Novel Tensor Decomposition-Based Approach for Cell-Type Deconvolution in Visium Datasets with Reference scRNA-Seq Data Containing Multiple Minor Cell Types

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Summary

The paper is good in general. The application of tensor decomposition-based unsupervised feature extraction has been well discussed. The procedure for achieving deconvolution under the proposed method is thorough and precise. However, it will be appropriate to consider the following points:

1. ABSTRACT:

The ABSTRACT fails to mention why the work was being carried out (at least the first two lines should give an idea of why the study is being done).

2. MATERIALS AND METHODS:

There is no justification or reason to support the choice of the Visium data set. Could there have been other possible datasets? Why the Visium dataset?

Granted that cell 2 location and RCTD are the top-ranked methods according to reference [18];

1. Did the study that led to the above conclusion use the Visium dataset?
2. What went into the study for there to be the conclusion that the other methods are inferior?
3. Is the decision not to compare the TD based on the inferiority of some of the methods

or on time constraints? (Lines 293 to 295).

Answers to points 1, 2, and 3 will enrich the manuscript, in my opinion.

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Novel tensor decomposition-based approach for cell type deconvolution in Visium datasets when reference scRNA-seqs include multiple minor cell types  -  is an interesting research paper in the field of gene expression and scRNA. At first, it seems that the materials methods  and results sections are mixed together, but as we delve deeper into the paper, it becomes clear that the research methodology needs to be mentioned in conjunction with the results. However, in the process of explaining the research methodology, it has become the main part of the paper.

Perhaps the weak point of the paper is in the discussion section, where the results are not sufficiently synthesized with those of other researchers. We expected that the four conventional methods mentioned in the introduction would also be compared and synthesized with the results of other researchers in the discussion section.

 

Minor comment:

Abstract: provide more information about methods.

Rewrite abstract, with distinct aim of study.

Keywords: select keywords different from title

L9: provide complete name for abbreviations.

Throughout the paper: verify abbreviations that be clearly defined.

L49: remove  -see the later part of the study –

L50: mention a reference for this sentence!

L56- 60: rewrite as the aim of study. Not conclusion.

L97: remove (for more…) . add the reference  , or cite?

Fig 1: replace action.explanation of regression analysis to the material method section.

Fig 14 : pi chart or pie chart?

L267: mention a reference for criteria of unsuccessfulness based on cell type fraction.

L282-287: replace to the discussion.

L288: discussion could be completed with adding the four conventional methods, their advantage and disadvantage considering your results.

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript presents a tensor decomposition (TD)-based unsupervised feature extraction approach for cell type deconvolution in Visium spatial transcriptomics datasets, particularly when reference scRNA-seq data include multiple minor cell types. The study demonstrates that the proposed TD method successfully integrates multiple Visium datasets and accurately infers cell type fractions under conditions where four established methods—RCTD, SPOTlight, SpaCET, and cell2location—fail. However, the following problems still exist:

1. The Methods section would benefit from a brief theoretical explanation of why TD is more robust to the presence of minor cell types compared to the other methods.

2. A discussion on the computational complexity and scalability of the method for very large datasets would be valuable. Reporting the comparative computational time would be helpful.

3. Please proofread for minor typographical errors.

4. The authors are encouraged to cite and discuss literature from bioinformatics on interactions prediction and network-based analysis.

Author Response

See attached.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript proposes a tensor decomposition (TD)-based unsupervised feature extraction (FE) method for cell-type deconvolution of 10x Visium spatial transcriptomics data, particularly in scenarios where the reference scRNA-seq dataset contains many minor cell types. The authors argue that such reference imbalance causes several standard deconvolution tools to perform poorly.

The TD-based unsupervised FE method applies tensor decomposition (HOSVD) to multiple Visium samples simultaneously to identify the singular vector that captures spatial structure shared across samples. This spatial component is then used to identify spatially informative genes, perform SVD on the filtered gene set, regress Visium expression onto scRNA-seq structure, and ultimately infer cell-type fractions.

During my review of the manuscript, I had several questions and suggestions:

1. If only one Visium sample is available, could this method still be applied for deconvolution? The entire framework appears to depend on constructing a tensor, which ideally requires multiple replicated or anatomically similar Visium slides. If the method cannot be reliably applied to a single Visium sample, please state this clearly in the Discussion section.
   
   
2. Because no benchmark or ground-truth comparison is provided, it is difficult to evaluate whether the proposed method truly performs well when the reference includes many minor cell types. It would be very helpful to release the code and a runnable demonstration so that readers can reproduce the results presented in the manuscript.
   
   
3. I feel the datasets used to demonstrate the method appear to be highly specific. Have you applied this method to other Visium datasets, and if so, did you observe similar performance? It would also strengthen the manuscript to include comparisons on additional datasets against existing methods such as RCTD and SPOTlight.
   
   
4. Grammar and typos:
1. Page 1 abstract, line 12: “Although it cannot be used in typical cases.” I don’t understand this sentence.
2. Page 2 line 70, visim should be visium 
3. Page 22 line 270: “but all had a have random abundance of cell types”, i don't understand this sentence, are you trying to say “but all had a random abundance of cell types”?
4. Page 8 line 169,170: “coincident with”, do you mean consistent with?

 

Comments on the Quality of English Language

Overall, I encourage the authors to seek professional proofreading assistance, as the manuscript contains many grammatical problems that hinder readability.

Author Response

see attached

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

N/A

Author Response

See attached

Author Response File: Author Response.pdf