acceptodds
Under review as a conference paper at ICLR 2027

Interior-Boundary Aware Graph Representation Learning for Spatially Resolved Transcriptomic

Abstract

Spatially resolved transcriptomics (SRT) can measure gene expression while preserving spatial information, providing an important tool for studying complex tissue structures and biological processes. In recent years, graph-based methods have been widely used to capture spatial relationships between spots for spatial domain identification. However, existing methods typically adopt the same strategy to model interior and boundary regions, without explicitly considering the structural differences between them. This may make local node representations overly similar and increase the mixing of information from heterogeneous neighbors. As a result, the distinctions between spatial domains can be weakened. Intuitively, interior spots and boundary spots exhibit distinct patterns in local neighborhood features and global relation structures, motivating us to distinguish these two types of spots and model their representations with different strategies. Based on this observation, we propose Boundary-aware Multi-level Relational Learning (BMRL), which estimates boundary characteristics by combining local expression differences with global semantic structure. Building on the identified interior and boundary spots, BMRL applies different representation learning constraints to them and further incorporates structural information at the region and slice levels. This allows structural information at different levels to complement each other and improves the integration of multi-slice data. Extensive experimental results show that BMRL achieves promising performance on various downstream tasks in both single-slice and multi-slice settings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatially Resolved Transcriptomics, Boundary-aware Representation Learning, Graph Neural Network

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_h8tNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Boundary-aware Multi-level Relational Learning for spatially resolved transcriptomics. The method estimates continuous boundary scores by combining local expression differences with global semantic information derived from optimal-transport assignments to prototypes. These scores guide different learning constraints: boundary-related spots receive stronger relational consistency supervision, while interior spots receive stronger spatial smoothing and local consistency constraints. The framework further incorporates region-level prototype relations, slice-level consistency, prototype supervision, and cross-slice similarity scaling.

Critical strengths

Heterogeneous neighborhoods near tissue interfaces can make uniform spatial smoothing inappropriate. Combining local expression variation with broader semantic information provides a reasonable basis for adapting the learning constraints. The continuous weighting scheme also avoids requiring a hard boundary classification during training.

Critical weaknesses
  1. The contribution beyond the closest existing framework is insufficiently isolated. Spotscape already introduces global cross-view relation consistency, multi-granularity prototype contrastive learning, and intra-/inter-slice similarity scaling. The unweighted form of Eq. (10) corresponds to its relation-consistency objective, while Eqs. (17) and (19) closely follow its prototype and scaling mechanisms. These relationships should be explained explicitly when positioning the new contributions. Although Spotscape is included in the multi-slice experiments, it is absent from the single-slice main comparison, where its global relational modeling is particularly relevant. This omission makes it difficult to determine how much of the reported improvement comes specifically from boundary adaptation and the newly added consistency terms.

  2. The central boundary-aware mechanism is not directly validated sufficiently. Figure 1 provides a qualitative comparison with annotation-derived boundaries, but the paper does not report quantitative boundary-estimation accuracy or clustering performance separately for boundary and interior spots.

Questions for authors
  1. Methodological distinction and comparison. Could you explicitly identify the differences between BMRL and Spotscape in the relational, prototype, and similarity-scaling objectives? Please include Spotscape in the single-slice comparison and, if possible, compare it with a variant that adds only the proposed boundary adaptation.

  2. Boundary-mechanism validation. Can you quantify the agreement between the estimated boundary scores and annotation-derived boundaries, using annotations only for evaluation? Please also report clustering performance separately for boundary and interior spots, together with a control that removes boundary dependence from all applicable losses.

  3. Optimization and reproducibility. What is the complete training objective for single-slice and multi-slice experiments, including all coefficients? Please specify the rank-fusion rule, prototype counts and granularities, Sinkhorn regularization, temperature, boundary enhancement coefficient, and prototype-update schedule.

  4. Evaluation fairness. Were the baseline results rerun under the same preprocessing, graph construction, and downstream clustering protocol, or adopted from previous publications? How were hyperparameters and cluster numbers selected, and were evaluation annotations used in that process?

  5. Preservation of biological differences. The cross-slice scaling loss encourages similar within-slice and between-slice similarity levels. How does the method handle tissue domains or biological states present in only one slice?

7 comments

We thank the reviewer for carefully reading our paper and appendix and for the valuable comments and constructive suggestions. We provide our responses to the questions below.

Weakness 1 and Question 1: Distinction from Spotscape and Contribution of New Modules

Answer: We sincerely appreciate your detailed assessment.

