Annotate cell types in desmoplastic small round cell tumor samples on the Portal

[danhtruong]

Proposed analysis

This analysis aims to annotate cell types for the existing desmoplastic small round cell tumor samples on the Portal. These can be found in SCPCP000013.

Specifically, we plan to add three levels of annotations:

The first will annotate cells as tumor cells and normal cells.
The second will annotate the normal cells, providing a distinct cell type (e.g., fibroblasts, T cells).
The last will classify cell states/lineage found in the tumor cells (e.g., epithelial, mesenchymal, neuroendocrine, or muscle)
In addition to providing human-readable names for cell types, we will also provide cell ontology identifiers where possible.

As part of this analysis, we will generate a reference of marker genes associated with desmoplastic small round cell tumor cells and cell states. We will also create a reference dataset containing well-annotated desmoplastic small round cell tumor cells. These references can be used to annotate other samples from desmoplastic small round cell tumor cells.

Scientific goals

The goal of this analysis is to provide validated cell type annotations that can be used for downstream analysis of the desmoplastic small round cell tumor samples. The analysis will accomplish the following goals:

• Provide annotations of normal cells found in the desmoplastic small round cell tumor samples
• Provide annotations of tumor cell states or lineages that may be present in the desmoplastic small round cell tumor samples
• Include cell ontology identifiers for all annotations where possible
• Create a reference of marker genes for desmoplastic small round cell tumor cells and states/lineage
• By providing standardized cell labels across all samples in the project, we can then perform joint analyses on all samples, such as differential expression or gene set enrichment.

Additionally, these annotations can be added to the existing objects available on the ScPCA Portal. This will give users the validated annotations without performing their own cell-type annotations.

Methods or approach

As part of this analysis, we will annotate all desmoplastic small round cell tumor samples and provide a reference that can be used by the community. To do this, we will perform the following steps on some of the samples (2-3) to first create a well-annotated reference dataset for desmoplastic small round cell tumors. Then, we can use this as a reference to annotate cells from all other samples using a reference-based approach like SingleR.

Step 1

Before starting to identify any cell types or cell states, it will be helpful to curate a list of marker genes that are associated with desmoplastic small round cell tumor cells or specific cell states. This may help assign cell types or can be used to validate assigned cell types and cell states. To do this, we will use our pre-existing data and literature to identify markers that can be used to identify:

Desmopastic small round cell tumor cells (e.g., ST6GALNAC5 and CACNA2D2)
Cell lineages (e.g., epithelial, mesenchymal)

Step 2

Next, we must identify which cells in each sample are tumor or normal.
There are multiple ways to do this:

• Use the list of marker genes mentioned in step 1 to identify tumor cells and a separate list of markers for normal cells
• Use a program such as SingleR to identify all cells, where we keep normal cells and re-annotate the tumor cells using the marker genes mentioned in step 1
• Use Seurat function addmodulescore to create a summary metric to identify tumor cells
• Use a program such as CopyKat to predict which cells are tumor or normal using copy number aberrations. This may not be as straightforward as using marker genes, but there are some chromosomal gains and losses associated with desmoplastic small round cell tumors.

In addition, we can perform unsupervised clustering, which should separate the tumor cells and normal cells. Next, we can annotate the clusters based on the marker genes.

Step 3

Next, we want to classify the types of normal cells that are present. Here, we can pull out any cells that are classified as normal cells and use a reference-based method to classify those cells, such as SingleR.

For the reference, we can use a publicly available reference from celldex containing both immune cells and other non-immune stromal cells (fibroblasts, endothelial cells).
The benefit of using celldex is the presence of cell ontology terms included in the reference datasets.

We will then need to validate these findings, which we can do by first curating a list of known markers for the normal cell types in our dataset. Then, we can plot the expression of those known markers across the cell types. We expect that the cells assigned to the cell type where that gene is a marker gene would have the highest expression. One thing to note is that there are some cases where we expect to see correlated expression of multiple marker genes, indicating a specific cell type. In addition, we will try not to subtype the normal cells unless there are verifiable markers.

Step 4

The last thing we will want to do is identify any tumor cells that can be further classified into known cell states or lineages. Desmoplastic small round cell tumors are driven by the pathognomonic EWS::WT1 fusion gene. However, literature has shown tumor cells can display multi-lineage expression, including epithelial, mesenchymal, muscle, and neuroendocrine markers.

Immunohistochemistry has identified markers associated with the different lineages in desmoplastic small round cell tumor specimens. Upon identifying the tumor cells, we will analyze classical markers associated with the lineages. Cells may express markers from only one lineage or multiple lineages. We can perform unsupervised clustering on tumor cells to better separate and identify these cell states.

