SAINT: Sequence-Aware Integration for Spatial Transcriptomics Multi-View Clustering

Q1: Although the manuscript mentions scenarios where spots have similar genomic profiles but differ in dominant gene expression, it lacks direct experimental evidence or case studies to validate this point. Current experiments mainly address overall clustering performance without explicitly analyzing or visualizing such cases. It is recommended to include examples highlighting instances where inconsistent dominant gene expression leads to misclustered spots to support this claim.

RQ1: Thanks for your valuable suggestions! We fully agree that explicitly validating SAINT's advantage in scenarios where spots have similar global expression profiles but differ significantly in dominant gene expression is essential. To this end, we designed an additional targeted experiment on the DLPFC dataset to examine this case in detail. We selected multiple spot pairs under two stringent criteria: (1) a Pearson correlation of global gene expression profiles greater than 0.90, ensuring that the overall expression patterns are highly similar; and (2) a fold-change in expression levels of dominant genes (highest expressed genes) greater than or equal to 2, ensuring a clear biological distinction in dominant gene activity.

We then compared the clustering outcomes for these spot pairs using both the baseline MAFN and our proposed SAINT model. The detailed results are presented in the following table.

The results demonstrate that MAFN often fails to separate biologically distinct spots with highly similar expression profiles, grouping them into the same clusters despite clear differences in their dominant genes. In contrast, SAINT successfully distinguishes the majority of these challenging pairs, confirming its ability to capture subtle biological differences.

For example, the pair TMSB10 (4.26) and TUBA1B (4.45) clearly illustrates this improvement.

• TMSB10 encodes thymosin beta-10, a regulator of actin cytoskeleton organization involved in cell motility, tissue remodeling, and tumor progression.
• TUBA1B encodes alpha-tubulin, a structural component of microtubules essential for cell architecture and intracellular transport.

These two genes participate in fundamentally different cytoskeletal systems and regulate distinct biological processes, implying divergent cellular states. Consequently, they should not be grouped into the same cluster. However, due to their similar overall expression levels, MAFN mistakenly clusters them together, whereas SAINT correctly separates them by leveraging sequence-derived biological information, thereby reflecting their functional divergence.

Beyond this case-specific observation, the experiment highlights an important property of SAINT:

• By integrating nucleotide-level biological priors, SAINT leverages functional information that expression-only models overlook.
• This enables SAINT to correctly separate spots that remain indistinguishable under conventional clustering strategies, particularly when expression similarity is misleading.

Overall, these findings strongly support our claim that SAINT enhances clustering accuracy in biologically complex scenarios where traditional methods struggle. They also confirm that incorporating nucleotide sequence information provides substantial benefits in capturing functional differences that are not evident from expression data alone. We will include this expanded analysis, together with additional visualization, in the revised manuscript to clearly demonstrate this advantage of SAINT.

Q2: The concerns on methodological novelty and contribution to the community.

RQ2: We sincerely thank the reviewer for raising this important concern. Below we clarify the novelty and broader contributions of SAINT:

1. SAINT explicitly incorporates nucleotide-level biological priors into spatial transcriptomics clustering.

While SAINT leverages established techniques such as graph convolutional networks and transformer-based encoders, its core innovation lies in systematically integrating nucleotide-level biological information. Existing methods rely solely on spatial and expression data, whereas SAINT uses sequence-derived embeddings to capture regulatory roles and functional properties of genes, enabling it to distinguish biologically distinct spots that expression-only models cannot separate.

2. SAINT addresses the previously overlooked challenge of fusing heterogeneous data modalities.

Integrating spatial, expression, and nucleotide sequence information is inherently difficult due to modality discrepancies and high-dimensional representations. SAINT introduces a biologically meaningful aggregation strategy at the spot level, effectively bridging these modalities to enhance clustering quality.

3. SAINT contributes to the community by constructing multiple sequence-augmented datasets.

Given the lack of publicly available sequence-annotated ST datasets, we created datasets covering 14 tissue sections from three benchmarks (DLPFC, HBC, MBA). These datasets not only support our validation but also provide valuable resources for future research in this field.