  1. Spotscape learns spot representations through cross-view relation consistency. However, it does not explicitly consider the differences in spatial neighborhood structures between interior and boundary spots. Interior spots usually have relatively homogeneous neighborhoods. In contrast, boundary spots are more likely to be affected by the propagation of heterogeneous information from different tissue regions. Applying the same relation learning constraints to both types of spots may lead to over-smoothing near tissue boundaries. This can weaken the distinctions between different tissue regions.

    Unlike Spotscape, BMRL combines local expression differences with global semantic information derived from optimal transport. It estimates a continuous boundary score for each spot and incorporates these scores into the relation learning process.

  2. To further evaluate the effectiveness of BMRL compared with Spotscape, we conducted additional comparative experiments on single-slice spatial domain identification.

    In addition, we tried to adjust the downstream clustering configuration of Spotscape to match that of BMRL. Spotscape achieved lower clustering performance under this setting than under the configuration used in its original paper. Therefore, we retained the results reported in the original Spotscape paper to avoid underestimating the baseline performance due to differences in clustering settings. We will clearly explain the sources of these results and the differences in evaluation settings in the revised manuscript. The results are shown in the following table.

Table. Comparison of Spotscape and BMRL on different datasets. Results are reported as mean (standard deviation). Spotscape results are taken from the original paper. Bold values indicate that BMRL outperforms Spotscape, while underlined values indicate equal performance.

Dataset / SliceMethodARINMICA
DLPFC (Patient 1)
151673Spotscape0.48 (0.02)0.64 (0.01)0.61 (0.02)
BMRL0.53 (0.02)0.68 (0.02)0.64 (0.01)
151674Spotscape0.47 (0.04)0.60 (0.02)0.60 (0.03)
BMRL0.52 (0.02)0.65 (0.01)0.64 (0.02)
151675Spotscape0.45 (0.02)0.60 (0.01)0.59 (0.02)
BMRL0.53 (0.02)0.66 (0.01)0.66 (0.02)
151676Spotscape0.42 (0.05)0.58 (0.04)0.57 (0.03)
BMRL0.45 (0.04)0.61 (0.03)0.62 (0.03)
DLPFC (Patient 2)
151507Spotscape0.60 (0.03)0.72 (0.01)0.76 (0.03)
BMRL0.63 (0.04)0.72 (0.03)‾\underline{0.72\ (0.03)}0.79 (0.03)
151508Spotscape0.48 (0.05)0.64 (0.03)0.63 (0.02)
BMRL0.53 (0.02)0.67 (0.01)0.67 (0.02)
151509Spotscape0.59 (0.01)0.71 (0.01)0.70 (0.02)
BMRL0.59 (0.01)‾\underline{0.59\ (0.01)}0.72 (0.01)0.70 (0.01)‾\underline{0.70\ (0.01)}
151510Spotscape0.53 (0.04)0.67 (0.02)0.64 (0.04)
BMRL0.55 (0.03)0.68 (0.01)0.66 (0.01)
DLPFC (Patient 3)
151669Spotscape0.46 (0.02)0.58 (0.01)0.65 (0.02)
BMRL0.51 (0.01)0.62 (0.01)0.68 (0.01)
151670Spotscape0.45 (0.04)0.56 (0.03)0.66 (0.03)
BMRL0.50 (0.02)0.60 (0.01)0.68 (0.01)
151671Spotscape0.68 (0.10)0.74 (0.04)0.79 (0.08)
BMRL0.76 (0.03)0.76 (0.01)0.85 (0.03)
151672Spotscape0.75 (0.04)0.74 (0.02)0.84 (0.05)
BMRL0.80 (0.03)0.76 (0.02)0.87 (0.03)
Other Datasets
MTG ControlSpotscape0.73 (0.02)0.78 (0.01)0.75 (0.03)
BMRL0.78 (0.04)0.80 (0.02)0.80 (0.04)
MTG ADSpotscape0.68 (0.02)0.75 (0.01)0.77 (0.03)
BMRL0.69 (0.01)0.76 (0.01)0.83 (0.01)
Mouse EmbryoSpotscape0.44 (0.01)0.63 (0.01)0.54 (0.01)
BMRL0.52 (0.01)0.67 (0.00)0.60 (0.01)
NSCLCSpotscape0.57 (0.02)0.57 (0.01)0.74 (0.01)
BMRL0.62 (0.01)0.58 (0.01)0.79 (0.01)

The experimental results show that BMRL outperforms Spotscape on 15 of the 16 cases in both ARI and NMI. For CA, BMRL also achieves better performance in 15 cases. These findings demonstrate the effectiveness of BMRL and suggest that boundary-adaptive representation learning and multi-level relation modeling help improve model performance.

