acceptodds
Under review as a conference paper at ICLR 2027

SpecCG: Spectrally Enhanced Joint Cell–Gene Representation Learning for Spatial Transcriptomics

Abstract

Spatial transcriptomics (ST) provides spatially resolved gene expression profiles for characterizing tissue organization and cellular heterogeneity. However, existing representation learning methods largely overlook global gene expression patterns obscured by noise and are predominantly cell-centric, thereby failing to adequately model tissue-scale expression organization and gene-level dependency structures. To overcome these limitations, we propose SpecCG, a Spectrally enhanced joint Cell–Gene representation learning framework for spatial transcriptomics. Specifically, SpecCG first performs adaptive spectral modulation on spatial maps of gene expression to enhance informative global expression patterns obscured by noisy observations. Further, it employs a multi-view architecture to integrate local spatial-transcriptional context with spectrally enhanced global expression patterns for discriminative cell representations, while modeling gene co-expression dependencies as complementary biological signals. Finally, SpecCG introduces a gene–cell joint decoding module to couple cell and gene embeddings through joint reconstruction, enabling reciprocal supervision and coordinated cell–gene representation learning. Extensive experiments on eight public ST datasets demonstrate that SpecCG consistently outperforms state-of-the-art baselines and yields spatially coherent, biologically interpretable domains.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spatial transcriptomics, clustering, graph neural networks, frequency-domain enhancement, joint cell–gene representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_h7xFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SpecCG for spatial domain identification in spatial transcriptomics. Its main contribution is the integration of learnable spectral enhancement with joint cell-gene modeling, bringing tissue-scale expression patterns together with local spatial relationships and gene co-expression. The framework couples cell and gene representations through expression reconstruction and learns clustering embeddings using contrastive and clustering objectives. Evaluation against ten baselines on five DLPFC sections and three additional tissue datasets reports improved clustering performance, supported by component ablations and marker-gene analyses.

Critical strengths

SpecCG’s main strength is its combination of spectral enhancement and cell–gene modeling. This is well motivated for spatial domain identification, where tissue-scale expression patterns and local spatial relationships are both relevant. The reported component ablations support the usefulness of this combination, making it a promising application of spectral methods to spatial transcriptomics.

Critical weaknesses
  1. The evaluation protocol leaves uncertainty about the source of the reported gains. Hyperparameters differ substantially across DLPFC sections, but their selection criterion and whether baselines received comparable tuning are unclear. The released code (main_test.py, lines 524-530) also contains logic to retain the highest-ARI result across epochs using ground-truth labels, but this selection procedure is not described in Algorithm 1. If this rule was used for the reported results, the gains may partly reflect label-informed selection rather than improvements attributable to the method.
  2. The methodological advance of SpecCG remains insufficiently established. Its integration of spectral enhancement and cell-gene modeling is reasonable, but GFE’s core filtering operation closely resembles GFNet (Rao et al., 2021). The component-removal ablations support the usefulness of the modules, but do not isolate what the proposed design adds over simpler alternatives.

[1] Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global Filter Networks for Image Classification. NeurIPS, 2021.

Questions for authors
  1. Given the similarity between GFE and existing learnable Fourier filtering approaches, what is the specific methodological contribution of GFE beyond adapting spectral filtering to spatial gene expression maps? Could the authors provide controlled comparisons with simpler spatial smoothing or alternative spectral filtering methods to substantiate the necessity of the proposed design?
  2. The DLPFC dataset contains 12 annotated sections, whereas the evaluation includes only five (151510, 151669, 151670, 151671, and 151674). What criteria motivated this selection? Could the authors report the results on all 12 sections, including the overall mean performance, to establish the representativeness of the evaluation and facilitate comparison with prior work?
  3. Appendix Table 4 reports substantially different hyperparameter configurations across DLPFC sections. Moreover, Figure 9 shows considerable and irregular fluctuations in clustering performance with respect to the temperature parameter (\tau), particularly on slices 151669, 151670 and 151674. Could the authors clarify how the optimal hyperparameters were selected and whether ground-truth labels or external clustering metrics (e.g., ARI) were used during tuning? Additionally, could the authors report the clustering performance under a shared hyperparameter configuration across DLPFC slices to assess the robustness of the reported improvements?
  4. For the results reported in Table 1, which checkpoint was used for each run: the final epoch, a fixed pre-specified epoch, or an epoch selected using an evaluation metric? If the best-ARI checkpoint was used, please report its performance alongside the final-epoch performance and state whether ground-truth labels were involved in checkpoint selection. Epoch-wise ARI/NMI and training-loss curves would further clarify the stability of the reported results.
10 comments
AuthorsOpenReview ↗

Contributions of the GFE module (1)

W2 [The methodological advance of SpecCG remains insufficiently established. Its integration of spectral enhancement and cell-gene modeling is reasonable, but GFE’s core filtering operation closely resembles GFNet (Rao et al., 2021). The component-removal ablations support the usefulness of the modules, but do not isolate what the proposed design adds over simpler alternatives.]

Q1 [Given the similarity between GFE and existing learnable Fourier filtering approaches, what is the specific methodological contribution of GFE beyond adapting spectral filtering to spatial gene expression maps? Could the authors provide controlled comparisons with simpler spatial smoothing or alternative spectral filtering methods to substantiate the necessity of the proposed design?]

Response.

Thank you for your suggestion. We agree that the two methods share the same operation principle, namely two-dimensional Fourier transform, frequency domain element-wise complex multiplication and inverse transform. We will add additional references in the methods and related work.

