Single-cell RNA-seq (scRNA-seq) changed biology by replacing the population average with the individual cell. Instead of one transcriptome per sample, you get thousands — each one a snapshot of a cell’s identity and state.
How droplet-based scRNA-seq works
The dominant platform (10x Genomics Chromium) uses microfluidics to encapsulate single cells in oil droplets along with a barcoded gel bead. Each droplet contains:
- One cell
- One bead carrying ~750,000 oligonucleotides with a unique cell barcode
- Reverse transcription reagents
Inside the droplet, mRNA from the cell hybridizes to the bead’s poly-T sequence and is reverse-transcribed. Each transcript receives the cell barcode plus a unique molecular identifier (UMI). After droplet breaking, all cDNA is amplified and sequenced together. The barcodes let you assign each read back to its cell of origin.
The data structure
The output is a gene-by-cell count matrix: rows are genes, columns are cells, values are UMI counts. A typical experiment captures 5,000–20,000 cells with 1,000–10,000 genes detected per cell. Most genes are zero in any given cell — this sparsity is normal.
The standard analysis pipeline
1. Quality control
Filter low-quality cells: those with too few or too many genes (suggests empty droplets or doublets), excessive mitochondrial RNA percentage (suggests dying cells, typically >20% in human, >10% in mouse). Filter genes detected in fewer than ~3 cells.
2. Normalization
Counts are scaled to account for differences in sequencing depth between cells. Methods: log-normalization (Seurat default), SCTransform, or scran’s deconvolution.
3. Feature selection
Identify the most variable genes (typically top 2,000) — these drive biological variation more than housekeeping genes.
4. Dimensionality reduction
Principal component analysis (PCA) reduces 2,000 genes to ~50 components. Then UMAP or t-SNE projects to 2D for visualization.
5. Clustering
Graph-based clustering (Leiden or Louvain algorithms) groups cells with similar transcriptomes. Each cluster typically corresponds to a cell type or state.
6. Cluster annotation
Identify marker genes for each cluster (genes specifically up-regulated in one cluster vs others) and compare to known cell type markers. Tools: SingleR, CellTypist, scType, manual annotation with gene panels.
7. Differential expression and pathway analysis
Compare conditions within or between cell types. Use specialized tests that account for sparsity (MAST, glmGamPoi).
Common applications
- Cell atlas building: Comprehensive maps of organs and tissues
- Tumor heterogeneity: Identify rare resistant clones, characterize immune infiltrate
- Developmental trajectories: Pseudotime analysis reconstructs differentiation paths
- Immune profiling: CITE-seq combines transcriptome with surface protein measurement
- Drug response: Identify which cell populations respond to a treatment
Common tools
- Cell Ranger: 10x Genomics’ processing pipeline (alignment, counting, basic QC)
- Seurat (R): Most widely used analysis package
- Scanpy (Python): Alternative for Python users; scales better to very large datasets
- Bioconductor SingleCellExperiment: Object framework for R-based pipelines
- scvi-tools: Deep learning-based methods for normalization, integration, batch correction
Caveats and pitfalls
- Dropout: Many genes are not detected in cells where they’re expressed
- Doublets: 1–10% of “cells” are actually two cells captured together
- Dissociation bias: Some cell types are more fragile and underrepresented
- Batch effects: Major source of artifacts; integration methods (Harmony, scVI, MNN) are essential when combining datasets
What about single-nucleus RNA-seq?
For tissues that don’t dissociate well (brain, heart, archived samples), snRNA-seq sequences nuclei instead of whole cells. It captures only nuclear RNA but works on flash-frozen tissue.
scRNA-seq is now standard. Most journals expect it where heterogeneity matters, and toolchains are mature enough that a small lab can run analyses end-to-end.