Weakness 2 and Question 2: Insufficient Quantitative Validation of Boundary Awareness.

Answer: Thank you for encouraging us to clarify this point.

To address this issue, we conducted additional quantitative experiments on boundary localization accuracy and clustering performance in different spatial regions.

  1. To further evaluate boundary localization performance, we calculated the Euclidean distance from each reference boundary spot to its nearest predicted boundary spot. We then normalized these distances by the typical spacing between adjacent spots. The results show an average distance of 2.5112 spot spacings.

    Previous studies (e.g., (a) and (b)) have used spatial distances to measure the proximity of spots to tissue boundaries. They also pointed out that annotations near spatial domain boundaries may contain some uncertainty. Therefore, strict spot-level matching may not fully reflect the spatial correspondence between predicted and reference boundaries.

    In addition, BMRL uses a multi-layer GNN to aggregate local spatial neighborhood information. Some spatial offsets may exist between the predicted and reference boundaries. However, the model may still capture structural information near tissue boundaries. More importantly, BMRL does not require precise localization of every boundary spot. Instead, it uses continuous boundary scores to capture tissue boundary patterns and adaptively adjust relation learning constraints at different spatial locations.

  2. We further divided the nodes into boundary and interior nodes based on the reference tissue annotations. We then calculated CA, ARI, and NMI for these two types of nodes separately. The results are shown in the following table.

    Table. Clustering performance of boundary and interior nodes. Results are reported as mean (standard deviation).

    MetricBoundary NodesInterior NodesOverall
    CA0.54 (0.02)0.91 (0.03)0.87 (0.03)
    ARI0.29 (0.01)0.86 (0.04)0.80 (0.03)
    NMI0.48 (0.01)0.82 (0.03)0.76 (0.02)

    The experimental results show that boundary nodes have much lower clustering performance than interior nodes. This indicates that clustering boundary regions is more challenging. These findings also support the need to distinguish different spatial regions during representation learning.

  3. We further removed boundary dependence from all applicable loss functions while keeping the other learning constraints unchanged. The results are presented in the following table.

    Table. Effect of removing all boundary-adaptive weights.

    MethodARINMICA
    BMRL (Full)0.7770.7750.871
    w/o Boundary Weighting0.7530.7700.840

    After removing all boundary-adaptive weights, ARI, NMI, and CA decrease by 0.024, 0.005, and 0.031, respectively. These results show that the boundary-aware mechanism provides additional performance improvements.

References

(a) Daniel C. Jones, Patrick Danaher, Youngmi Kim, Joseph M. Beechem, Raphael Gottardo, and Evan W. Newell. An information theoretic approach to detecting spatially varying genes. Cell Reports Methods, 3(6), 2023.

(b) Yunhak Oh, Junseok Lee, Yeongmin Kim, Sangwoo Seo, Namkyeong Lee, and Chanyoung Park. Global context-aware representation learning for spatially resolved transcriptomics. arXiv preprint arXiv:2506.15698, 2025.

Question 3: Optimization and Reproducibility

Answer: Thank you for your helpful suggestion.

(1) BMRL uses different training objectives for single-slice and multi-slice experiments.

Single-slice Training Objective. As shown in the following equation,

Lsingle=λRELRE+λBRCLBRC+λISLIS+λlocalLlocal+λprotoLproto+λsliceLslice.\mathcal{L}_{\mathrm{single}} = \lambda_{\mathrm{RE}}\mathcal{L}_{\mathrm{RE}} + \lambda_{\mathrm{BRC}}\mathcal{L}_{\mathrm{BRC}} + \lambda_{\mathrm{IS}}\mathcal{L}_{\mathrm{IS}} + \lambda_{\mathrm{local}}\mathcal{L}_{\mathrm{local}} + \lambda_{\mathrm{proto}}\mathcal{L}_{\mathrm{proto}} + \lambda_{\mathrm{slice}}\mathcal{L}_{\mathrm{slice}}.

Multi-slice Training Objective. As shown in the following equation,

Lmulti=Lsingle+λSSLSS+1t≥TwarmupλPCLLPCL.\mathcal{L}_{\mathrm{multi}} = \mathcal{L}_{\mathrm{single}} + \lambda_{\mathrm{SS}}\mathcal{L}_{\mathrm{SS}} + \mathbf{1}_{t\geq T_{\mathrm{warmup}}}\lambda_{\mathrm{PCL}}\mathcal{L}_{\mathrm{PCL}}.

