Single-Cell RNA-seq Analysis: From Droplets and Matrices to Biological Claims

QC · Normalization · Batch integration · Clustering · Annotation · Trajectory & communication · Validation

Published 2026 · Last updated 2026-09-26 · Hailu Zhou

Single-cell RNA-seq promised to remove the averaging lie of bulk transcriptomics. Instead it replaced one averaging problem with a noisier, richer, more seductive one: every cell is its own experiment, and the analyst must decide which variation is biology, which is technical, and which is an artifact of how the cell was dissociated, captured, and sequenced. A scRNA-seq pipeline is therefore less a fixed recipe than a chain of judgment calls.

Before clustering: is the data worth clustering?

Raw data usually arrives as FASTQ from a 10x Chromium, BD Rhapsody, or similar platform, or as count matrices from a core facility. The first step is not clustering—it is asking whether the data is worth clustering. Empty droplets, dead cells, doublets, mitochondria-heavy lysates, and stress-response programs from dissociation all enter before any biology. Filtering by counts, genes, and mitochondrial fraction is necessary but dangerous: aggressive filtering can delete rare activated immune cells that happen to look stressed. The analyst needs domain knowledge, not just thresholds.

Normalization and feature selection

Normalization is not "log it." scRNA-seq counts are confounded by sequencing depth, gene length (less so than bulk), batch, cell size, and amplification efficiency. Options include log-normalization, sctransform, and platform-specific models. Feature selection—usually highly variable genes—feeds downstream dimensionality reduction. PCA is the workhorse; UMAP and t-SNE are visualizations, not clustering algorithms. A common beginner mistake is treating the UMAP layout as a phylogenetic tree. It is not. Neighborhood graph structure is somewhat real; absolute distances and branch order mostly are not.

Batch integration

Batch integration is where pipelines live or die. Two samples processed on different days will separate by batch even if they are biologically identical. Harmony, Seurat integration, FastMNN, and related methods attempt to remove technical covariance while preserving condition effects. The test is not "the batches mixed." The test is: known cell types align, condition differences remain, marker genes stay differential, and rare populations are not washed away by over-correction.

Clustering and annotation

Clustering uses graph-based community detection (Louvain, Leiden) on a neighbor graph. Resolution is arbitrary: high resolution splits one T-cell state into ten; low resolution merges macrophages and monocytes. There is no universal "correct" cluster number. Good practice is to sweep resolutions, inspect marker genes at each, and merge/split using biology. Marker identification via Wilcoxon, MAST, or AUROC tells you what distinguishes a cluster, not what defines a cell type. Annotation requires reference atlases, literature, and sometimes projection onto a curated dataset.

Downstream analysis

Downstream analysis is where the science happens. Differential expression in single cells needs methods that model zeros, dispersion, and cell-type composition. Trajectory inference (Monocle, slingshot, PAGA) proposes ordered states, not fate proven by time. Cell–cell communication tools (CellChat, CellPhoneDB) infer ligand–receptor enrichment, not demonstrated signaling. Multiomic extensions—CITE-seq, ATAC, spatial—add constraints but also add batch dimensions. The field's biggest trap is algorithmic certainty: a polished UMAP with arrows looks like mechanism, but it is a model of cells, not cells themselves.

What a defensible scRNA-seq manuscript reports

A defensible scRNA-seq manuscript reports: dissociation protocol, viability, capture efficiency, mapped reads, estimated doublet rate, filtering rules, normalization method, integration method and confirmation metrics, clustering resolution, marker statistics, annotation rationale, and validation plan. Ideally, key populations are confirmed by flow sorting, smFISH, IHC, bulk deconvolution, or independent cohort. Without orthogonal validation, "novel cell state" often ages badly.

The deeper skill is restraint

scRNA-seq is extraordinary for generating hypotheses: new fibroblast subsets, tumor ecosystems, developmental bifurcations. It is weaker as standalone proof of function. The best analysts use it to decide which 20 cells to sort, which receptor to block, which marker to validate by flow or IF, which population to test in a mouse. Single-cell data should shrink the experiment space, not replace the bench.

References

  1. Luecken MD, Theis FJ. Current best practices in single-cell RNA-seq analysis: a tutorial. Mol Syst Biol. 2019;15(6):e8746.
  2. Stuart T, Butler A, Hoffman P, et al. Comprehensive integration of single-cell data. Cell. 2019;177(7):1888–1902.
  3. Hao Y, Stuart T, Kowalski MH, et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat Biotechnol. 2024;42(2):293–304.
  4. Korsunsky I, Millard N, Fan J, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat Methods. 2019;16(12):1289–1296.
  5. Heumos L, Schaar AC, Lance C, et al. Best practices for single-cell analysis across modalities. Nat Rev Genet. 2023;24:685–705.
  6. Wolf FA, Angerer P, Theis FJ. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 2018;19(1):15.
  7. Traag VA, Waltman L, van Eck NJ. From Louvain to Leiden: guaranteeing well-connected communities. Sci Rep. 2019;9(1):5233.
  8. Saelens W, Cannoodt R, Todorov H, Saeys Y. A comparison of single-cell trajectory inference methods. Nat Biotechnol. 2019;37(5):547–554.