Methodological contributions of the GFE module

Our GFE module and GFNet [1] differ significantly in specific motivations and methodological details, GFNet learns arbitrary filter weights on image feature channels. In SpecCG, each channel of the input corresponds to a gene expression map with a biological identity. For frequency (u,v)(u,v) and gene gg, we learn Z^g′(u,v)=W(u,v)Z^g(u,v)\widehat Z'_g(u,v)=W(u,v)\widehat Z_g(u,v), where<strong>same one</strong>Complex response W(u,v)W(u,v) acts on all genes. Each gene retains its own spectral amplitude and phase, so shared responses can also produce different spatial outputs.

  • The motivation for this constraint is that co-expressed genes may carry similar tissue-scale changes, and most beneficial patterns tend to have a wide range of tissue responses. Learning frequency responses that are generally beneficial to different genes can avoid estimating the entire filter separately for each sparse gene, and also reduces the computational loss and difficulty of frequency domain modulation to a certain extent.

Extended ablation experiments with simple spatial smoothing methods

We performed comparisons in conventional clustering and spatial expression pattern with reference and frequency domain signal recovery tasks respectively, and conducted them under the same parameter adjustment and training conditions to illustrate the unique advantages of the GFE module in recovering biological expression patterns compared with simple spatial smoothing alternative methods.

  • Simple spatial smoothing method:
    • Simple smooth baseline 1: Frequency hard interception filtering based on 2D DFT, intercepting the first 10%, 30% and 50% of the frequency domain components respectively.
    • Simple smooth baseline 2: Graph Laplacian filter: X~=(I−αL)X\widetilde{X} = (I - \alpha L)X, We adopt one-step graph Laplacian smoothing as the baseline, with the smoothing step size fixed at 0.5. Therefore, each spot with neighbors retains half of its own expression, and then adds half of the weighted average expression of its neighbors.

General clustering tasks:

Conclusion: Relevant empirical evidence shows that the GFE module is superior to the simple spatial smoothing method in improving the overall clustering performance of the model.

  • Experimental results.
    • We report the performance of different spatial smoothing alternatives on 4 datasets in Table 1. We can observe that GFE is better than graph Laplacian filtering, while fixed DFT is the worst, which shows that the learnable response pattern of GFE gives the model the opportunity to attenuate unhelpful components while capturing the inherent expression patterns of the data and providing clearer clustering results.
    • It is worth pointing out that the DFT in Table 1 and Table 2 respectively combines the average results of the 10%, 30% and 50% frequency components before interception to report a more representative DFT filtering performance.

Table 1. ARI of Different Spatial Smoothing Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation80.63 ± 0.1265.66 ± 0.4941.98 ± 0.1737.95 ± 0.23
Graph Laplacian smoothing82.9665.9844.5637.82
SpecCG88.3667.2349.0943.81

Table 2. NMI of Different Spatial Smoothing Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation76.94 ± 0.6271.32 ± 0.2471.05 ± 1.4354.56 ± 0.35
Graph Laplacian smoothing77.6571.9672.1454.03
SpecCG82.7573.2472.0357.72
  • Paper Revision: We have added the corresponding material to Appendix A.6.3 of the manuscript.
AuthorsOpenReview ↗

Contributions of the GFE module (2)

Spatial expression patterns and frequency domain signal recovery tasks

To quantify the unique advantages of the GFE module in terms of spatial expression pattern and spectral recovery, we selected the marker gene S100A11 with a clear white matter (WM) enrichment pattern on real DLPFC slice 151671, and regarded it as an evaluation marker with clean expression and a quantitative reference. Then add noise with the following settings:

Drawing on DUSTED's edge-emphasized dropout idea [2], apply spatially correlated dropout to non-zero counts.

Table 3. Noise Configurations

SettingBaseline Dropout RateAdditional Edge Dropout RateAffected Edge Distance (units)
-0.150.355

in The zero-setting probability outside the edge band is 0.15. Within the edge band, the basic probability and the additional probability are added together, which is 0.50. The bandwidth is 5 coordinate units. Nonzero counts were independently set to zero with location-dependent probabilities. This allows the model to face both loss of expression and stronger loss of tissue edges.

  • Spatial expression pattern recovery index
    • AUROC takes the target domain as positive and the remaining domains as negative, and measures whether expression enrichment can distinguish the target domain.
    • Spearman’s rank correlation was computed across matched spots.
  • Spectral Error Measurement
    • We directly quantify the spectral recovery error by performing graph Fourier transform GFT on clean, noisy and reconstructed representations using spatial neighbor graphs to avoid gridding. (noisy-clean, reconstructed-clean)
    • If the clean, noisy and reconstructed expressions are re-interpolated to a regular grid and then DFT is performed, the error will be mixed into the additional interpolation noise introduced by the gridding process.
    • Create GFT on the orthogonal eigenbase of the same graph Laplacian. The low frequency on the graph represents the slower changing mode between adjacent spot, so the global changes can be compared on the original tissue geometry.

Specifically, z=(x−xˉ)/xˉz=(x-\bar{x})/\bar{x} is taken for the same gene with clean, noisy and reconstructed expression, respectively, and the mean contains zero expression positions. Compute the signed coefficients c=U⊤zc=U^\top z using the same orthonormal basis UU. For each pattern and each marker gene, according to the frequency of the graph from low to high, and select the number of frequency bands that cumulatively cover 60%, 80% and 100% of non-zero frequency energy (the number of frequency bands is determined according to clean expression. This reference selection is shared by noisy and reconstruction, and we use the relative error of frequency domain recovery:

EK=∥c1:K−c1:Kclean∥2∥c1:Kclean∥2.E_K = \frac{\lVert c_{1:K}-c^{\mathrm{clean}}_{1:K}\rVert_2}{\lVert c^{\mathrm{clean}}_{1:K}\rVert_2}.