Here, tt denotes the current training epoch, and TwarmupT_{\mathrm{warmup}} denotes the number of warm-up epochs. The Multi-scale Prototype Contrastive Loss is only used in multi-slice experiments and is activated after the warm-up stage.

The results are shown in the following table.

Table. Ranges of loss coefficients and warm-up epochs used in BMRL.

ParameterValue Range
λRE\lambda_{\mathrm{RE}}0.1–0.3
λBRC\lambda_{\mathrm{BRC}}0.65–1.5
λIS\lambda_{\mathrm{IS}}5×10−45\times10^{-4}–10−110^{-1}
λlocal\lambda_{\mathrm{local}}0.02–0.06
λproto\lambda_{\mathrm{proto}}0.07 (fixed)
λslice\lambda_{\mathrm{slice}}0.001–0.1
λSS\lambda_{\mathrm{SS}}0.1–1.0 (multi-slice only)
λPCL\lambda_{\mathrm{PCL}}0.01–0.05 (multi-slice only)
TwarmupT_{\mathrm{warmup}}100–500 epochs

(2) In Eqs. (3)–(8), we compute the local and global boundary scores, respectively. We then use percentile-rank fusion to obtain the final continuous boundary score, as defined in the following equation:

bi=βR(bilocal)+(1−β)R(biglobal).b_i = \beta R(b_i^{\mathrm{local}}) + (1-\beta)R(b_i^{\mathrm{global}}).

where R(⋅)R(\cdot) denotes the normalized rank transformation. We rank the scores in ascending order and divide each rank by the total number of spots NN. The fusion weight β\beta ranges from 0.2 to 0.9 across different experiments.

(3) Prototype-update schedule.

As shown in Eqs. (13)–(14), we perform multi-granularity K-means on the second augmented view Z(2)Z^{(2)} to generate shared region prototypes. These prototypes are used to maintain consistent semantic relations between the two augmented views and the region prototypes. Lproto\mathcal{L}_{\mathrm{proto}} is optimized from the beginning of training. The region prototypes are recomputed at each parameter update.

As shown in Eqs. (16)–(17), we use the stop-gradient representations of the second augmented view to generate prototype pseudo-labels. These labels supervise the prototype predictions of the first augmented view. LPCL\mathcal{L}_{\mathrm{PCL}} uses the same shared region prototypes. It is only used in multi-slice experiments and is activated after a warm-up period of 100–500 epochs.

(4) Other hyperparameter settings are summarized in the following table, including prototype counts and granularities, Sinkhorn regularization, temperature, and boundary enhancement coefficients.

Table. Other hyperparameter settings of BMRL.

ParameterValue
OT prototype countKK
Prototype granularities[1,1.5,2]×K[1, 1.5, 2]\times K or [1,1.25,1.5]×K[1, 1.25, 1.5]\times K
Sinkhorn regularization ϵ\epsilon0.1,median⁡(C)0.1\\,\operatorname{median}(C)
Maximum Sinkhorn iterations100
Sinkhorn stopping tolerance10−610^{-6}
PCL temperature τ\tau0.75–1.0
Boundary enhancement coefficient α\alpha0.35–1.0

Here, KK denotes the base number of clusters, and CC denotes the OT cost matrix defined in Eq. (4).

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

The paper proposes Boundary-aware Multi-level Relational Learning (BMRL), a graph representation learning framework for spatially resolved transcriptomics.

Critical strengths
  • Biologically sound design
  • Comprehensive empirical validation
Critical weaknesses

W1. Incremental methodological novelty: Combines existing techniques (graph augmentation, optimal transport, prototype learning, similarity coordination); relies heavily on paradigms closely related to prior work like Spotscape.

W2. Unvalidated boundary detection accuracy: Relies on qualitative visualizations rather than quantitative metrics to evaluate how accurately its boundary score identifies actual tissue transitions.

W3. Marginal isolated gain from boundary weighting: Ablations show that removing boundary weights while keeping relation consistency only drops ARI from 0.777 to 0.763, suggesting the core relation loss drives most performance gains rather than boundary awareness itself.

W4. Unanalyzed computational overhead: Pairwise relation consistency (O(N2)\mathcal{O}(N^2) complexity) and multi-step optimal transport introduce substantial computational load, but the paper lacks systematic runtime or memory scaling benchmarks for large datasets.

