTON depth are +0.064 and −0.002. Moreover, HodgeDepth alone also outperforms expression alone and GASTON depth alone in both settings. The HodgeDepth advantage is already present without additional expression features and spatial voting. Spatial voting adds ARI in each input setting, but more significantly for expression only.

With HodgeDepth as the sole GMM input, mean ARI is 0.523 without voting, matching GraphST (with voting in its pipeline), and 0.556 with voting, exceeding the two strongest expression-based benchmarks, SEDR (0.533) and GraphST (0.523). Without voting, adding HodgeDepth to expression raises mean ARI to 0.584, exceeding every baseline in Table 1 (all include gene expression features except Gaston and Spaceflow). This gain indicates that single HodgeDepth effectively captures spatial structure for clustering.

5. Ablation of projector smoothing and confidence weighting. We ablated each component separately in the E2 encoder across all 12 DLPFC sections, holding preprocessing, local regression, graph construction, and the remaining settings fixed. In the E2 encoder for HodgeDepth, node confidence is ci=ri(1−Ai)2c_i=r_i(1-A_i)^2, where rir_i is eigen-contrast, and AiA_i is neighborhood asymmetry.

The joint ablation combines λP=0\lambda_P=0 with unit confidence, which also removes the asymmetry factor. The graph-only check uses wijw_{ij} alone as the edge weight. We evaluate with the same continuous-depth metrics.

VariantSpearman ∣ρ∣\lvert\rho\rvertKendall ∣τb∣\lvert\tau_b\rvertAdjacent-pair AUC (mean / min)Pairs orderedMonotone R2R^2Common-weight RHR_H
E2 encoder (reference)0.8870.7790.902 / 0.75462/640.9180.127
No projector smoothing0.8870.7780.901 / 0.75362/640.9180.129
No asymmetry correction0.8880.7790.905 / 0.75663/640.9220.130
No confidence weighting (all stages)0.8860.7770.901 / 0.74863/640.9190.134
All three removed0.8860.7760.899 / 0.75063/640.9180.133
No eigen-contrast0.8850.7740.894 / 0.74262/640.9120.137
Confidence off in smoothing only0.8870.7780.900 / 0.75262/640.9180.128
Confidence off in synchronization only0.8880.7800.906 / 0.75763/640.9230.127
Confidence off in integration only0.8860.7760.897 / 0.74562/640.9160.134
Graph-only edge weights0.8840.7740.897 / 0.74863/640.9130.151

The three individual removals yield nearly identical depth rankings to E2. Among the stage-specific confidence ablations, disabling confidence in integration produces the largest change in the common-weight residual.

AuthorsOpenReview ↗

Hyperparameter Sensitivity Analysis

6. Sensitivity Analysis on Hyperparameters.

6.1. Local regression neighborhood size, PCs and smoothing. We varied the local regression neighborhood and the number of expression PCs at the default smoothing coefficient λP=0.01\lambda_P=0.01. The table reports the mean absolute Spearman correlation with the annotated layer order.

Local regression neighbors4 PCs8 PCs14 PCs20 PCs
240.6700.8160.8430.820
360.7840.8460.8630.885
48 (default)0.8120.8900.8870.887
720.8280.8890.8920.892
960.8880.8900.8920.892
1440.8890.8900.8920.892

Most of the results are pretty robust, especially with enough PCs (>4). In addition, at the default neighborhood and PC count, mean ∣ρ∣|\rho| ranges from 0.887 to 0.891 across the tested λP\lambda_P values between 0 and 1, and falls to 0.849 at λP=10\lambda_P=10.

6.2. Sensitivity to depth-feature weight

Depth-feature multiplier γ\gamma0 (expression only)0.250.51 (default)248
Mean ARI with spatial voting0.5030.6030.6180.6180.6180.6180.619

Adding depth at the default weight raises mean ARI by 0.115 over expression alone. Across the tested weights from 0.5 to 8, mean ARI stays between 0.618 and 0.619, indicating little sensitivity to the exact weight in this range.

6.3. Spatial voting neighborhood

Voting neighborsNone122550 (default)100150300
Mean ARI0.5840.6110.6140.6180.6190.6180.582

Mean ARI stays between 0.614 and 0.619 for 25–150 voting neighbors. At 300 neighbors, it falls to 0.582, consistent with oversmoothing.

Configuration transfer across cohorts

We fixed the HodgeDepth hyperparameters and gate thresholds across sections and cohorts. The Batiuk evaluation directly reused the DLPFC configuration because both cohorts contain Visium sections at comparable spatial resolution. Both used 48 neighbors and 14 expression PCs for local regression, λP=0.01\lambda_P=0.01 for projector smoothing, and γ=1\gamma=1 with 50-neighbor spatial voting for clustering. The gate applies the same thresholds from Appendix B.6 independently to each section using expression and spatial information.

AuthorsOpenReview ↗

Three new non-layered cancer cohorts

7. Additional analyses on colorectal, breast, and lung cancer datasets

To assess whether HodgeDepth generalizes beyond layered cortex, we first evaluated it on colorectal and breast cancer datasets (response to Review 3, Point 1). HodgeDepth achieved the highest ARI on both benchmarks: 0.2635 across two colorectal cancer sections, compared with 0.1651 for SpatialPCA, and 0.3913 on the breast cancer dataset, compared with 0.2136 for SpaGCN-noHE. SpatialPCA and SpaGCN-noHE were the strongest baselines on their respective datasets. We also evaluated HodgeDepth on a Xenium section of invasive lung adenocarcinoma containing 531,165 cells (response to Review 2, Point 3). HodgeDepth led on every reported metric, achieving an ARI of 0.8631 compared with 0.5192 for BayesSpace, the strongest baseline by ARI.

The consistent gains across all three cancer studies support the use of HodgeDepth for spatial domain recovery beyond layered tissues.

Comments on the paper

1 comment
Area_Chair_Y1RCOpenReview ↗

Please check authors' rebuttal and participate in the discussion

Dear reviewer,

The discussion phase is coming to an end. Please read the authors’ rebuttal and post a brief follow-up in the discussion thread. Let us know which of your concerns the response addresses and which remain open.

Please post any follow-up questions in the thread visible to the author as soon as possible, so that the author has time to respond and provide clarification. I also encourage you to engage with one another’s assessments—especially where your perspectives differ—and to note whether any counterarguments or discussions lead you to revise your own assessment.

Thank you for your effort and thoughtful review. Thank you again!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.