Identify a CNV inference method to use for annotating tumor and normal cells

[jashapiro]

If we can't do a great job with cell types, what can we do with assigning tumor/normal status?

Especially inferring CNV status.

[allyhawkins]

Since we have decided to include tumor/ normal designations on the portal, I'm updating this issue to reflect the next steps. To label cells as tumor cells, we need to identify a method we like for CNV inference. We will probably want to use CopyKAT for this, mainly because the other methods I looked at require a reference of tumor/normal, which we will not have. I think a good place to start is to use CopyKAT with some of the RMS libraries and compare those annotations to the submitter-provided annotations.

[sjspielman]

Noting from the RMS paper, which used inferCNV

Malignant cells were confirmed in patient tumor data by inference of copy-number variation using inferCNV version 1.1.3 of the TrinityCTAT Project (https://github.com/broadinstitute/infercnv). An average read count per gene cutoff of 0.1 was used. Reference cells were defined by those cell clusters in the Seurat analysis that expressed markers of hematopoietic or endothelial cells. Analysis included denoising, implementation of inferCNV’s i6 HMM model, and subclustering using quantile normalization.

Edit - should also note here that inferCNV, like most methods out there, requires a normal reference which we generally do not have! copyKAT does not have this requirement, hence our preference.

[sjspielman]

Just sharing a highly initial result for one of the RMS datasets..

> table(sce$celltype, sce$copykat_prediction)
                       aneuploid diploid not.defined
  Fibroblast                  30      32           0
  Lymphocyte                  29      88          10
  Monocyte                    93      96           6
  Neuron                       8      81           0
  Tumor_Mesoderm              83     137           8
  Tumor_Myoblast-A            10      65           0
  Tumor_Myoblast-B             6      32           4
  Tumor_Myoblast-C           213    3196          66
  Tumor_Myoblast-D             8     109           8
  Tumor_Myocyte-A             57     483           5
  Tumor_Myocyte-B             49     583           6
  Vascular Endothelium       179      87           3

And at a more birds-eye view, where submitter is condensed to just tumor/normal for "known" RMS cell types.

  submitter copykat         n
  <chr>     <chr>       <int>
1 normal    aneuploid     339
2 normal    diploid       384
3 normal    not.defined    19
4 tumor     aneuploid     426
5 tumor     diploid      4605
6 tumor     not.defined    97
7 NA        aneuploid      14
8 NA        diploid        60
9 NA        not.defined     4

[allyhawkins]

What parameters did you use to generate these results? I think the one I'm specifically interested in is the distance parameter based on this paragraph from the readme.

One struggling observation is that none of one clustering method could fit all datasets. In this version, I add a distance parameters for clustering that include "euclidean" distance and correlational distance, ie. 1-"pearson" and "spearman" similarity. In general, corretional distances tend to favor noisy data, while euclidean distance tends to favor data with larger CN segments.

It looks like, the example there uses euclidean, but based on this statement we probably want to use correlational. I would expect the pediatric tumors to be noisier because of their quiet genomes.

[sjspielman]

This was my first run through of copyKAT to see how it all works using defaults (including euclidian) and it took several hours so I am going to be running this today on the cluster with more cores for hopefully better runtimes 🤞 . I was actually planning on looking at correlational today too, based on this README paragraph! We're on the same page 😎

[sjspielman]

I want to note here also that I looked a bit into SCEVAN, but got persistent (and different!) errors when trying to run it on a couple different SCEs, including a few from ScPCA & HCA. For example one such error...

Error in data.frame(..., check.names = FALSE) : 
  arguments imply differing number of rows: 7565, 7568

The code itself has some aspects which are a non-standard for R packages (such as, loading libraries in package functions), but since the DESCRIPTION file doesn't specify any package versions (and is missing some remotes..) I wondered if maybe my environment was not compatible. They do offer a docker container, which I pulled down to try out instead. Sadly, when I went try out the same libraries through in docker, the same errors persisted all around.

I also considered* (grammar edit) whether we might look into ikarus (a python library), but I don't think it's well-suited for our data since its ML model was trained on a bunch of adult cancers. A set of genes identified from these adult cancers are used to support calling tumor/normal, and this gene set is probably not biologically meaningful for our pediatric data. So, I did not look into this library further b/c I assume a lot of mis-specification.

[sjspielman]

Jotting down here a couple notes, where all bold in quotes was added by me:

• I had missed this little note all the way at the end of the copyKat README

Final final note, CopyKAT had difficulty in predicting tumor and normal cells in the cases of pediatric and liquid tumors that have a few CNAs. CopyKAT provides two ways to bypass this to give certain output instead of being dead staright: 1) input a vector of cell names of known normal cells from the same dataset 2) or try to search for T cells.

