multimodal-single-cell

Multimodal Single-Cell Analysis

Method selection and execution guidance for multi-modal single-cell assays (RNA+protein, RNA+ATAC, 3+ modalities).

Data Container

Always use MuData (.h5mu) as the container for multi-modal data. Each modality is a separate AnnData accessible via mdata.mod['rna'], mdata.mod['prot'], mdata.mod['atac'], etc. Do NOT store multiple modalities in a single AnnData object.

Technology Detection and Method Selection

Assay type?
├── CITE-seq (RNA + surface protein)
│   ├── Joint embedding (default)
│   │   ├── TOTALVI (probabilistic, handles protein background noise, default)
│   │   └── WNN via muon (quick baseline, weighted nearest neighbors)
│   └── Per-modality analysis → process RNA and protein separately, then integrate
│
├── Multiome (RNA + ATAC)
│   ├── Joint embedding (default)
│   │   ├── MultiVI (probabilistic, handles missing modalities, default)
│   │   └── GLUE (graph-linked embedding, best for regulatory inference)
│   │       PREREQUISITE: rna.var needs chrom/chromStart/chromEnd, which
│   │       normally require a GTF — none is available. Check whether the
│   │       data already carries them; if not, report the blocker and
│   │       fall back to MultiVI. See references/scglue-api.md.
│   └── Per-modality analysis → scanpy for RNA, muon.atac for ATAC, then combine
│
├── TEA-seq / DOGMA-seq (RNA + protein + ATAC, 3 modalities)
│   └── WNN via muon (most flexible for >2 modalities)
│       ├── Compute per-modality neighbors
│       ├── mu.pp.neighbors(mdata, key_added="wnn", ...) with multi-modal weights
│       └── Cluster on WNN graph
│
└── Other combinations
    └── WNN via muon (generalizes to any number of modalities)

Per-Modality QC

Each modality has distinct noise characteristics. QC must be run separately before integration.

Modality QC:
├── RNA
│   ├── Standard scRNA-seq QC (MAD-based thresholds)
│   ├── n_genes, total_counts, pct_mito
│   └── Doublet detection (scrublet or SOLO)
│
├── Protein (CITE-seq)
│   ├── Isotype control check: background level from isotype control antibodies
│   ├── Ambient protein correction: DSB normalization or CLR (centered log-ratio)
│   ├── Filter proteins with low detection across cells
│   └── Check for antibody aggregation artifacts (unusually high counts across all proteins)
│
└── ATAC
    ├── TSS enrichment score (>2 acceptable, >5 good)
    ├── Nucleosome signal (<4 good, banding pattern in fragment size distribution)
    ├── Fraction of reads in peaks (FRiP > 0.3)
    ├── Total fragments (>1000)
    └── mu.atac.tl.nucleosome_signal(mdata.mod['atac'])