Existing modules

Yes, this module is based on an existing module of Ewing sarcoma samples in discussion.

Input data

This analysis will use the processed SingleCellExperiment objects for SCPCP000013. Depending on the methods and tools implemented, we may also need to use the processed AnnData objects.

Additionally, we will obtain a reference dataset from the celldex package to annotate normal cells.

Scientific literature

The following papers may be helpful in curating marker gene lists to use for this analysis:

New transcriptional-based insights into the pathogenesis of desmoplastic small round cell tumors (DSRCTs)
Desmoplastic small round cell tumor is dependent on the EWS-WT1 transcription factor
Single-cell multiomics profiling reveals heterogeneous transcriptional programs and microenvironment in DSRCTs
Comprehensive Molecular Profiling of Desmoplastic Small Round Cell Tumor

Other details

Computational resources

All analysis will be done on my local machine if possible.

Timeline

This analysis will be done in stages, where each stage is, at minimum, a single pull request:

Stage 1: Curate lists of marker genes from the literature
Stage 2: Classify cells as tumor vs. normal in a single sample
Stage 3: Classify normal cells in a single sample
Stage 4: Classify tumor cell states in a single sample
Stage 5: Apply this analysis to 2-3 samples to create a reference dataset
Stage 6: Create the code to run SingleR on all remaining desmoplastic small round cell tumor samples using our well-annotated reference dataset

[Jen-OMalley]

Hi @danhtruong. I'm Jen, the Scientific Community Manager at the Data Lab. Thank you for sharing your proposed analysis! Our team is currently reviewing your proposal, and we're looking forward to discussing it with you soon. We'll get back to you with next steps within 3 business days.

We will also be setting up an Amazon Web Services account for you. Once we do, you should receive an email with an invitation to finish setting up your account. I'll reach out again when you should be expecting to see this!

In the meantime, please let me know if you have any questions about OpenScPCA. We look forward to working together!

[Jen-OMalley]

@danhtruong your AWS account has been created and you should receive an email to complete setup. Here are instructions for setting up AWS!

[allyhawkins]

This sounds great @danhtruong and we are excited to have you on board! I had a few questions regarding your analysis plan.

You mention curating a marker gene list and using that marker gene list to annotate the tumor cells. How exactly are you planning on doing that? You mention using Seurat::AddModuleScore() so do you plan on using that or another method?

Based on my experience with the Ewing data, classifying cells solely based on gene expression can be quite challenging, especially if there's no clear separation between tumor and normal cells. However, if you have distinct populations of normal and tumor cells, using a method like AUCell with the set of marker genes for DSRCT could be effective. It's important to note that this method works best with a bimodal distribution of gene set scores, and not as well when the gene set is expressed in most of the cells. UCell is also promising, but it requires you to define a cutoff for cell classification, so some manual investigation will still be necessary.

I was curious if you expect to find normal cells in these samples. I will note that for these particular samples, four of them came directly from the tumor, and the other three are matched PDX samples. My naive expectation is that you will be more likely to see a mix of tumor and normal cells in the samples that came directly from the tumor over the PDX samples, so those are probably the better samples to start with. But I'm not a DSRCT expert, so I'm sure you have a better idea of what to expect than I do!

I did also notice that in Henon et al they looked at the expression of two different DSRCT gene signatures to label tumor cells, and they have a nice supplemental table (Table S2A) of the genes that are differentially expressed in the tumor cells and normal cell types across all 12 patients. Those should be great starting points! That list of DEGs that they provide could potentially be used to create a reference that you might be able to use with a marker gene based assignment method like CellAssign. Note that this is not the most computationally efficient method, but we have found it works well if you have a good reference. Something to think about if you end up having trouble relying on gene expression alone.

I know you mentioned looking at gene expression of marker genes across clusters, which I think is a very reasonable approach. I just want to caution you that if you do plan on using that approach, you will need to spend time first doing the clustering. We do provide clustering assignments in the processed SingleCellExperiment objects, but we use default clustering parameters and don't expect those to work well across all 500 samples on the Portal. Because of that, I think if you want to use clusters we ask that you first perform clustering on the samples you will be working with.

You also mention using CopyKAT or other CNV methods. I was just curious if you have used these with DSRCT samples and had any luck. It looks like there are a small number of recurrent CNVs that you could look for, but they are similar to the Ewing genome, where they are very quiet. Has this worked for you in the past?

I would also encourage you to check out these helpful sections of our documentation to get started:

Technical setup
Contributing to Analysis

Please let me know if you have specific questions on how to get started!