Here, the two vectors contain the first K nonzero-frequency GFT coefficients of the evaluated signal and clean reference. Their discrepancy is normalized by the reference norm, with lower error indicating better recovery.

Conclusion: The listed examples show that GFE not only effectively helps the model restore the spatial expression pattern and frequency domain signal masked by noise (DFT cannot restore the expression pattern at all and the frequency domain restoration error becomes large, graph Laplacian filtering improves layer discernibility, but the correspondence with the clean expression order has declined, and the frequency domain signal cannot be restored), but also is better than other simple spatial smoothing baselines in many aspects, indicating that the GFE module has a unique value that is different from other simple spatial smoothing baselines.

  • Experimental results:
    • The results shown in Table 4 show that the GFE module can achieve effective recovery of spatial expression patterns and frequency domain signals while improving clustering performance.
    • The other two filtering baselines did not achieve effective improvements in all indicators. The graph Laplacian filter improved the layer discriminability, but the correspondence with the clean expression order declined. The data reconstructed by DFT worsened, indicating that it not only failed to restore the original signal, but also could not solve the impact of gridding.
    • The comparison between the GFE module and DFT once again illustrates the important value of learnable frequency domain modulation and the necessity of design.
    • The results in Table 5 once again demonstrate the above observations. DFT not only fails to restore the frequency domain signal, but also further increases the error. The GFE module uses the learnable frequency domain modulation module to effectively re store the original expressed frequency domain signal from the noise, significantly reducing the frequency domain recovery error.
AuthorsOpenReview ↗

Contributions of the GFE module (3)

Table 4. Quantification of Spatial Expression Pattern Recovery on Real Datasets

METHODAUROC (%) ↑Spearman ↑ARI (%) ↑NMI (%) ↑
Noisy input56.400.3092--
DFT + KNN interpolation49.53 ± 23.910.1120 ± 0.120142.04 ± 7.4555.89 ± 0.97
Laplacian filtering86.400.199651.7457.01
SpecCG99.410.643371.2974.05

Table 5. Relative Spectral Error on Real Datasets

METHODRSE: 60% Energy ↓RSE: 80% Energy ↓RSE: 100% Energy ↓
Noisy input0.92130.96471.0738
DFT + KNN interpolation1.01121.00861.0070
Laplacian filtering0.97150.97870.9831
SpecCG0.76050.94150.9535

References

[1] Y. Rao, W. Zhao, Z. Zhu, J. Lu, and J. Zhou. Global Filter Networks for Image Classification. Advances in Neural Information Processing Systems, 34 (2021).

[2] J. Zhu, Y. Li, Z. Tang, and C. Chang. DUSTED: Dual-Attention Enhanced Spatial Transcriptomics Denoiser. Proceedings of the AAAI Conference on Artificial Intelligence, 39, 1219–1227 (2025).

  • Paper Revision: We have added the corresponding material to Sections 4.3–4.4 of the main text..
Reviewer_xUoROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies spatial domain identification from gene expression and spatial coordinates. SpecCG uses learnable two-dimensional spectral modulation to enhance global expression patterns, combines spatial-neighborhood and expression-similarity graphs to learn multiview cell representations, and incorporates a gene co-expression graph and joint cell–gene reconstruction during training. The evaluation covers eight entries, including five cortical sections and breast cancer, mouse brain, and embryo data. The method reports improvements in ARI and ACC, with ablations and marker-gene visualizations supporting the full framework. Its main contribution is to integrate global spatial information and gene dependencies into a unified model for spatial domain identification.

Critical strengths
  1. The method addresses a practically relevant problem with a design that connects global spatial patterns to gene co-expression. This combination is well motivated by the need to identify tissue domains using both spatial organization and molecular relationships.

  2. Results across multiple tissue settings are complemented by ablations and marker-gene visualizations. These provide different forms of evidence for the practical value of the complete framework.

Critical weaknesses
  1. The fairness of hyperparameter selection is unclear. SpecCG is tuned separately for each dataset, whereas the baselines largely use default settings, and it is not specified whether evaluation annotations inform parameter selection. This makes it difficult to determine how much of the reported advantage reflects the method rather than differences in tuning or possible use of evaluation labels.

  2. The evidence does not isolate the benefit of learnable spectral modulation. Removing the module supports its contribution within the full framework, but does not show that it outperforms a simpler spatial-smoothing alternative. Consequently, the claim that it recovers biological expression patterns is stronger than the current evidence supports.

Questions for authors
  1. Clarification of model selection: How were hyperparameters selected for SpecCG and each baseline? Were the annotations used for final evaluation also used in tuning, and how were differences in tuning effort handled?

  2. Evidence for the spectral contribution: Is there a comparison with a simple spatial-smoothing alternative under matched tuning and training conditions? What evidence distinguishes the claimed recovery of biological expression patterns from the effect of smoothing alone?