4. SAINT achieves significant experimental gains over existing methods.

Comprehensive experiments show that SAINT consistently outperforms state-of-the-art models. For example, compared to MAFN, SAINT improves the average ARI by up to 10.3% and NMI by up to 16.2% on certain slices (e.g., DLPFC 151507). These results demonstrate that embedding nucleotide sequence-derived knowledge enriches the semantic and biological relevance of learned representations.

Overall, SAINT offers more than a simple integration of existing components. It introduces a new biological perspective, provides novel datasets, and achieves substantial empirical improvements. We believe these contributions provide substantial value and impact to both the computational biology and broader machine learning communities. Nonetheless, we recognize the importance of this comment and will emphasize these conceptual and empirical advancements more explicitly in our revised manuscript.

We sincerely thank the reviewer again for the constructive suggestions and hope that our responses address your concerns.

The rebuttal effectively addresses my concern about similar gene embeddings with differing biological functions. As a result, I have decided to increase my score.

Thanks for your valuable suggestions! We respond to your questions point by point carefully, and we hope our responses can address your concerns.

Q1: How do you determine ground-truth clusters? Do you trust the data and annotation quality? Are there any dataset-specific considerations e.g. variations in annotation methodology or annotator agreement that could influence your results?

RQ1: We appreciate the reviewer’s thoughtful question. In our experiments, we followed standard practice from previous works (e.g., Spatial-MGCN[1], MAFN[2], and GraphST[3]) by using the manual anatomical annotations provided by the original dataset authors as ground-truth labels.These expert-defined annotations are widely adopted benchmarks in the field. For example, DLPFC labels rely on neuropathologist-defined histological landmarks, whereas Visium Mouse Brain datasets use atlas-derived anatomical boundaries, which may introduce inconsistencies. To mitigate these biases, we evaluated SAINT on datasets with diverse annotation sources (DLPFC, HBC, MBA, and additional Visium samples), and the consistent improvements observed suggest our conclusions are not overly sensitive to annotation schemes. Nevertheless, annotation variability remains an inherent limitation of current ST benchmarks.

Q2: Could your report your analysis for a larger set of spatial datasets and include uncertainty estimates in your evaluation?

RQ2: We appreciate the reviewer's valuable suggestion. To address this concern, we conducted additional evaluations on two larger and more diverse 10x Genomics Visium datasets. The key statistics of these datasets are summarized in the following table.

The comparative performance results (mean ± standard deviation) for ARI and NMI on the two datasets are summarized below.

The results clearly demonstrate the superior performance of SAINT-SA compared to existing SOTA methods on both datasets, with statistically significant improvements observed in both ARI and NMI metrics. Specifically, on the H&E dataset, SAINT-SA achieved approximately a 3.8% improvement in ARI over Spatial-MGCN and about a 1.2% improvement over SAINT-G. For the Fluorescent dataset, SAINT-SA notably enhanced ARI performance by about 5.5% compared to Spatial-MGCN and approximately 1.9% compared to SAINT-G.

These improvements demonstrate the robustness and effectiveness of our proposed model across larger-scale spatial samples and more complex datasets.

Q3: Lack of uncertainty evaluation, making it difficult to assess the significance and reliability of the reported results in the figures and tables.

RQ3: We appreciate the reviewer’s suggestion. To address this, we conducted additional experiments using five different random seeds and reported the mean performance along with the standard deviation. The results are summarized below.

These results show that the performance variations across different seeds are small (standard deviations are typically below 0.015), confirming that the proposed model produces stable and reliable clustering results.

Q4: To what extent does the proposed model build on existing works like Spatial-MGCN and MAFN in terms of architecture and experimental evaluation? Were the experimental setups, dataset choices, and benchmarking protocols directly taken from prior studies, and how are these similarities and differences addressed in the manuscript?

RQ4: We appreciate the reviewer's insightful comments. Here, we thoroughly clarify and emphasize the distinctiveness of our contributions. Our evaluation adopts the benchmarking setup from MAFN[2] in terms of datasets, metrics, and protocols to ensure fair comparison. However, SAINT introduces several fundamental innovations beyond these works:

Firstly, Spatial-MGCN[1] and MAFN[2] primarily rely on integrating spatial and gene expression information through GCNs and adaptive fusion strategies, without incorporating nucleotide-level biological knowledge encoded within gene sequences. In contrast, SAINT explicitly integrates nucleotide sequence information of expressed genes into the spatial clustering process.

