single-cell

Single-Cell Analysis

Full pipeline guidance for scRNA-seq, snRNA-seq, and CyTOF (mass cytometry) analysis.

Pipeline Decision Tree

QC and Filtering

QC strategy?
├── Thresholds → Adaptive MAD-based (NOT arbitrary cutoffs)
│   ├── n_genes_by_counts: median +/- 5 MADs
│   ├── total_counts: median +/- 5 MADs
│   ├── pct_counts_mt: median + 3 MADs (upper only; >20% suspicious)
│   └── snRNA-seq: also check pct_counts_ribo (should be low)
├── Doublet detection
│   ├── Single sample → scrublet (sc.pp.scrublet)
│   └── Multiple samples / higher accuracy → SOLO (scvi-tools, deep generative)
└── Ambient RNA (optional, pre-loaded)
    └── If raw + filtered matrices available → SoupX (R via rpy2)
        Estimates contamination from the empty-droplet profile, then corrects
        counts. Report the estimated contamination fraction alongside the
        corrected matrix — a correction whose magnitude goes unstated cannot
        be judged downstream.

Normalization

Data type?
├── Standard scRNA-seq → sc.pp.normalize_total(target_sum=1e4) + sc.pp.log1p
├── Heterogeneous populations (very different sizes) → scran pooling via rpy2
└── Planning to use scVI downstream → skip normalization (scVI takes raw counts)

Highly Variable Genes (HVG)

Input state?
├── Raw counts → flavor='seurat_v3' (variance-stabilizing, works on counts)
└── Log-normalized → flavor='seurat' (default, log-normalized expected)
Always: n_top_genes=2000-3000, subset to HVGs for PCA

Integration / Batch Correction

Batch effects present? (check PCA colored by batch)
├── No batch effect → skip integration
├── Moderate batch effect (same tissue, same protocol)
│   └── Harmony (fast, operates in PCA space, corrects embeddings only)
├── Complex batch effect (different donors, protocols, tissues)
│   └── scVI (deep generative model, corrects latent space, preserves counts)
├── Alignment only (no shared latent space needed)
│   └── scanorama (fast alignment, good for simple batch structures)
└── Cell types already known, and over-correction is the risk
    └── STACAS (R via rpy2) — anchor-based, and can be told which labels
        must NOT be merged, so integration stops collapsing genuinely
        distinct populations into one