3 comments
AuthorsOpenReview ↗

Hyperparameter selection strategies for SpecCG and each baseline method

W1[The fairness of hyperparameter selection is unclear. SpecCG is tuned separately for each dataset, whereas the baselines largely use default settings, and it is not specified whether evaluation annotations inform parameter selection. This makes it difficult to determine how much of the reported advantage reflects the method rather than differences in tuning or possible use of evaluation labels.]

Q1[Clarification of model selection: How were hyperparameters selected for SpecCG and each baseline? Were the annotations used for final evaluation also used in tuning, and how were differences in tuning effort handled?]

Response.

Thank you for requesting clarification on model selection and parameter tuning strategies. The manuscript does not fully explain how to select parameter configurations for the baseline, which makes the "default settings" easily ambiguous.

Hyperparameter selection strategies for SpecCG and each baseline method

First of all, it is certain that the annotations of each dataset used for the final evaluation are also used to calculate the ARI when adjusting parameters. We use ARI to select hyperparameters. SpecCG performs a grid search on contrast loss weight β\beta, clustering guidance weight γ\gamma, and temperature τ\tau; candidate values ​​for the first two are [0.001,0.01,0.1,1,10,100] and temperature [0.1,0.2,...,1.0]. For the baseline, if there is an existing default configuration that is the optimal configuration called out according to the same criteria on the corresponding dataset, we use that configuration; otherwise, the same grid search and ARI selection criteria are used in their respective applicable candidate spaces. This process unifies the selection criteria and conditions, enables SpecCG and baselines to be compared under fair parameter configurations, ensures the fairness of performance comparisons and the rationality of performance gain sources, and keeps the parameter adjustment workload between methods on the same scale because they all use annotation information to select optimal hyperparameters.

  • Paper Revision: We have added the corresponding material to Appendix A.5.3 of the manuscript.
AuthorsOpenReview ↗

Extended ablation experiments with simple spatial smoothing methods (1)

W2[The evidence does not isolate the benefit of learnable spectral modulation. Removing the module supports its contribution within the full framework, but does not show that it outperforms a simpler spatial-smoothing alternative. Consequently, the claim that it recovers biological expression patterns is stronger than the current evidence supports.]

Q2[Evidence for the spectral contribution: Is there a comparison with a simple spatial smoothing alternative under matched tuning and training conditions? What evidence distinguishes the claimed recovery of biological expression patterns from the effect of smoothing alone?]

Response.

Extended ablation experiments with simple spatial smoothing methods

We performed comparisons in conventional clustering and spatial expression pattern with reference and frequency domain signal recovery tasks respectively, and conducted them under the same parameter adjustment and training conditions to illustrate the unique advantages of the GFE module in recovering biological expression patterns compared with simple spatial smoothing alternative methods.

  • Simple spatial smoothing method:
    • Simple smooth baseline 1: Frequency hard interception filtering based on 2D DFT, intercepting the first 10%, 30% and 50% of the frequency domain components respectively.
    • Simple smooth baseline 2: Graph Laplacian filter: X~=(I−αL)X\widetilde{X} = (I - \alpha L)X, We adopt one-step graph Laplacian smoothing as the baseline, with the smoothing step size fixed at 0.5. Therefore, each spot with neighbors retains half of its own expression, and then adds half of the weighted average expression of its neighbors.

General clustering tasks:

Conclusion: Relevant empirical evidence shows that the GFE module is superior to the simple spatial smoothing method in improving the overall clustering performance of the model.

  • Experimental results.
    • We report the performance of different spatial smoothing alternatives on 4 datasets in Table 1. We can observe that GFE is better than graph Laplacian filtering, while fixed DFT is the worst, which shows that the learnable response pattern of GFE gives the model the opportunity to attenuate unhelpful components while capturing the inherent expression patterns of the data and providing clearer clustering results.
    • It is worth pointing out that the DFT in Table 1 and Table 2 respectively combines the average results of the 10%, 30% and 50% frequency components before interception to report a more representative DFT filtering performance.

Table 1. ARI of Different Spatial Smoothing Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation80.63 ± 0.1265.66 ± 0.4941.98 ± 0.1737.95 ± 0.23
Graph Laplacian smoothing82.9665.9844.5637.82
SpecCG88.3667.2349.0943.81

Table 2. NMI of Different Spatial Smoothing Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation76.94 ± 0.6271.32 ± 0.2471.05 ± 1.4354.56 ± 0.35
Graph Laplacian smoothing77.6571.9672.1454.03
SpecCG82.7573.2472.0357.72
  • Paper Revision: We have added the corresponding material to Appendix A.6.3 of the manuscript.

Spatial expression patterns and frequency domain signal recovery tasks

To quantify the unique advantages of the GFE module in terms of spatial expression pattern and spectral recovery, we selected the marker gene S100A11 with a clear white matter (WM) enrichment pattern on real DLPFC slice 151671, and regarded it as an evaluation marker with clean expression and a quantitative reference. Then add noise with the following settings:

Drawing on DUSTED's edge-emphasized dropout idea [1], apply spatially correlated dropout to non-zero counts.

Table 3. Noise Configurations

SettingBaseline Dropout RateAdditional Edge Dropout RateAffected Edge Distance (units)
-0.150.355

in The zero-setting probability outside the edge band is 0.15. Within the edge band, the basic probability and the additional probability are added together, which is 0.50. The bandwidth is 5 coordinate units. Nonzero counts were independently set to zero with location-dependent probabilities. This allows the model to face both loss of expression and stronger loss of tissue edges.

