Please complete a declaration of competing interests, considering the following questions: I would suggest the authors to include more details. Is it possible to combine multiple reference dataset to generate a reference dataset that contains all required cell type and use GEDIT to predict only once instead? Two recent papers can do this:įigure 1, which shows the workflow of GEDIT, is not very informative. Facing this issue, as stated in line 311, the authors predicted proportions multiple times using different references. It is a common problem that no single reference contains all relevant cell types. I would suggest the authors to compare GEDIT with some popular deconvolution methods that use single-cell RNA-seq data as reference, for example, MuSiC, as mentioned in paper. Since more and more single-cell RNA-seq data are becoming available, many researchers are using well-labeled single-cell RNA-seq data as reference for deconvolution. For example, is it due to GEDIT's unique way of gene selection, the step of row scaling or use of non-negative linear regression? In addition to benchmarking data, how do other methods perform on the other 3 datasets? I recommend the author to provide some insights on why GEDIT is better than other methods. One limitation is that the comparison is only using pearson correlation and average error. The analyses show good results in comparison to existing methods on benchmarking experiment. some genes have zero or very little expression in some cell type is also an important and helpful signal for cell type abundance estimation. It seems this model can only utilize the information from positive marker genes and cannot take the advantage of negative marker genes, e.g. GEDIT performs a non-negative linear regression to estimate the cell type abundances. How is this new selection strategy compared to classical highly variable gene selection? Is there any evidence to support that this entropy-based selection has some advantage on highly variable selection over commonly used highly variable gene selection method? GEDIT has its unique way of selecting genes, that is, using a newly-define signature score based on information entropy. These genes are very useful for deconvolution. I think this strict filtering criterion may have excluded some marker genes which are only highly expressed in one cell type.
As stated in line 339 to 340, GEDIT first excludes genes with zero detected expression in half or more of cell types. There are two possible issues that I'm concerned about on the gene selection step in GEDIT: a. The authors should make this clear in the paper. Since cells of different types may have different size and GEDIT does not take this into consideration, I am not sure that the predicted proportion by GEDIT is cell type abundance or molecular abundance. GEDIT is able to utilize data from different platforms as reference, including bulk RNA-seq, microarray, and single-cell RNA-seq. The evaluations in the paper are convincing to some extent and the paper is well-written. It compares with other computational approaches for cell type deconvolution and shows higher accuracies and greater versatility on data from a variety of tissue types and technological platforms. The authors present an algorithm called GEDIT, which uses information from a reference dataset to estimate cell type abundances in a target dataset.
0 kommentar(er)