W5. Insufficient biological validation: Focuses primarily on domain label overlap rather than biological integrity—reconstruction is shown for only one gene (NR4A2), and heterogeneous disease integration risks erasing true biological variation as batch effects.

Questions for authors

Provided above

5 comments

We thank the reviewer for carefully reading our paper and appendix and for the valuable comments and constructive suggestions. We provide our responses to the questions below.

Weakness 1: Regarding Methodological Novelty

Answer: We sincerely appreciate your constructive feedback.

  • Existing graph-based methods usually learn spot representations by aggregating spatial neighborhood information. However, they often do not explicitly distinguish between interior and boundary spots during relation learning. As Reviewer h8tN pointed out, tissue boundaries are particularly vulnerable to over-smoothing caused by neighborhood aggregation.

    The main motivation of BMRL is to adopt different learning strategies for these two types of spots. For interior spots with relatively homogeneous neighborhoods, we emphasize spatial smoothing and local relation consistency. This helps preserve spatial continuity within tissue regions. For boundary spots with more heterogeneous neighborhoods, we place greater emphasis on cross-view semantic relation consistency. This helps reduce the interference from heterogeneous neighborhood information and preserve the distinctions between spatial domains.

    Unlike previous methods, BMRL explicitly incorporates continuous boundary characteristics into its relation learning objectives. This allows different spatial locations to receive different learning constraints.

  • Existing graph-based methods usually focus on neighborhood relations at the spot level. BMRL further considers structural information at the region and slice levels. As Reviewer h8tN summarized, we use multi-granularity region prototypes to model region-level semantic relations. We also use slice-level consistency constraints to preserve the overall structure of each slice. These relations at different levels complement each other. As a result, BMRL can learn local spatial structures while also using semantic information at the region and slice levels.

Weakness 2: Quantitative Evaluation of Boundary Scores

Answer: Thank you for your careful review and useful suggestions.

To further evaluate boundary localization performance, we calculated the Euclidean distance from each reference boundary spot to its nearest predicted boundary spot. We then normalized these distances by the typical spacing between adjacent spots. The results show an average distance of 2.5112 spot spacings.

Previous studies (e.g., [a] and [b]) have used spatial distances to measure the proximity of spots to tissue boundaries. They also pointed out that annotations near spatial domain boundaries may contain some uncertainty. Therefore, strict spot-level matching may not fully reflect the spatial correspondence between predicted and reference boundaries.

In addition, BMRL uses a multi-layer GNN to aggregate local spatial neighborhood information. Some spatial offsets may exist between the predicted and reference boundaries. However, the model may still capture structural information near tissue boundaries. More importantly, BMRL does not require precise localization of every boundary spot. Instead, it uses continuous boundary scores to capture tissue boundary patterns and adaptively adjust relation learning constraints at different spatial locations.

References

[a] Daniel C Jones, Patrick Danaher, Youngmi Kim, Joseph M Beechem, Raphael Gottardo, and Evan W Newell. An information theoretic approach to detecting spatially varying genes. Cell Reports Methods, 3(6), 2023.

[b] Vipul Singhal, Nigel Chou, Joseph Lee, Yifei Yue, Jinyue Liu, Wan Kee Chock, Li Lin, Yun-Ching Chang, Erica Mei Ling Teo, Jonathan Aow, et al. Banksy unifies cell typing and tissue domain segmentation for scalable spatial omics data analysis. Nature Genetics, 56(3):431–441, 2024.

Weakness 3: Contribution of the Boundary-aware Mechanism.

Answer: We appreciate your valuable comments on this aspect.

  • The original ablation experiments show that the cross-view relation consistency constraint plays an important role in model performance. However, removing the boundary weighting from LBRCL_{\mathrm{BRC}} (Eq. (10)) does not mean removing the entire boundary-aware mechanism. This is because boundary information is also used in the internal spatial smoothing loss LISL_{\mathrm{IS}} (Eq. (11)) and the local relation consistency loss LlocalL_{\mathrm{local}} (Eq. (12)).

  • To evaluate the overall contribution of the boundary-aware mechanism, we conduct an additional ablation experiment by removing all boundary-adaptive weights. The results are shown in the following table.

    Table. Effect of removing boundary-adaptive weights.

    MethodARINMICA
    BMRL (Full)0.7770.7750.871
    w/o BRC Boundary Weighting0.7630.7750.858
    w/o Boundary Weighting0.7530.7700.840

    It should be noted that the main methodological contribution of BMRL is not to replace general cross-view relation consistency learning. Instead, BMRL introduces a boundary-adaptive representation learning strategy based on it. This strategy adjusts learning constraints based on boundary scores to further improve spatial domain identification performance.