Secondly, SAINT introduces a sequence-aware encoder utilizing the pretrained Nucleotide Transformer. Unlike Spatial-MGCN and MAFN, SAINT converts gene sequences into meaningful high-dimensional embeddings, capturing functional and regulatory genomic signals otherwise inaccessible from expression data alone.

Thirdly, while both Spatial-MGCN and MAFN incorporate spatial and expression graphs, SAINT uniquely integrates sequence-derived representations using a late fusion module combined with cross-modal decorrelation loss (DICR).

Finally, another important contribution of SAINT is the creation of sequence-augmented datasets across multiple widely-used benchmarks (DLPFC, HBC, MBA). Neither Spatial-MGCN nor MAFN has addressed dataset augmentation at the nucleotide sequence level.

In summary, while we indeed employ the evaluation framework proposed by MAFN to facilitate robust comparative assessments, SAINT's core methodological advancements deserve special emphasis. We agree with the reviewer on clearly highlighting these methodological distinctions in the revision, explicitly referencing relevant prior works to transparently contextualize SAINT's innovation.

Q5: How did you decide which samples to include in the main table versus the appendix? It might be fairer to report all results in the main paper, including those where your method is not the best. Also, why were some methods from Table 4 not visualized in Figure 5 (A–C)?

RQ5: We thank the reviewer for this helpful comment. In the original submission, some results were placed in the appendix due to space constraints. Including full tables in the main text would have required reducing row height and spacing, making them less readable. This choice was not based on performance. For instance, Table 4 shows that SAINT is not the top performer on slices 151673 and 151674, yet results for these and other slices (e.g., 151675 and 151676) were also moved to the appendix.

For Figure 5 (A–C), we displayed only a subset of methods to keep the figure concise, while the full visualizations are provided in Figures 8 and 9 of the appendix. These extended results show trends consistent with the main text, confirming SAINT’s strong clustering performance. We will update the manuscript to clarify and improve the presentation of these results.

Q6: Have you considered any further downstream tasks beyond clustering spatial observations?

RQ6: We appreciate the reviewer’s insightful question. While our current work focuses on clustering spatial observations, the embeddings learned by SAINT provide a foundation for several downstream tasks in spatial transcriptomics.These include disease subtype classification by distinguishing pathological regions, spatial domain-guided differential expression analysis to identify genes driving heterogeneity, and mutation impact assessment linking sequence variation to spatial gene regulation. We plan to explore these directions in future work to further demonstrate the utility of our framework.

Q7: Figure 5 (A) What does tissue hires image mean? Maybe it is a typo?

RQ7: We thank the reviewer for catching this ambiguity. The label “tissue hires image” in Figure 5(A) was intended to refer to the high-resolution histology image provided with the DLPFC dataset. We agree that the current wording may cause confusion, and we will revise it to “high-resolution tissue image” in the updated manuscript to ensure clarity.

Q8: The authors should mention that the limitations section should more clearly highlight the similarities and differences with existing works, discuss annotation and data quality.

RQ8: We thank the reviewer for this comment. As discussed in our response to RQ4, we have already clarified how SAINT differs from existing methods while sharing some evaluation protocols for fair comparison. In the revised manuscript, we will explicitly highlight these distinctions, ensuring these aspects are clearly stated in the Limitations section.

References.

[1] Wang, Bo, et al. "Spatial-MGCN: a novel multi-view graph convolutional network for identifying spatial domains with attention mechanism." BiB (2023).

[2] Zhu, Yanran, et al. "Multi-view adaptive fusion network for spatially resolved transcriptomics data clustering." TKDE (2024).

[3] Long, Yahui, et al. "Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST." Nature Communications (2023).

Thanks for your valuable suggestions! We respond to your questions point by point carefully, and we hope our responses can address your concerns.