AuthorsOpenReview ↗

Extended ablation experiments with simple spatial smoothing methods (2)

  • Spatial expression pattern recovery index
    • AUROC takes the target domain as positive and the remaining domains as negative, and measures whether expression enrichment can distinguish the target domain.
    • Spearman’s rank correlation was computed across matched spots.
  • Spectral Error Measurement
    • We directly quantify the spectral recovery error by performing graph Fourier transform GFT on clean, noisy and reconstructed representations using spatial neighbor graphs to avoid gridding. (noisy-clean, reconstructed-clean)
    • If the clean, noisy and reconstructed expressions are re-interpolated to a regular grid and then DFT is performed, the error will be mixed into the additional interpolation noise introduced by the gridding process.
    • Create GFT on the orthogonal eigenbase of the same graph Laplacian. The low frequency on the graph represents the slower changing mode between adjacent spot, so the global changes can be compared on the original tissue geometry.

Specifically, z=(x−xˉ)/xˉz=(x-\bar{x})/\bar{x} is taken for the same gene with clean, noisy and reconstructed expression, respectively, and the mean contains zero expression positions. Compute the signed coefficients c=U⊤zc=U^\top z using the same orthonormal basis UU. For each pattern and each marker gene, according to the frequency of the graph from low to high, and select the number of frequency bands that cumulatively cover 60%, 80% and 100% of non-zero frequency energy (the number of frequency bands is determined according to clean expression. This reference selection is shared by noisy and reconstruction, and we use the relative error of frequency domain recovery:

EK=∥c1:K−c1:Kclean∥2∥c1:Kclean∥2.E_K = \frac{\lVert c_{1:K}-c^{\mathrm{clean}}_{1:K}\rVert_2}{\lVert c^{\mathrm{clean}}_{1:K}\rVert_2}.

Here, the two vectors contain the first K nonzero-frequency GFT coefficients of the evaluated signal and clean reference. Their discrepancy is normalized by the reference norm, with lower error indicating better recovery. All methods use the same graph Fourier basis and reference-selected components for each marker and spatial layout.

Conclusion: The listed examples show that GFE not only effectively helps the model restore the spatial expression pattern and frequency domain signal masked by noise (DFT cannot restore the expression pattern at all and the frequency domain restoration error becomes large, graph Laplacian filtering improves layer discernibility, but the correspondence with the clean expression order has declined, and the frequency domain signal cannot be restored), but also is better than other simple spatial smoothing baselines in many aspects, indicating that the GFE module has a unique value that is different from other simple spatial smoothing baselines.

  • Experimental results:
    • The results shown in Table 4 show that the GFE module can achieve effective recovery of spatial expression patterns and frequency domain signals while improving clustering performance.
    • The other two filtering baselines did not achieve effective improvements in all indicators. The graph Laplacian filter improved the layer discriminability, but the correspondence with the clean expression order declined. The data reconstructed by DFT worsened, indicating that it not only failed to restore the original signal, but also could not solve the impact of gridding.
    • The comparison between the GFE module and DFT once again illustrates the important value of learnable frequency domain modulation and the necessity of design.
    • The results in Table 5 once again demonstrate the above observations. DFT not only fails to restore the frequency domain signal, but also further increases the error. The GFE module uses the learnable frequency domain modulation module to effectively restore the original expressed frequency domain signal from the noise, significantly reducing the frequency domain recovery error.

Table 4. Quantification of Spatial Expression Pattern Recovery on Real Datasets

METHODAUROC (%) ↑Spearman ↑ARI (%) ↑NMI (%) ↑
Noisy input56.400.3092--
DFT + KNN interpolation49.53 ± 23.910.1120 ± 0.120142.04 ± 7.4555.89 ± 0.97
Laplacian filtering86.400.199651.7457.01
SpecCG99.410.643371.2974.05

Table 5. Relative Spectral Error on Real Datasets

METHODRSE: 60% Energy ↓RSE: 80% Energy ↓RSE: 100% Energy ↓
Noisy input0.92130.96471.0738
DFT + KNN interpolation1.01121.00861.0070
Laplacian filtering0.97150.97870.9831
SpecCG0.76050.94150.9535

References

[1] J. Zhu, Y. Li, Z. Tang, and C. Chang. DUSTED: Dual-Attention Enhanced Spatial Transcriptomics Denoiser. Proceedings of the AAAI Conference on Artificial Intelligence, 39, 1219–1227 (2025).

  • Paper Revision: Sections 4.3–4.4 of the main text.
Reviewer_Lap6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SpecCG, a spatial transcriptomics representation learning framework combining Fourier-domain gene expression enhancement, multi-view cell graph encoding, gene co-expression modeling, and joint cell–gene reconstruction. The method applies learnable spectral modulation to gridded gene expression maps and integrates the enhanced features with graph-based cell and gene representations. The method reports substantial improvements across eight benchmark datasets.

Critical strengths
  1. Introducing frequency-domain modeling to spatial transcriptomics is a novel approach; rather than being limited to local adjacency graphs, it attempts to model spatial expression structures over longer distances.

  2. The strategy of modeling both cells and genes is sound. Extending from cell-only representation learning to gene-cell coupled decoding is well-motivated from a biological perspective.

Critical weaknesses
  1. The authors first map real-space coordinates onto a regular grid, then perform a 2D DFT on the expression map of each gene. However, real tissues have irregular shapes, a large number of unsampled locations, and varying spatial resolutions. Standard DFT implicitly involves periodic extension. The paper’s appendix also links this operation to circular convolution. Consequently, spectral domain modulation may mathematically misconnect the two sides of the tissue edge or interpret high-frequency components arising from missing regions as genuine biological signals. The authors did not fully evaluate the impact of zero-padding, interpolation, grid resolution, and tissue rotation on the final results.