[danhtruong]

This sounds great @danhtruong and we are excited to have you on board! I had a few questions regarding your analysis plan.

You mention curating a marker gene list and using that marker gene list to annotate the tumor cells. How exactly are you planning on doing that? You mention using Seurat::AddModuleScore() so do you plan on using that or another method?

I will use this as well as explore the genes individually. We have internal DSRCT data and markers that distinguish tumor vs normal. I believe these are the same as Henon et al and others. Unlike Ewing Sarcoma, DSRCT appears to be much more distinct from normal cells.

Based on my experience with the Ewing data, classifying cells solely based on gene expression can be quite challenging, especially if there's no clear separation between tumor and normal cells. However, if you have distinct populations of normal and tumor cells, using a method like AUCell with the set of marker genes for DSRCT could be effective. It's important to note that this method works best with a bimodal distribution of gene set scores, and not as well when the gene set is expressed in most of the cells. UCell is also promising, but it requires you to define a cutoff for cell classification, so some manual investigation will still be necessary.

We can try all methods described, AUCell and UCell.

I was curious if you expect to find normal cells in these samples. I will note that for these particular samples, four of them came directly from the tumor, and the other three are matched PDX samples. My naive expectation is that you will be more likely to see a mix of tumor and normal cells in the samples that came directly from the tumor over the PDX samples, so those are probably the better samples to start with. But I'm not a DSRCT expert, so I'm sure you have a better idea of what to expect than I do!

Our internal samples were mostly DSRCT cells, but we found fibroblasts and immune cells as well. The PDX may or may not have fibroblasts.

I did also notice that in Henon et al they looked at the expression of two different DSRCT gene signatures to label tumor cells, and they have a nice supplemental table (Table S2A) of the genes that are differentially expressed in the tumor cells and normal cell types across all 12 patients. Those should be great starting points! That list of DEGs that they provide could potentially be used to create a reference that you might be able to use with a marker gene based assignment method like CellAssign. Note that this is not the most computationally efficient method, but we have found it works well if you have a good reference. Something to think about if you end up having trouble relying on gene expression alone.

I will compare our DEG in our internal DSRCT data set with Henon et al. I am sure they will be quite similar.

I know you mentioned looking at gene expression of marker genes across clusters, which I think is a very reasonable approach. I just want to caution you that if you do plan on using that approach, you will need to spend time first doing the clustering. We do provide clustering assignments in the processed SingleCellExperiment objects, but we use default clustering parameters and don't expect those to work well across all 500 samples on the Portal. Because of that, I think if you want to use clusters we ask that you first perform clustering on the samples you will be working with.

Thanks. I will use both the default values and may generate our own clustering.

You also mention using CopyKAT or other CNV methods. I was just curious if you have used these with DSRCT samples and had any luck. It looks like there are a small number of recurrent CNVs that you could look for, but they are similar to the Ewing genome, where they are very quiet. Has this worked for you in the past?

Yes. It has worked for us because we had WES, and the inferCNV values were concordant. There are recurrent CNVs as you mention which we will investigate too.

I would also encourage you to check out these helpful sections of our documentation to get started:

Technical setup Contributing to Analysis

Please let me know if you have specific questions on how to get started!

Thanks!

[allyhawkins]

Hi @danhtruong, I wanted to respond to provide some additional guidance as you are ready to get started.

In an effort to keep the analysis across modules uniform and transparent, we ask that you start your analysis with the processed objects available on the ScPCA Portal (_processed.rds or _processed_rna.h5ad). These objects have already undergone removal of empty droplets, filtering of low quality cells, and normalization. See the documentation on the ScPCA Portal for more information about how these objects were processed. This ensures that all analyses in OpenScPCA have the same starting point.

When you are ready to get started, you can download the data using the download-data.py script in the repository, which will download the processed objects by default.

Please follow the below steps to start contributing to the project:

• Follow the technical setup instructions found in the OpenScPCA documentation.
• File an issue to track the initiation of your analysis module.
• Initiate your module and file a pull request.

After you have initiated your module, you will be ready to continue with the rest of the analysis that you proposed. I would recommend that you break up your work into the following steps, where each bullet point would be an issue and at least one subsequent pull request:

• Curate a list of marker genes associated with DSRCT cells
• Identify clusters in the DSRCT samples
• Use the list of marker genes to identify tumor cells
• Run copy number inference to identify tumor cells
• Annotate normal cells
• Perform clustering on tumor cells and identify tumor cell states

For more information on contributing to the project, I recommend you review these sections of the documentation:

• More on analysis modules
• Working with Git
• Scoping and creating pull requests