Q1: Although the paper references the MFAN work for hyperparameter settings, the sensitivity analysis provided in Section 4.4 (RQ3) is not sufficiently comprehensive. While this section examines the impact of embedding dimensions across different datasets, it does not analyze the influence of the model's own loss-related hyperparameters, such as $𝛼$ and $𝛾$, on clustering performance. Expanding the analysis to include these factors would provide a more complete understanding of the model's behavior.

RQ1: We sincerely thank the reviewer for the helpful suggestion to expand the sensitivity analysis of our model's hyperparameters. In response, we not only examined the impact of embedding dimensions as requested but also performed an additional analysis on the two loss-related hyper-parameters, $α$ and $γ$.

For the embedding dimensions $d_1$ and $d_2$, the results show that ARI and NMI fluctuate within a narrow range, with variations not exceeding approximately 7%. This confirms that the model exhibits stable performance and is not particularly sensitive to the choice of embedding dimensions.

Meanwhile, for the loss hyper-parameters $𝛼$ and $𝛾$, the results exhibit somewhat larger fluctuations, with ARI and NMI varying by up to around 12% across the tested settings. This variation is expected given the wide scale of parameter values explored. The best performance is observed when both $𝛼$ and $𝛾$ are set to 0.1, suggesting this configuration is particularly effective.

Overall, these results confirm that our model maintains robust clustering performance over reasonable variations in hyper-parameter settings.

Q2: The experimental setup section is insufficiently detailed. In particular, SAINT has two variants: SAINT-G and SAINT-SA. However, in the subsequent experimental analysis (e.g., in section 4.3 Ablation Study, RQ2), it is not clearly specified which model was used.

RQ2: We thank the reviewer for the insightful comments. We acknowledge that some parts of the experimental setup and the corresponding analyses were not sufficiently clear, particularly in Section 4.3 (Ablation Study, RQ2). In the original Figure 3, there was a labeling mistake: the middle column labeled w. SA should actually be w. G, which corresponds to the variant SAINT-G that uses averaged gene-level embeddings. The rightmost column labeled SAINT correctly refers to SAINT-SA, our final model incorporating sequence-aware attention.

In the revised manuscript, we have corrected this labeling error and added clearer descriptions of the experimental settings, including the selection of model hyperparameters, the construction of graphs, and so on. We have also unified the naming conventions across all figures and tables to avoid any further confusion. Thank you again for highlighting this important point!

Q3: Some acronyms (e.g., "STC") are not defined, which lowers the readability of the paper.

RQ3: We thank the reviewer for pointing out this issue. We acknowledge that some acronyms, such as STC, were not explicitly defined in the original submission. In the revised manuscript, we have clarified that STC stands for Spatial Transcriptomics Clustering and have ensured that all acronyms are properly introduced upon first use to improve readability.

Q4: Please provide a clear explanation of how the model derives the clustering results from the final embedding representations. Is a clustering algorithm (e.g., K-means, Leiden, etc.) directly applied to generate the results?

RQ4: We appreciate the reviewer's interest in clarifying how clustering results are obtained from the learned embeddings. In our framework, after the model outputs the final spot-level embeddings, we apply the K-means algorithm directly on these embeddings to obtain the clustering results. The number of clusters $𝐾$ is set to match the number of annotated tissue regions for evaluation purposes. The predicted cluster assignments are then compared with the ground-truth labels to compute ARI and NMI scores. We will clarify this procedure explicitly in the revised manuscript.

Q5: The manuscript does not clearly specify the parameter settings for the comparison methods. Is the default parameter setting used? Furthermore, are the experimental results based on a single experiment or repeated multiple experiments? If the results are from a single experiment, how does the author ensure the fairness and repeatability of the results?

RQ5: We sincerely appreciate the reviewer’s detailed question. For all comparison methods, we used the default parameter settings reported in their original publications to ensure fairness. For our proposed model, the reported results are based on multiple repeated experiments rather than a single run, thereby accounting for randomness during training. Similar to previous works such as Spatial-MGCN [1], MAFN [2], and GraphST [3], we fixed the random seed in all experiments to guarantee reproducibility.