  2. The authors claim that GFE can recover global expression patterns masked by noise. However, the current experiments primarily demonstrate improvements in clustering metrics and embedding visualizations. Better clustering results do not necessarily imply that the original gene expression has been accurately recovered. The spectral domain module may simply have performed some form of filtering or smoothing suitable for the current clustering task, or it may even have amplified technical artifacts associated with tissue regions. Furthermore, the noise analysis in the appendix relies on idealized assumptions and cannot directly prove that biological signals can be recovered under real ST noise structures. Consequently, there is a lack of synthetic experiments with known clean expression profiles, as well as quantitative validation of spatial-spectral recovery errors and real gene expression patterns.

  3. The author's primary comparison is with w/o GFE. However, simply removing GFE does not demonstrate the unique value of learnable Fourier filtering. Therefore, comparisons with additional simple filtering baselines are needed.

Questions for authors

See the weaknesses

5 comments
AuthorsOpenReview ↗

evaluation of grid completion methods + simple filtered baselines + grid resolution + tissue rotation (1)

W1 [The authors first map real-space coordinates onto a regular grid, then perform a 2D DFT on the expression map of each gene. However, real tissues have irregular shapes, a large number of unsampled locations, and varying spatial resolutions. Standard DFT implicitly involves periodic extension. The paper’s appendix also links this operation to circular convolution. Consequently, spectral domain modulation may mathematically misconnect the two sides of the tissue edge or interpret high-frequency components arising from missing regions as genuine biological signals. The authors did not fully evaluate the impact of zero-padding, interpolation, grid resolution, and tissue rotation on the final results.]

W3[Simply removing GFE does not demonstrate the unique value of learnable Fourier filtering, comparisons with additional simple filtering baselines are needed.]

Response.

Thanks for your advice! We have used additional experiments to quantify performance changes under different grid filling methods and filtering methods, and to evaluate the impact of grid resolution and tissue rotation on model performance.

Standard DFT implicitly involves periodic extension

We fully agree with your point: the irregular tissue shapes, unsampled locations, and the inherent periodic extension of standard DFT can indeed cause erroneous connections across tissue edges and misclassify high-frequency artifacts from missing regions as true biological signals.

  • To mitigate these effects, we have adopted the following strategies in our model design:
  • Nearest-neighbor grid interpolation can partially alleviate the spurious high-frequency signals caused by directly zero-padding missing locations.
  • For Visium data, we align the staggered spot rows before gridding to reduce the risk of introducing spurious high-frequency components when mapping the hexagonal sampling lattice onto a rectangular grid.
  • We also use learnable spectral modulation to capture frequency responses that help mitigate these effects. In future work, we will investigate how to adapt the DFT to irregularly sampled spatial data to address these issues more directly.

Combined evaluation of grid completion methods + simple filtered baselines

It is worth pointing out that SpecCG itself uses the nearest neighbor interpolation method for meshing, so it is only compared with the zero-padding method. Two simple filter baselines are used instead of the GFE module to perform frequency domain enhancement operations.

  • Filter baseline 1: By performing frequency hard interception filtering of 2D DFT on the gridded expression spectrum, the first 10%, 30% and 50% of the frequency domain components are intercepted respectively. The average performance of the three frequency preservation levels is reported in Table 1.
  • Filter baseline 2: Graph Laplacian filter: X~=(I−αL)X\widetilde{X} = (I - \alpha L)X, We adopt one-step graph Laplacian smoothing as the baseline, with the smoothing step size fixed at 0.5. Therefore, each spot with neighbors retains half of its own expression, and then adds half of the weighted average expression of its neighbors.

Conclusion: Our extended experiments show that zero-padding does cause a certain degree of performance degradation, because directly setting 0 at missing positions can easily cause additional high-frequency signal interference, and neighborhood interpolation reduces this impact to a certain extent. Compared with the simple filtering method, the clustering performance gain of using GFE for frequency domain enhancement is verified.

  • Experimental results.
    • We report the combined performance of different grid completion + filtering methods on 4 datasets in Table 1. Whether based on fixed DFT filtering or GFE, the interpolation method is basically better than zero-padding. This shows that at least the interpolation method can reduce the additional high-frequency noise introduced during the gridding process to a certain extent.
    • After removing the influence of different grid completion methods, it can be observed that GFE is better than graph Laplacian filtering, while fixed DFT has the worst effect. This shows that the learnable response mode of GFE gives the model the opportunity to attenuate unhelpful components, while capturing the inherent expression pattern of the data and providing clearer clustering results.

Table 1. ARI of Different Grid Filling and Filtering Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation80.63 ± 0.1265.66 ± 0.4941.98 ± 0.1737.95 ± 0.23
DFT + zero-padding78.03 ± 0.7465.04 ± 1.8542.15 ± 0.2837.81 ± 0.16
Graph Laplacian smoothing82.9665.9844.5637.82
GFE + zero-padding84.2766.6943.1439.25
SpecCG88.3667.2349.0943.81
AuthorsOpenReview ↗

evaluation of grid completion methods + simple filtered baselines + grid resolution + tissue rotation (2)

Table 2. NMI of Different Grid Filling and Filtering Methods (%) ↑

Method151671HBCMBAE1S1
DFT + KNN interpolation76.94 ± 0.6271.32 ± 0.2471.05 ± 1.4354.56 ± 0.35
DFT + zero-padding74.41 ± 0.1371.75 ± 0.7371.82 ± 0.3654.92 ± 0.12
Graph Laplacian smoothing77.6571.9672.1454.03
GFE + zero-padding78.3473.1571.8354.86
SpecCG82.7573.2472.0357.72

In addition, in answering your W2 (Table 9-10), we still compared GFE with the above two simple filtering baselines, and used multiple indicators to verify the advantages of the proposed model from the perspective of global expression pattern and frequency domain recovery error.

