MuData

A MuData (muon, mudata, SCENIC+) holds one AnnData per modality under mod/, beside global elements that span them: a joint clustering in obs, a WNN embedding in obsm, the joint graph in obsp. obsmap and varmap say which global rows each modality has. Both an .h5mu and a MuData Zarr store are read; an output is an .h5mu.

adata convert multiome.h5mu -o multiome.h5ad
adata convert scenicplus.h5mu -o rna.h5ad --mod scRNA_counts
adata convert multiome.h5ad -o multiome.h5mu
adata convert multiome.h5ad -o one.h5mu --mod rna

Flag	Meaning
--mod	Reading: the modality to write (default: each, to OUT.<mod>.<ext>; a file with one modality goes to OUT). Writing: the name to write the whole AnnData under, instead of splitting it
--feature-types-column	Writing: the var column to split by (default feature_types, when there is one)

Reading writes each modality as it is stored, then adds the global parts onto its cells (or features), in the modality’s order, through the maps:

MuData	Each modality’s AnnData
mod/<name>	the AnnData itself: X, layers, obs, var, obsm, varm, obsp, varp, uns, raw
global obs and var columns	obs and var columns, on the modality’s rows
<name>:col (a column pulled from this modality)	col, when the modality has no col of its own
<other>:col (pulled from another modality)	left out, with a note
global obsm, varm	obsm, varm, on the modality’s rows; the per-modality masks are left out
global obsp, varp	obsp, varp, both axes on the modality’s rows
global uns	uns

Where a global element has a name the modality already has, the modality’s own is kept and the global one is left out, with a warning. Without obsmap/varmap (an older writer), rows are matched by name. A map that does not place each of the modality’s rows once is refused before anything is written. Global rows are read a block at a time, sorted for the read when the modality’s order differs; a graph’s entries are remapped as they stream and rebuilt through on-disk buckets, keeping CSR or CSC.

Writing splits var by its feature types, as muon reads a 10x Multiome matrix: Gene Expression becomes rna, Peaks atac, Antibody Capture prot, and any other type a modality named after it (CRISPR Guide Capture → crispr_guide_capture). Without a feature-types column the AnnData is one modality, rna.

AnnData	MuData
X, layers, var, varm, varp, raw	each modality’s, for its features; every matrix read once for all of them
obs	global obs; each modality’s obs holds the cell names only
obsm, obsp, uns	global obsm, obsp, uns
	obsmap, varmap and the obsm/varm masks, root encoding-type: mudata, axis: 0, and the 512-byte MuData header mudata reads first

An obsm entry named like a modality is refused, since a MuData keeps that name for the modality’s mask. Converting the .h5mu back gives each modality with the cells’ annotations and embeddings again.