Running Xclone on multi-region data

[tpjones15]

Hello,

I was wondering what the best approach would be for multi-region data from the same tumour.

Would it be:

1. run the baf preprocessing separately per sample:

xcltk baf  \
    --label        {sample name}    \
    --sam          {BAM file}       \
    --barcode      {barcode file}   \
    --snpvcf       {genome1K.phase3.SNP_AF5e2.chr1toX.hg38.vcf.gz}  \
    --region       {annotate_genes_hg38_update_20230126.txt}        \
    --outdir       {output folder}          \
    --gmap         {Eagle_v2.4.1/tables/genetic_map_hg38_withX.txt.gz}  \
    --eagle        {Eagle_v2.4.1/eagle}      \
    --paneldir     {1000G_hg38}              \
    --ncores       10

2. Combine sample level adatas into a tumour level object for the RDR module
3. Combine the output of the sample level xcltk runs:

import xclone
data_dir = "xxx/xxx/xxx/"
AD_file = data_dir + "AD.mtx"
DP_file = data_dir + "DP.mtx"
mtx_barcodes_file = data_dir + "barcodes.tsv" # cell barcodes
# use default gene annotation
BAF_adata = xclone.pp.xclonedata([AD_file, DP_file], 'BAF',
                                 mtx_barcodes_file,
                                 genome_mode = "hg19_genes")
BAF_adata = xclone.pp.extra_anno(BAF_adata, anno_file, barcodes_key = "cell",
            cell_anno_key = ["Clone_ID", "Type", "cell_type"], sep = ",")

Doing the above for each sample and then combining the BAF_adata objects ?
4. Run the BAF and RDR modules on these combined objects

Any help on this would be much appreciated

Best,
Tom

[James-Qiao]

Hello Tom,

Thank you for bringing up this issue!

Currently, XClone is not optimized for handling multiple samples. We have done some tests to combine BAM files of multiple samples and follow the normal XClone pipeline, which had reasonable results. The approach that you proposed also makes sense to me, and the two methods should have little difference.

Of course, it's always good to test different approaches. Any further feedback on running XClone on multiple samples is welcome!

Best,
James

[tpjones15]

Hi James,

just coming back to this, sorry for such a long delay.

I went ahead with merging the sample level RDR and BAF objects into a patient level object. (for this case we have 3 tumour regions - which I've split for the purposes of plotting.

The RDR output looks good: (R3 has an additional chr7p gain and is lacking the chr21 gain we also see in matched WES)

The BAF profile follows the same trend - additionally picking up the chr10p CNLOH in R2 and R4 we find in WES.

As does the combined module:

However, running the new releases subclone detection, xclone is unable to find these differences. it only-detects 2 subclones (just the normal and tumour compartments).

If I force 3 subclones - they don't look to be as convincing as the plots above would suggest (again not finding the R3 specific events).

I appreciate this is quite a qualitative approach. I was going to try clustering on the probability matrix itself which I think is what was done in your manuscript? compared to what is done now in the clustering function. Alternatively, do you think the approach you suggest of merging bams and performing everything together once could help here?

Appreciate any feedback or suggestions!
Best,
Tom

[James-Qiao]

Hi Tom,

Thank you for your feedback on running the new releases of XClone!

Currently, the subclonal detection (or clustering) of XClone uses K-means clustering on the RDR, BAF, and combined probability matrix. This method is simple but may fail to detect clonal differences in some cases. Therefore, I would recommend first analyzing the heatmaps pre-clustering, then assign appropriate number of clones to reduce noise while preserving CNA signals. A useful strategy is to assign more number of clones than you have to retain more details, say 5 clones or more in your case. Then it is easy to merge clones that have similar CNAs to obtain the final clones. Indeed, it is quite a qualitative approach and may require some trials.

For the question regarding whether using merged bam files would help, I would say it probably wouldn't. Based on your results, the CNA calling part seems to work well, and only the clustering has some issues. Using merged bam probably won't affect much on this.

Any other feedback is appreciated!

Best,
James