CellBender

The output of CellBender’s remove-background (output.h5 and output_filtered.h5, 0.2 and later) is recognised by its latent variables, whatever its name. Its matrix is a 10x one and is read as a 10x HDF5 file is; the latents are kept, under the names CellBender’s own loaders give them, so code written against those keeps working. Those loaders name features by symbol; pass --var-names gene_symbols for the same var names.

CellBender	AnnData
denoised counts	X
per-droplet latents (cell_probability, cell_size, background_fraction, droplet_efficiency; latent_* in 0.2)	obs columns, NaN for droplets CellBender did not analyse
which droplets were analysed	obs["cellbender_analyzed"]
gene_expression_encoding (latent_gene_encoding in 0.2)	obsm, same name
ambient_expression	var["ambient_expression"]
which features were analysed (features_analyzed_inds, 0.3+)	var["cellbender_analyzed"]
everything else (fitted constants, learning curve, settings)	uns["cellbender"]

adata convert output.h5 -o cellbender.h5ad
adata convert output.h5 -o cells.h5ad --min-cell-probability 0.5
adata convert output_filtered.h5 -o cb.h5ad --raw raw_feature_bc_matrix.h5

Flag	Meaning
--min-cell-probability	Keep droplets whose cell_probability is at least this; CellBender’s own cutoff is 0.5
--raw	Add the raw Cell Ranger matrix CellBender ran on as layers["cellranger"], and the denoised counts as layers["cellbender"], as load_anndata_from_input_and_output does

The full output.h5 keeps every droplet, as read_10x_h5 would. --raw matches barcodes, so it works for output_filtered.h5 too, whose cells CellBender orders by count; the raw file must list the same features in the same order, and every output barcode, or it is refused. The 10x HDF5 flags apply as well. --from 10x-h5 reads a CellBender file as a plain 10x matrix, without the latents.