Weakness 4: Computational Cost and Scalability

Answer: We thank the reviewer for this helpful suggestion.

  • To further evaluate the computational efficiency and scalability of BMRL, we conducted additional runtime experiments on datasets of different sizes. Specifically, we compared BMRL with six methods: Spotscape, SEDR, STAGATE, SpaCAE, SpaceFlow, and GraphST. The number of spots ranged from 1,000 to 75,000. The results are shown in the following table.

    Table. Runtime comparison across different numbers of spots (seconds).

    SpotsBMRLSpotscapeSEDRSTAGATESpaCAESpaceFlowGraphST
    1k18.011.513.010.115.825.410.3
    2k32.023.327.620.529.730.821.3
    5k104.381.592.978.9108.888.286.0
    10k258.4211.1236.2222.2297.5199.9250.7
    20k683.4596.0675.6660.0896.1556.5756.2
    30k1234.61100.01348.01245.01921.71133.11458.8
    50k2442.62146.32945.52429.74470.72474.52987.0
    75k4044.03696.05340.04140.08616.04494.05106.0

    The experimental results show that the runtime of all methods increases with the number of spots. At 50k spots, BMRL takes 2442.6 seconds, which is approximately 13.8% longer than Spotscape (2146.3 seconds). However, BMRL is faster than SEDR, SpaCAE, SpaceFlow, and GraphST. Its runtime is also close to that of STAGATE.

  • We further evaluated the resource usage of seven methods on a dataset containing 75,000 spots. All methods used batch-adapted versions with the same input of 5,000 genes and a batch size of 4,096 target nodes. Each method was run in a separate process for five training epochs and one complete inference pass. The results are presented in the following table.

    Table. Resource usage comparison of seven methods on 75,000 spots. All methods use batch-adapted implementations.

    MethodParametersParameter Storage (MiB)Peak CPU Memory (GiB)Peak GPU Allocated (GiB)Peak GPU Reserved (GiB)
    BMRL2,599,4969.927.642.674.43
    Spotscape2,599,4969.923.032.634.38
    SEDR489,6561.873.513.586.91
    STAGATE5,162,82819.692.942.975.23
    SpaCAE145,0140.555.802.534.58
    SpaceFlow255,2000.973.662.626.16
    GraphST644,0972.463.133.737.43

    The experimental results show that BMRL and Spotscape have the same number of trainable parameters. Both methods require 9.92 MiB of parameter storage. The peak allocated GPU memory of BMRL is 2.67 GiB, which is close to the 2.63 GiB used by Spotscape. This represents an increase of only about 1.5%.

  • To further analyze the additional computational cost of the boundary-aware mechanism, we separately measured the time required to compute continuous boundary scores on a dataset with 75,000 spots. We used a subgraph sampling strategy. The experiments were conducted on a device equipped with an NVIDIA RTX 4070 Laptop GPU. The results are summarized in the following table.

    Table. Runtime of continuous boundary score computation on 75,000 spots.

    Computation StageRuntime (s)
    Local Expression Boundary Score1.98
    OT Boundary Score (Prototype Clustering and Spatial Adjustment)1.39
    Rank Fusion0.012
    Total Boundary Score Computation3.39

    For 75,000 spots, the computation of continuous boundary scores takes only 3.39 seconds. This accounts for approximately 0.084% of the total BMRL runtime of 4044.0 seconds.

Reviewer_tyDTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Boundary-aware Multi-level Relational Learning (BMRL) for spatial transcriptomics. It combines local expression differences and global prototype-based optimal transport to estimate boundary scores, then applies boundary-aware and interior-oriented representation learning constraints. Additional prototype-level and slice-level objectives support multi-slice integration. The overall motivation is relevant, and the method demonstrates improvements on several spatial transcriptomics benchmarks.

Critical strengths
  1. In spatial transcriptomics, tissue boundaries are indeed susceptible to over-smoothing caused by neighborhood aggregation; therefore, distinguishing between "interior" and "boundary" spots is a sound research direction.

  2. Integrating neighborhood expression differences, the entropy of prototype assignments derived via Optimal Transport (OT), and local semantic inconsistency offers a more flexible approach than relying solely on spatial distance.