  • Paper Revision: We have added the corresponding material to Appendix A.6.3 of the manuscript.

Grid resolution and tissue rotation

Although different spatial transcriptomic data have different spatial resolutions, we adopted an adaptation strategy during the gridding process based on the data from different platforms and their spatial resolutions. For 10x Visium data, we aligned the staggered array rows to construct a regular grid. For irregularly sampled data, we performed grid mapping based on actual spatial coordinates. Therefore, the default grid has an explicit sampling basis, eliminating the need to pre-specify uniform pixel dimensions for all slice. We understand that you want to compress or expand the adapted grid size under a fixed dataset to test the impact of grid resolution on model performance. At the same time, the tissue rotation itself may also affect the size and completion effect of the constructed mesh.

  • Tissue rotation: Rotate the original spatial coordinates by 60 degrees, 90 degrees and 180 degrees around the tissue center respectively.
  • Grid resolution: In the original orientation and within the same tissue range, the grid height and width are respectively scaled to the default 0.5 or 2 times. When multiple spot fall into the same grid, the average is first taken, and then mapped back to all original spot after enhancement.
  • Conclusion:
    • Empirical results on two datasets show that the model is not sensitive to tissue rotation and the final performance is only slightly affected by it.
    • Reducing the size of the default grid resolution has a significant impact on final performance, indicating that multiple spot Merging into the same coarse grid may lose local difference information. Therefore, it also verifies the effectiveness of the data-led gridding method we adopted. The grid resolution should be determined based on the resolution of the data itself and the spatial sampling method.
  • Experimental results:
    • There are only minor fluctuations in the overall clustering performance after tissue rotation on 151671 and HBC
    • Changing the grid resolution size on 151671, especially the compressed resolution, results in significant performance degradation.

Table 3. ARI of Different Tissue Rotations and Grid Resolutions (%) ↑

DatasetRotated 60°Rotated 90°Rotated 180°0.25x resolution4x resolutionSpecCG
15167187.8188.1687.6981.9286.4888.28
HBC71.0069.8267.3967.9567.0567.24

Table 4. NMI of Different Tissue Rotations and Grid Resolutions (%) ↑

DatasetRotated 60°Rotated 90°Rotated 180°0.25x resolution4x resolutionSpecCG
15167181.8481.6082.0378.1081.2682.73
HBC71.7670.5372.4572.7072.9073.21
AuthorsOpenReview ↗

Synthetic experiments for quantifying spectral recovery errors and gene expression patterns (1)

W2 [There is a lack of synthetic experiments with known clean expression profiles, as well as quantitative validation of spatial-spectral recovery errors and real gene expression patterns.]

W3 [Simply removing GFE does not demonstrate the unique value of learnable Fourier filtering, comparisons with additional simple filtering baselines are needed.]

Response.

Synthetic experiments to quantify gene expression patterns

It is worth clarifying that in the 4.3 chapter (global gene expression distribution) in our manuscript, we have visualized the original and reconstructed spatial expression patterns of the 5 layer marker genes on 151510, and compared their distribution among the annotation layers, which provides evidence at the expression level. But it cannot give a quantified error from a known clean reference. Therefore, your proposed synthesis experiment is of great significance to our contribution to further quantification of GFE.

We first save clean expressions in three spatial layouts, and then introduce preset noise perturbations:

  • Construction of synthetic datasets
    • We constructed three synthetic data sets with different spatial layouts, containing 3,000, 3,000, and 1,500 cells respectively. Each data set has five spatial domains and 2,000 genes.
    • First use SymSim's SimulateTrueCounts() [1] to generate clean expressions of five types of cells, and then construct the spatial trend based on the spatial correlation Gaussian field of scCube [2], and assign coordinates to cells with corresponding expression profiles.
    • Each layout saves clean expression and applies two levels of spatially correlated dropout to non-zero counts, so that every noisy spot and gene has a known clean control.

Table 5. Noise Configurations

SettingBaseline Dropout RateAdditional Edge Dropout RateAffected Edge Distance (units)
10.150.253
20.150.355

The zeroing probability outside the edge band is 0.15. Within the edge zone, the basic probability and the additional probability are added, the two levels are 0.40 and 0.50 respectively, and the bandwidth increases from 3 to 5 coordinate units. Nonzero counts were independently set to zero with location-dependent probabilities. This allows the model to face both loss of expression and stronger loss of tissue edges.

  • Marker gene selection
    • Sorting by the average enrichment of each gene in the three layouts, one highest-ranking gene was fixed for each of the first three domains as a common domain marker gene across the three patterns (which can be considered to have a clean expression profile).
    • Table 6 gives the lowest enrichment value among the three modes, which is used to illustrate that fixed markers are identifiable in each layout.

