How to design a single-cell experiment so the analysis succeeds
Plan scRNA-seq studies that succeed: biological replicates, cells per sample, sequencing depth, cells vs nuclei, multiplexing and batch layout.
Key takeaways
- Biological replicates determine statistical power; extra cells per sample cannot replace them.
- Spread conditions across every capture batch; never process one condition per day.
- Sample multiplexing (hashing or genotype) cuts cost per sample and makes balanced designs easier.
- Talk to the analyst before you load the chip. Many analysis problems are design problems.
Most of the single-cell analyses we rescue were limited by decisions made before sequencing. A clever analysis cannot add replicates that were never collected or separate a condition from the day it was processed. Here is how we advise labs to plan, so the data can answer the question.
Start from the question
Write down the comparison that matters, for example "macrophage expression in treated vs untreated tumors", and the claim you hope to make. That decides almost everything else:
- Atlas or discovery questions (what cell types exist?) need breadth: more cells, more tissue regions, fewer conditions.
- Comparison questions (what changes with treatment?) need replicates: more independent donors or animals per condition.
- Dynamics questions (how do cells differentiate?) need time points, and sometimes lineage tracing.
Replicates beat cells
The statistical unit for condition comparisons is the biological replicate. Ten thousand cells from one mouse per condition support a description, not a statistical comparison. Pseudobulk benchmarks (Squair and colleagues) make this explicit.
Our rules of thumb:
- Minimum: 3 independent biological replicates per condition. Better: 4 to 6.
- Human cohorts: more, often 6 to 10 or more per group, because people vary more than inbred mice.
- Cells per sample: 3,000 to 10,000 recovered cells is usually plenty for major types. Target more only if a rare population (under 1%) is central to the question.
If budget forces a trade-off, cut cells per sample before cutting replicates.
Design batches so they can be removed
Every capture date, chip, reagent lot and sequencing run is a potential batch. The goal is not to avoid batches but to make them orthogonal to condition:
- Process at least one sample from every condition in every batch.
- Randomize processing order within a day.
- Record everything: dissection time, time to processing, operator, kit lot, cell viability.
When we see a design where all controls were processed one week and all treated samples the next, we can describe the data. We cannot cleanly test the treatment.
Use sample multiplexing
Cell hashing (TotalSeq hashtags, CellPlex), lipid-tagged barcodes (MULTI-seq) and genotype-based demultiplexing (souporcell, vireo for genetically distinct donors) let you pool samples in one lane. Pooling:
- puts several conditions in the same batch, so batch effects cannot be confounded with condition,
- identifies cross-sample doublets directly, and
- lowers cost per sample, which buys more replicates.
Cells or nuclei?
| Single cells | Single nuclei | |
|---|---|---|
| Sample type | Fresh, dissociable | Frozen, archived, hard to dissociate |
| Transcripts | Cytoplasmic and nuclear | Mostly nuclear, unspliced |
| Dissociation stress | Can induce stress genes | Avoided |
| Ambient RNA | Moderate | Often higher |
| Multiome compatible | No | Yes (RNA + ATAC) |
Sequencing depth
For 10x 3' gene expression, 20,000 to 50,000 reads per cell is typical. For snATAC-seq, we suggest at least 25,000 read pairs per nucleus. Check sequencing saturation in the Cell Ranger summary. If it is low and the question involves subtle expression changes, resequencing the same library is often cheaper than a new experiment.
Controls and references
- Include a reference or healthy control that matches tissue, age and handling.
- For tumors or disease, plan for matched public atlases to support annotation.
- For spatial follow-up, keep serial sections or blocks from the same samples. See our spatial platform comparison.
Talk to the analyst first
A thirty-minute design conversation costs nothing and can save an experiment. Our free scoping calls regularly cover replicate numbers, batch layout, multiplexing and depth before any money is spent on sequencing. When the data arrives, the same duo team runs the analysis.
Frequently asked questions
How much sequencing depth do I need for scRNA-seq?
For 10x 3' gene expression, 20,000 to 50,000 reads per cell is typical. 20,000 is usually sufficient for cell typing; deeper sequencing helps detect lowly expressed genes and subtle states. Check sequencing saturation in the QC report.
Should I use single cells or single nuclei?
Use cells when tissue can be dissociated quickly and fresh, and when cytoplasmic transcripts matter. Use nuclei for frozen archives, hard-to-dissociate tissues (brain, heart, muscle, adipose) and to avoid dissociation stress. Nuclei need intron-aware counting and stronger ambient correction.
Work with OmicsQuest. Related services:
References
- Squair JW, Gautier M, Kathe C, et al. Confronting false discoveries in single-cell differential expression. Nat Commun. 2021;12:5692.
- Zimmerman KD, Espeland MA, Langefeld CD. A practical solution to pseudoreplication bias in single-cell studies. Nat Commun. 2021;12:738.
- Leek JT, Scharpf RB, Bravo HC, et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat Rev Genet. 2010;11:733-739.
- Zheng GXY, Terry JM, Belgrader P, et al. Massively parallel digital transcriptional profiling of single cells. Nat Commun. 2017;8:14049.
- Heumos L, Schaar AC, Lance C, et al. Best practices for single-cell analysis across modalities. Nat Rev Genet. 2023;24:550-572.