Critical weaknesses
  1. The actual contribution of the core "boundary-aware" mechanism is significantly less substantial than the paper’s narrative suggests. Experimental results in the appendix show that removing the entire "relation-consistency objective" leads to a performance drop of approximately 0.137, whereas removing only its "boundary-aware weighting" component results in a decline of just about 0.014. This implies that the primary performance gains likely stem from general cross-view relational regularization, rather than the boundary-aware mechanism emphasized by the authors.

  2. Equations (3)–(8) calculate boundary scores based on local expression distances and the uncertainty of prototype assignments; however, they do not explicitly demonstrate that these scores can accurately localize true tissue boundaries. High entropy in prototype assignments does not necessarily imply that a spot lies on a tissue boundary; it could instead arise from factors such as noise in expression measurements or the presence of rare cell states. Although the paper provides visualizations of the identified boundaries, it lacks a quantitative analysis, such as precision, recall, boundary F1-score, or distance error, evaluated against manually annotated ground-truth boundaries.

  3. The optimization objective in Eq. (10) does not fully align with the "relation consistency" claimed by the authors. In essence, this objective constrains the symmetry of cross-view similarity matrices, rather than directly ensuring that the internal relationships between corresponding node pairs are preserved across the two views. The authors need to explain why cross-view symmetry serves as an appropriate proxy for maintaining boundary semantic relations, and whether it is superior to a more direct relation-preservation loss.

Questions for authors

If standard relation consistency can already account for the vast majority of performance gains, why is boundary awareness significant enough to constitute an independent, major contribution?

4 comments

We thank the reviewer for carefully reading our paper and appendix and for the valuable comments and constructive suggestions. We provide our responses to the questions below.

Weakness 1: Contribution of the Boundary-aware Mechanism.

Answer: Thank you for your valuable suggestions.

  • Removing the boundary weighting from LBRCL_{\mathrm{BRC}} (Eq. (10)) does not mean removing the entire boundary-aware mechanism. This is because boundary information is also used in the internal spatial smoothing loss LISL_{\mathrm{IS}} (Eq. (11)) and the local relation consistency loss LlocalL_{\mathrm{local}} (Eq. (12)).

  • We further evaluated the overall contribution of the boundary-aware mechanism. For this purpose, we conducted an additional ablation experiment. Specifically, we removed all boundary-adaptive weights from LBRCL_{\mathrm{BRC}} (Eq. (10)), LISL_{\mathrm{IS}} (Eq. (11)), and LlocalL_{\mathrm{local}} (Eq. (12)). The model architecture, loss terms, and other training settings remained unchanged. As a result, all spots were constrained in a uniform manner. The results are shown in the following table.

Table. Effect of removing boundary-adaptive weights.

MethodARINMICA
BMRL (Full)0.7770.7750.871
w/o BRC Boundary Weighting0.7630.7750.858
w/o Boundary Weighting0.7530.7700.840

These results show that boundary-aware weighting is an important component for improving model performance. Compared with removing only the boundary-aware weights in LBRCL_{\mathrm{BRC}} (Eq. (10)), removing all boundary-adaptive weights leads to a larger performance drop. This suggests that the boundary-aware mechanism improves performance by applying different constraints to different spatial locations.

Weakness 2: Quantitative Evaluation of Boundary Scores.

Answer: We appreciate your helpful comments.

  • We agree with the reviewer that high entropy in prototype assignments does not necessarily indicate a true tissue boundary. Previous studies have shown that combining local gene expression and spatial neighborhood information helps identify tissue structures [a]. Information-theoretic divergence has also been used to characterize spatial expression relationships [b].

    Inspired by these studies, BMRL does not rely solely on prototype assignment entropy for boundary estimation. Specifically, we combine local expression differences (Eq. (3)), prototype assignment uncertainty (Eq. (6)), and neighborhood semantic differences (Eq. (7)) to estimate continuous boundary scores. These complementary signals help reduce the limitations of using prototype assignment entropy alone.

  • Previous studies (e.g., [c] and [d]) have pointed out that spot-level spatial transcriptomics data have limited spatial resolution. Therefore, the estimated boundaries may have some spatial offsets from the reference tissue boundaries, and strict spot-level matching may not fully reflect their spatial correspondence. Based on this consideration, we evaluate boundary localization using both strict matching and different spatial tolerance distances.

  • We conducted boundary localization experiments. First, following the spatial neighborhood label-based boundary definition strategy used in [e], we constructed reference boundaries using manually annotated spatial domain labels. Specifically, for each spot, we considered its six nearest spatial neighbors. A spot was defined as a reference boundary spot if at least two of its neighbors had different labels.