Table 6. Marker gene enrichment level

Target domainShared markerMinimum log2 enrichment across patterns
Domain 1SYMSIM_GENE_13803.93
Domain 2SYMSIM_GENE_09322.47
Domain 3SYMSIM_GENE_05533.02
  • Spatial expression pattern recovery index
    • AUROC takes the target domain as positive and the remaining domains as negative, and measures whether expression enrichment can distinguish the target domain.
    • Spearman’s rank correlation was computed across matched spots.

Conclusion: Quantitative results show that the gene expression (marker gene) reconstructed by SpecCG improves significantly in both AUROC and Spearman indicators compared with the noisy data, indicating that the reconstruction not only retains the domain identification of markers, but also is closer to the spatial ordering of clean expression. This verifies that the proposed model can effectively restore the global spatial expression pattern disturbed by noise.

  • Experimental results:
    • We quantified the average recovery effect of marker genes in the three model data sets under two noise settings respectively. The means and standard deviations across data modes and marker genes are given in Table 7 and compared with the corresponding noise data.
    • The reconstruction performance under noise setting 1 is better than noise setting 2. This is obvious. The higher the noise level, the greater the challenge for reconstruction recovery. However, the overall recovery effects are very significant, indicating the effectiveness of the model in restoring gene expression patterns.

Table 7. Quantification of Spatial Expression Patterns Recovery

Noise SettingTypeAUROC (%) ↑Spearman ↑
Setting 1Noisy77.34 ± 3.370.5233 ± 0.0412
Setting 1Reconstructed98.15 ± 2.490.8105 ± 0.1785
Setting 2Noisy69.60 ± 3.590.3654 ± 0.0643
Setting 2Reconstructed96.74 ± 7.870.7301 ± 0.1331

Comments on the paper

2 comments
AuthorsOpenReview ↗

We look forward to your prompt response.

Dear Reviewer,

We have carefully and sincerely considered all the Weaknesses and Questions you raised. We have designed and added relevant experiments and analyses, and provided point-by-point responses to your concerns and suggestions. We have also updated the manuscript and marked the revisions. We would greatly appreciate it if you could take the time to review these updates.

As the author discussion period is nearing its deadline, we would appreciate your feedback on our responses and revisions at your earliest convenience. If you find that our revisions adequately address your concerns, we would be grateful if you would consider increasing your score accordingly. If you have any further questions or suggestions, we would be happy to respond and continue the discussion.

Thank you again for your valuable suggestions and constructive criticism, which have helped us improve the quality of our manuscript.

AuthorsOpenReview ↗

Summary of Supplementary Experiments and Related Revisions

Dear AC and Reviewers,

We thank the reviewers for their constructive feedback. We have provided point-by-point responses, conducted additional experiments, and updated the manuscript with marked revisions. The main additions are summarized below.

Reviewer Lap6

  • Gridding and geometric sensitivity. We compared KNN interpolation with zero-padding and evaluated tissue rotations of 60°, 90°, and 180°, together with coarser and finer computational grids. These experiments characterize sensitivity to gridding choices and support the use of platform-adapted grids. The interpolation and filtering comparisons have been added to Appendix A.6.3.

  • Quantitative expression and spectral recovery. We constructed three synthetic datasets with known clean expression profiles and two levels of spatially varying dropout. Across shared domain markers, SpecCG improved AUROC and Spearman and reduced relative GFT coefficient error at all evaluated retained-energy levels. This provides expression-level evidence beyond clustering and embedding visualizations. Dataset construction is described in Appendix A.5, and recovery results and spatial visualizations are included in Section 4.3.

  • Value beyond simple spatial smoothing. We replaced GFE with fixed DFT low-pass filtering using hard truncation and graph Laplacian smoothing. On four datasets, GFE achieved higher ARI than both alternatives. On DLPFC section 151671, the S100A11 experiment further showed improvements in spatial discrimination, spot-wise expression ranking, and graph-spectral recovery. These controlled comparisons are included in Section 4.4 and Appendix A.6.3.

Reviewer xUoR

  • Hyperparameter selection. We clarified that SpecCG and the baselines use ARI-based grid search, including the use of evaluation annotations during tuning. Existing baseline configurations are retained when they were selected using the same criterion; otherwise, the baselines are tuned within their applicable search spaces.

  • Value beyond simple spatial smoothing. We replaced GFE with fixed DFT low-pass filtering using hard truncation and graph Laplacian smoothing. On four datasets, GFE achieved higher ARI than both alternatives. On DLPFC section 151671, the S100A11 experiment further showed improvements in spatial discrimination, spot-wise expression ranking, and graph-spectral recovery. These controlled comparisons are included in Section 4.4 and Appendix A.6.3.

Reviewer h7xF

  • Methodological contribution. We clarified the relationship to GFNet and the motivation for sharing a learnable frequency response across gene expression maps. The additional filtering comparisons provide evidence for the value of learnable spectral modulation within SpecCG.

  • DLPFC coverage and parameter robustness. We extended evaluation to all 12 annotated DLPFC sections. SpecCG achieved mean ARI and NMI of 72.61% and 73.73%, respectively. We also reported all-section results under a shared hyperparameter configuration in the rebuttal to quantify sensitivity to section-specific tuning.

  • Checkpoint selection and convergence. We clarified that the reported results use the prespecified final checkpoint at epoch 300, with no ground-truth labels used for checkpoint selection. The mistakenly uploaded development script has been replaced with the correct code in the supplementary material. Epoch-wise clustering metrics and training-loss curves have been added to Appendix A.6.6 to document training behavior.

We hope these additions help clarify the contribution of SpecCG and the evidence supporting its spatial domain identification and global expression pattern recovery.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.