To further evaluate stability, we conducted experiments with five different random seeds and reported the mean and standard deviation of ARI and NMI. The results, summarized in the table below, show that variations across seeds are small (standard deviations typically below 0.015), confirming the reliability and fairness of our reported results.

Q6: Paper Formatting Concerns: The tables presented in the manuscript have insufficient row height, resulting in partial data being obscured. We recommend increasing the table height appropriately to ensure complete display of all results.

RQ6: We sincerely thank the reviewer for pointing out this issue. We acknowledge that, due to space constraints in the original submission, the row height and spacing of some tables were compressed, which resulted in partial data appearing unclear. In the revised manuscript, we have adjusted the formatting of all tables and figures to increase row height and spacing, ensuring that all results are displayed clearly and are easy to read.

References.

[1] Wang, Bo, et al. "Spatial-MGCN: a novel multi-view graph convolutional network for identifying spatial domains with attention mechanism." BiB (2023).

[2] Zhu, Yanran, et al. "Multi-view adaptive fusion network for spatially resolved transcriptomics data clustering." TKDE (2024).

[3] Long, Yahui, et al. "Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST." Nature Communications (2023).

Thanks for your valuable suggestions! We respond to your questions point by point carefully, and we hope our responses can address your concerns.

Q1: The paper does not compare against scGPT‑spatial, which is a relevant pre-trained foundation model for spatial transcriptomics.

RQ1: We appreciate the reviewer's valuable suggestion. We fully agree that comparing our method with scGPT-spatial would strengthen our manuscript and provide valuable insights.

Indeed, scGPT-spatial represents an important advancement with its continual pretraining and innovative architecture tailored to spatial transcriptomic data. Regrettably, only limited zero-shot inference demo is publicly available, which prevents us from conducting a comprehensive evaluation. We have reached out to the authors to inquire about the availability of the full codebase. Upon its release, we will promptly integrate detailed comparative analyses into our revised manuscript.

Q2: Runtime and computational cost are not discussed, which is important given the use of large pretrained models and multi-view graphs.

RQ2: Thanks for your suggestions. To address this concern, we provide both a theoretical complexity analysis and empirical runtime measurements.

(1) Theoretical complexity analysis. The computational complexity of SAINT is mainly determined by the multi-view GCN layers, where each branch processes a feature matrix $X \in \mathbb{R}^{N \times F}$ with adjacency $A \in \mathbb{R}^{N \times N}$, resulting in $O(3|E|F)$, where $N$ is the number of spots, $F$ the feature dimension, and $|E|$ the number of edges. The sequence embedding branch only adds a lightweight projection from precomputed embeddings of dimension $d_s$ to $d_p$ with complexity $O(N d_s d_p)$, and the ZINB decoder adds a negligible $O(N d_p)$ term. Thus, the total complexity remains $O(|E|F)$, comparable to GCN-based methods like MAFN, with only a small linear overhead from the sequence branch.

(2) Empirical runtime measurements. We further measured the actual runtime on DLPFC slice 151507, using MAFN and SAINT variants under the same hardware environment (Intel Core i9-9900K CPU, 64GB RAM, with an NVIDIA RTX 3090Ti GPU). The results are summarized in the table.

These results show that the sequence-aware variants of SAINT incur only a marginal increase in runtime (less than 1% compared to MAFN), demonstrating that the integration of nucleotide-derived features introduces negligible computational overhead.

Q3: The graph construction details (e.g., k-NN, radius, hyperparameters) are not clearly specified.

RQ3: Thanks for your suggestions. In our implementation, the spatial graph is built by connecting each spot to its $k = 14$ nearest neighbors based on Euclidean distances, with an additional radius threshold of $r = 560$ pixels to restrict edges to local neighborhoods. For training, we set the number of epochs to 250, the learning rate to 0.001, and the weight decay to $5 \times 10^{-4}$. The hidden dimensions of the two GCN layers are 128 and 64, respectively, and dropout is set to 0. We will explicitly describe these parameters in the revised manuscript to ensure clarity and reproducibility.

Q4: No ablation isolating the three graph types (spatial, feature, combined) to assess their individual contributions.