    Next, we constructed predicted boundaries using the continuous boundary scores estimated by BMRL. Under a given spatial tolerance, we calculated Precision, Recall, and F1 for boundary and interior spots. The results are shown in the following table.

    Table. Boundary and interior spot evaluation results.

    Spot TypePrecisionRecallF1
    Interior Spots85.69%85.69%85.69%
    Boundary Spots56.11%79.80%62.75%

    In addition, we calculated the average distance from each reference boundary spot to its nearest predicted boundary spot. The average distance is 2.51122.5112 spot spacings. This indicates a certain degree of spatial proximity between the predicted and reference boundaries.

    These results show that BMRL can capture spatial structural information related to tissue boundaries. This further supports the feasibility of using boundary scores to adaptively adjust relation learning constraints.

References

[a] Tim Stuart, Andrew Butler, Paul Hoffman, Christoph Hafemeister, Efthymia Papalexi, William M Mauck, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija. Comprehensive integration of single-cell data. Cell, 177(7):1888–1902, 2019.

[b] Eloise Withnell and Maria Secrier. SpottedPy quantifies relationships between spatial transcriptomic hotspots and uncovers environmental cues of epithelial-mesenchymal plasticity in breast cancer. Genome Biology, 25(1):289, 2024.

[c] Daniel C Jones, Patrick Danaher, Youngmi Kim, Joseph M Beechem, Raphael Gottardo, and Evan W Newell. An information theoretic approach to detecting spatially varying genes. Cell Reports Methods, 3(6), 2023.

[d] Vipul Singhal, Nigel Chou, Joseph Lee, Yifei Yue, Jinyue Liu, Wan Kee Chock, Li Lin, Yun-Ching Chang, Erica Mei Ling Teo, Jonathan Aow, et al. Banksy unifies cell typing and tissue domain segmentation for scalable spatial omics data analysis. Nature Genetics, 56(3):431–441, 2024.

[e] Edward Zhao, Matthew R Stone, Xing Ren, et al. Spatial transcriptomics at subspot resolution with BayesSpace. Nature Biotechnology, 39(11):1375–1384, 2021.

Weakness 3: Difference Between Cross-view Symmetry and Direct Relation Preservation.

Answer: Thank you for pointing out this issue.

We further explain the design motivation of the cross-view relation consistency constraint in Eq. (10).

  • The main reason for using the cross-view symmetry constraint is that interior and boundary nodes have different spatial structural characteristics. Therefore, different relation learning strategies are needed.

    Specifically, we first identify nodes with high boundary scores. For these nodes, we constrain the symmetry of cross-view similarities. This helps maintain stable semantic relations between boundary and interior nodes across different augmented views. For nodes with low boundary scores, we directly constrain the consistency of relations between spatially adjacent nodes across the two augmented views. This helps maintain the stability of local structures in interior regions.

  • To further evaluate the effectiveness of this design, we replaced the original cross-view symmetry constraint with a direct relation preservation constraint. All other model components and experimental settings remained unchanged.

    The results are shown in the following table.

Table. Comparison of different relation constraints.

MethodARINMICA
Direct Relation Preservation0.7470.7540.849
Cross-view Symmetry (BMRL)0.7770.7750.871

The results show that the cross-view symmetry constraint achieves better performance on all three metrics. BMRL distinguishes boundary and interior nodes through boundary-adaptive learning. Under this framework, cross-view symmetry is more effective than direct relation preservation.

Questions 1: Necessity and Independent Contribution of Boundary Awareness.

Answer: Thank you for raising this important question.

General relation consistency learning can indeed improve model performance. However, stable relations do not necessarily mean that the learned representations can fully preserve differences between tissue regions. As Reviewer h8tN pointed out, heterogeneous neighborhoods near tissue boundaries may make uniform spatial smoothing unsuitable. Interior nodes usually have homogeneous neighbors. However, boundary nodes may connect nodes from different tissue regions. Therefore, applying the same learning constraints to all spatial locations may lead to over-smoothing near tissue boundaries.

Based on this consideration, BMRL uses continuous boundary scores to adaptively adjust learning constraints. This helps maintain spatial consistency in interior regions and reduce over-smoothing near tissue boundaries. The additional ablation experiments also show that the boundary-aware mechanism provides further performance improvements.

Dear Reviewer tyDT,

Thank you again for your valuable comments and suggestions.

We have carefully addressed your concerns and provided additional experiments and explanations in our rebuttal.

We would greatly appreciate it if you could take a moment to review our responses and share any further feedback.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.