Basically, they are suggesting we either a) know a priori some subset of normal cells (which IRL we probably will not know!), or b) perform cell-typing, identify confidence T-cells (probably as an "easier" cell population to label with confidence?), and provide those to copyKat as a diploid baseline.

• From the copyKat paper:

To do this, we pool single cells into several small hierarchical clusters and estimate the vari- ance of each cluster using a Gaussian mixture model (GMM) (Fig. 1b). The cluster with minimal estimated variance is defined as the ‘confident diploid cells’ by following a strict classification criterion. Potential misclassifications may occur when the data have only a few normal cells or when the tumor cells have near-diploid genomes with limited copy number aberration (CNA) events. In this case, CopyKAT provides a ‘GMM definition’ mode to identify the diploid normal cells one by one, where a mixture of three Gaussian models of gene expression in single cells is assumed to represent genomic gains, losses and neutral states. A single cell is then defined as a confident diploid cell when genes in neutral states account for at least 99% of the expressed genes.

I can't find anything about this "mode" in the actual package though, so I don't know how to invoke this approach which might be better for our data! Perhaps I'll open an issue over in their repo. Edit - I did the thing navinlabcode/copykat#100

• I did a bit of exploration as well for whether there's anything we can get out of the CNA matrix (see Explore copyKat performance on RMS data #230 (comment)). It turns out, maybe? I don't have anything really definitive on this, hence posting as a comment in this issue. I compared the CNA values between one of our RMS results and the test result that copyKat provides. I've plotted, for each of the first 45 genomic bins (genome is split into 220KB bins), the CNA distributions between all cells that are eventually labeled diploid vs. aneuploid. I would expect to see some separation between these distributions! Indeed, we see separation in many of the genomic bins for the test data result, but not for our RMS data. This is highly cursory but suggestive of a potential avenue for defining confidence in some way.

• Related to above, clustering these CNA results is, in part, how copyKat identifies diploid/aneuploid. Nothing is really preventing us from clustering these numbers ourselves?

[allyhawkins]

Basically, they are suggesting we either a) know a priori some subset of normal cells (which IRL we probably will not know!), or b) perform cell-typing, identify confidence T-cells (probably as an "easier" cell population to label with confidence?), and provide those to copyKat as a diploid baseline.

I just wanted to note that the idea of labeling T-cells is an interesting one. However, the very obvious caveat that I can think of is when T-cells aren't in the sample.... which is likely. Also, what about the other normal cells? Would copyKat then automatically label these as non-normal?

Another thing I've been thinking about yesterday and today is the fact that we don't really know how all of these samples were prepped. It's likely that there was sorting involved for many of them and if the sorting was specific for tumor cells, then I wouldn't expect this to work. Without a true normal population present, I would think copyKat would break or not work as well. Just some food for thought as we think about CNV information. How do we know that we aren't just dealing with all tumor cells for samples?

From the copyKat paper:
To do this, we pool single cells into several small hierarchical clusters and estimate the vari- ance of each cluster using a Gaussian mixture model (GMM) (Fig. 1b). The cluster with minimal estimated variance is defined as the ‘confident diploid cells’ by following a strict classification criterion. Potential misclassifications may occur when the data have only a few normal cells or when the tumor cells have near-diploid genomes with limited copy number aberration (CNA) events. In this case, CopyKAT provides a ‘GMM definition’ mode to identify the diploid normal cells one by one, where a mixture of three Gaussian models of gene expression in single cells is assumed to represent genomic gains, losses and neutral states. A single cell is then defined as a confident diploid cell when genes in neutral states account for at least 99% of the expressed genes.

I think this statement in part addresses that concern I have about sample prep. But also, for many of these tumors the proportion of non-tumor cells is going to be low? Does that mean there aren't enough "normal" cells for copyKat to properly classify and assign the normal population?

Just some Friday afternoon brain dumping....

[sjspielman]

Noting #230 has been merged where we explore copyKat performance, but I do not think that this accomplished our goal of identifying a CNV inference method to use. So, we'll keep this open for now, but may want to break out a new issue to track next steps more precisely.

[jaclyn-taroni]

I'm removing this from the epic it's currently in!

[MagpiePKU]

@sjspielman did you had any success in forcing copyKAT to consider some cells as diploid? Thanks.

[sjspielman]

HI @MagpiePKU, what you see in this issue from last year is the extent of the analysis/exploration we performed with copyKAT, and you can find the notebook we wrote for this exploration here. The data used in this notebook is in our private AWS S3 storage, but it can be obtained from our Single Cell Pediatric Cancer Atlas. Please bear in mind that since we are analyzing pediatric tumor genomes which are fairly quiet compared to adult tumor genomes, we did expect it will be even more challenging for methods like copyKAT, which use ploidy, to discriminate between normal and tumor cells. We have not had a chance to explore further methods since. Best wishes with your analyses!