RQ4: Thanks for your valuable advice. To address this concern, we performed an ablation study isolating the spatial, feature, and combined graphs, and compared these settings with the full SAINT-SA model on three datasets (DLPFC-151507, HBC, and MBA). The results, reported in terms of ARI and NMI, are summarized in the table below.

These results reveal several insights. Firstly, using only the feature graph yields the weakest performance across all datasets, indicating that expression similarity alone is insufficient to capture spatial domains. Secondly, the spatial graph consistently performs better, confirming the importance of local spatial connectivity. Thirdly, combining the two graphs further improves ARI and NMI, demonstrating that integrating both spatial and feature information is beneficial. Finally, the full SAINT-SA model, which incorporates sequence-aware features in addition to the combined graph, achieves the highest performance on all datasets.

Q5: No ablation studying the impact of each loss term.

RQ5: We appreciate the reviewer’s request. To address this, we conducted an ablation study where each loss component, including the ZINB reconstruction loss, the graph regularization loss, and the dcir contrastive loss, was removed individually while keeping the others unchanged. The ARI and NMI results on the DLPFC-151507, HBC, and MBA datasets are summarized below.

The ablation results demonstrate that removing any loss term reduces performance, confirming that all three components are essential. In particular, excluding the ZINB loss causes a clear drop in ARI, highlighting its role in reconstructing gene expression distributions. Omitting the graph regularization term weakens the capture of local structures, while removing the dcir loss reduces cross-view consistency. The full SAINT-SA model, integrating all losses, consistently achieves the best ARI and NMI, demonstrating their complementary contributions.

Q6: The meaning of $d_1$ and $d_2$ in Figure 4 is unclear and should be explained. It’s ambiguous whether w. SA in Figure 3 and SAINT-G in Table 1 refer to the same variant.

RQ6: Thanks for your valuable advice. Regarding the first point, in Figure 4 $d_1$ corresponds to the embedding dimension used in SAINT-G, where gene-level sequence embeddings are averaged, while $d_2$ corresponds to the dimension used in SAINT-SA, where expression-aware attention is applied during aggregation.

For the second point, “w. SA” in Figure 3 and “SAINT-G” in Table 1 refer to different variants. The former is a simplified version with sequence embeddings but without the full fusion design, whereas the latter corresponds to the gene-level averaged sequence embedding variant. We will make this distinction explicit in the revision.

Q7: How does SAINT perform when the number of genes with valid nucleotide sequences is low?

RQ7: Thanks for your question. To address this, we conducted an experiment where we progressively reduced the proportion of genes with valid nucleotide sequences and evaluated the performance of SAINT-SA. The reduction was simulated by randomly masking a fraction of the sequence embeddings, as shown in the table below.

The results show that although performance decreases as fewer valid sequences are available, SAINT-SA remains competitive even with only 20% of genes retaining sequence information. When sequence data are entirely removed, its performance approaches that of the non-sequence variant, demonstrating the model’s robustness to missing nucleotide information.

Q8: Does the method degrade gracefully?

RQ8: To investigate whether our method degrades gracefully, we tracked the performance of both SAINT-G and SAINT-SA during training, as shown in the table below.

The results show a larger drop in the early stage (up to 50 epochs), followed by only minor decreases as training continues, and eventually stabilizing at epoch 200–250. This consistent trend demonstrates that the proposed method degrades smoothly over time rather than experiencing abrupt performance collapses.

Q9: How exactly is the spatial graph constructed? Is it k-NN in spatial coordinates, a radius threshold, or another method? What specific hyperparameters are used (e.g., k)?

RQ9: Thanks for your question. In our framework, the spatial graph is built using a radius-based neighbor search, connecting spots whose Euclidean distance is below a predefined threshold. We set the radius to 560 pixels to ensure each spot has a reasonable number of neighbors without creating overly dense connections. The adjacency matrix is then symmetrized to form an undirected graph. We will clarify this construction and its hyperparameters in the revised manuscript.

References.

[1] Wang, Chloe, et al. "scGPT-spatial: Continual pretraining of single-cell foundation model for spatial transcriptomics." bioRxiv (2025).