Compatibility with 10X Multiome data?

[AndiL04]

Hi,

I really like your tool and the tool is really easy to use!

I am currently working on 10X Multiome data, and I am wondering if I need to split the h5 files generated from CellRanger-arc to two matrices as the input for CellBender? I saw it has mentioned in the publication that 10X Multiome assay can be used but I didn't find specific instructions.

I tried it on one sample and the output.log shows:

cellbender:remove-background: Command: cellbender remove-background --cuda --input raw_feature_bc_matrix.h5 --output output.h5 cellbender:remove-background: CellBender 0.3.0 cellbender:remove-background: (Workflow hash 1840fd242a) cellbender:remove-background: 2023-09-15 22:31:33 cellbender:remove-background: Running remove-background cellbender:remove-background: Loading data from raw_feature_bc_matrix.h5 cellbender:remove-background: CellRanger v3 format cellbender:remove-background: Features in dataset: 36601 Gene Expression, 75643 Peaks cellbender:remove-background: Trimming features for inference. cellbender:remove-background: 107369 features have nonzero counts. cellbender:remove-background: Prior on counts for cells is 8583 cellbender:remove-background: Prior on counts for empty droplets is 267 cellbender:remove-background: Excluding 12101 features that are estimated to have <= 0.1 background counts in cells. cellbender:remove-background: Including 95268 features in the analysis. cellbender:remove-background: Trimming barcodes for inference. cellbender:remove-background: Excluding barcodes with counts below 133 cellbender:remove-background: Using 2237 probable cell barcodes, plus an additional 8971 barcodes, and 40753 empty droplets.

With the output (output_filtered.h5), I used scCustomize to read in the data and split the matrix to GEX part and ATAC part. I am not sure if that is the correct way to do.

Really appreciate your help!

Best, Andi

[sjfleming]

Hi @AndiL04 , thanks for your question.

We do think that cellbender can be directly used on multiome data with both gene expression and ATAC data, just the way you've done it. You've done it exactly the way you're supposed to.

A few caveats and things to think about:

• The denoised output for the ATAC data will be "correct" to the extent that ATAC data obeys the cellbender noise model, which includes "ambient" noise and "barcode swapping" noise. While this noise model is pretty comprehensive (we think), we have also not conclusively proven that this is exactly the right noise model for ATAC data. Does ATAC data actually have "ambient" genomic DNA fragments, where chopped up DNA is actually floating in solution outside cells? Or is ATAC data more susceptible to "barcode swapping" (which includes chimera formation during PCR)? Is there some other kind of noise (not part of our model) that affects ATAC data? (Currently we don't think so...)
• Some of our collaborators believe that the ATAC data modality is much less noisy to begin with. Some users think that it's not worth removing noise from the ATAC data, so they just want to run the gene expression data alone. (You can try that by using the --exclude-feature-types Peaks input argument, still giving cellbender the entire raw data file.) See some discussion here, although some of the problems mentioned have already been solved in v0.3.0. #167 #121
• If you run out of GPU memory (because ATAC has so many features), then you can use the --projected-ambient-count-threshold input argument. The default value is 0.1, but larger values will lead to the inclusion of fewer features in the analysis. The meaning of the number is "estimate the noise counts summed over all cells, for each feature... then exclude features (leave them alone) with estimated noise counts less than this value". The idea is that really low-count features don't contribute much noise, and so they can be left alone, i.e. output=input.
• You might also need a lot of CPU memory, depending on how large the input file is. If you see the process get "Killed" after the training part of the log (where it's showing the loss per epoch), then you've probably run out of CPU memory.

[AndiL04]

Thank you so much for the insightful suggestions! Regarding the points 3&4, I was able to run on default setting without errors on a few samples. Now I am testing adding --exclude-feature-types Peaks and compare the results.

Thanks again for the detailed explainations!

[sjfleming]

You're welcome! Yeah if you use --exclude-feature-types Peaks then cellbender will not touch the ATAC data (output = input).

[YiweiNiu]

Hey @AndiL04, did you get a chance to compare the results? I am quite curious about this. Thanks.

[AndiL04]

Hi @YiweiNiu, regarding to our data, the results were similar. We chose to add --exclude-feature-types Peaks argument, limiting the application of CellBender exclusively to the RNA-seq component. After QC, we observed an approximate 15% increase in the number of nuclei compared to the results obtained using the original CellRanger output.

Hope this info help.