Single-Cell RNA-seq Quality Control: Cells, Genes, and Mitochondrial Reads

Single cell RNA seq quality control is not about applying universal thresholds. It is about making context aware filtering decisions based on your tissue of origin, dissociation method, and biological question. This guide is for bench scientists and bioinformaticians who want to filter cells and genes with confidence and document their choices transparently. You will learn why mitochondrial read percentage varies across cell types, how to set gene and UMI cutoffs that match your data, and how to report your decisions so that others can reproduce your analysis.

At a Glance

Decision Criteria

Why Universal Thresholds Fail

Standard cutoffs such as “keep cells with fewer than 20% mitochondrial reads” appear in many tutorials from sources like the Galaxy Training Network. These numbers work for common cultured cell lines. They break for tissues with high metabolic activity. For example, a study on human ovarian aging (source [7]) found that granulosa cells naturally express more mitochondrial transcripts because they support follicular development. Applying a 10% mito cutoff would discard healthy cells in that context.

Tissue Specific Considerations

Your filtering thresholds must reflect the biology of your sample. Use these questions to decide:

• Is your tissue hypoxic or metabolically active? Sources [6] and [11] describe conditions where mitochondrial reads are elevated even in viable cells.
• Did you use enzymatic dissociation? Harsh dissociation increases stress and mito reads. Compare with single nucleus protocols such as the minimized bias profiling in source [10].
• What is your expected cell size and complexity? Immune cells are small and have fewer transcripts. Epithelial cells are larger and may have more genes.

Gene and UMI Cutoffs

Do not set gene cutoffs by guesswork. Plot the number of genes per cell against the number of UMIs per cell. Cells with very low gene counts are likely empty droplets. Cells with very high gene counts may be doublets. Use the knee plot of cumulative UMI counts to identify the inflection point. The NCBI Bookshelf provides background on UMI based quantification that supports this approach. Document the inflection point value and the number of cells retained.

Mitochondrial Read Fraction

Mitochondrial reads measure cell health, but they also reflect cell type. Cells that rely on oxidative phosphorylation have higher baseline mito fractions. For example, a study on keloid infiltrating zones (source [6]) reported stromal cells with higher mitochondrial activity than immune cells. Your cutoff should be a percentile based decision, not a fixed number. Examine the bimodal distribution of mito fraction. Set the threshold at the upper tail of the main population, not at an arbitrary constant.

Practical Workflow

Step 1: Generate Raw Counts

Begin with filtered feature barcode matrices from your alignment software. Ensure that you have removed ambient RNA using tools such as EmptyDrops or SoupX. The EMBL-EBI Training resource includes practical modules on preprocessing scRNA seq data that explain ambient RNA correction.

Step 2: Calculate QC Metrics

Compute per cell metrics: number of UMIs, number of genes, and fraction of reads mapped to mitochondrial genes. Also compute per gene metrics: total counts across cells and number of cells expressing the gene. Use libraries from Bioconductor such as scater or scran.

Step 3: Plot and Inspect

Generate violin plots for each metric. Look for outlier populations. Create scatter plots of UMI count vs gene count and color by mito fraction. Identify clusters of cells that deviate from the main cloud. Do not rely on automated outliers alone. Manual inspection is essential. Refer to the workflow described in source [10] for a protocol that includes quality inspection steps.

Step 4: Apply Contextual Thresholds

Set lower bounds for genes and UMIs based on the knee point. Set upper bounds based on the 95th or 99th percentile, not on a universal value. For mito fraction, set the threshold at the inflection point where the density curve drops. Alternatively, use a MAD (median absolute deviation) approach. For example, keep cells within 3 MADs of the median for each metric. Document every cutoff value.

Step 5: Filter and Validate

Remove cells that fail any filter. Then check whether expected cell types remain. For example, if you lose all T cells after filtering mito reads, your threshold may be too strict. Reexamine the biological basis. Source [7] discusses cell type specific stress responses that can mimic poor quality. Iterate until you retain a diverse population consistent with your tissue.

Step 6: Document Your Choices

Write a QC report that includes the following:

• Number of cells before and after filtering.
• Threshold values and the justification for each.
• Number of genes retained (typically genes expressed in at least 3 to 10 cells).
• Software versions and parameters.
• Any doublet removal steps.

This documentation is critical for reproducibility. The Galaxy Training Network offers templates for tracking analysis steps.

Common Mistakes

Overfiltering Healthy Cells

The most frequent error is applying a strict mitochondrial filter to tissues with high oxidative metabolism. Cells from sources [6] and [11] show that mito reads can reach 30% in viable cells. Remove these cells and you remove the biological signal you want to study.

Ignoring Doublet Detection

Gene and UMI filters alone do not remove doublets. Doublets look like large healthy cells with high gene counts. You must use computational doublet detection. Source [8] describes cell interaction studies that require clean single cell data. Doublets would confound those analyses.

Using the Same Thresholds for Different Libraries

Batch effects are real. Each sequencing run may have different quality distributions. Apply QC per sample or per batch, not globally. Source [3] on the Galaxy Training Network emphasizes per sample quality checks.

Not Checking Ambient RNA

Empty droplets contain ambient RNA that looks like real cells with low UMIs. Prefiltering by UMI count removes many empties, but some can persist. Use dedicated tools before QC filtering.

Limits and Uncertainty

Technical Noise from Droplet Based Methods

Droplet protocols produce variable capture efficiency. Cells with low UMI counts may be real cells with low expression or technical failures. You cannot distinguish them with certainty. Use a conservative lower bound to retain small cells such as lymphocytes.

Biological Heterogeneity Within the Same Sample

Tissues are heterogeneous. A single threshold for gene count may discard a rare cell type with low transcriptional activity. For example, quiescent stem cells express fewer genes than proliferating ones. Evaluate whether your filters disproportionately remove one cluster.

Uncertainty in Mitochondrial Read Origin

High mitochondrial reads can indicate cell stress, but they can also indicate contamination from platelet aggregates or cell fragments. Source [9] in plant single cell studies shows that mitochondrial reads vary by cell type even in healthy tissue. Your filter may inadvertently select for one cell type over another.

Iterative Filtering Is Required

Do not expect to set filters once. After initial filtering and clustering, reexamine the QC metrics per cluster. If one cluster has unusually high mito reads, decide whether it is a dying cell population or a genuine metabolic state. This iterative approach is described in the context of liver and adipose tissue profiling in source [10].

Frequently Asked Questions

Why do some tutorials recommend a 5% mitochondrial cutoff while others say 20%?

The recommended cutoff depends on the tissue and protocol. Cultured cell lines and immune cells from blood often have low mitochondrial reads. Solid tissues with high energy demand contain cells with naturally higher mito fractions. Always inspect your own data distribution before setting a cutoff.

Should I remove cells with very high gene counts?

Not automatically. High gene counts can indicate doublets, but they can also indicate large cells such as macrophages or megakaryocytes. Use doublet detection software to distinguish. If doublet scores are low, keep those cells.

What do I do if my data shows a bimodal distribution in mitochondrial reads?

Examine which cell types fall into each mode. The high mode may represent a genuine cell population such as stressed or metabolically active cells. Consult tissue specific references like source [6] or [7] to see if that pattern is expected.

Is it better to filter cells by gene count or by UMI count?

Both metrics are needed. Gene count reflects complexity. UMI count reflects capture efficiency. A cell with many UMIs but few genes may be a technical artifact. A cell with many genes but few UMIs may be a damaged cell with high ambient contamination. Use both in combination.

References and Further Reading

1. NCBI Bookshelf provides foundational background on single cell sequencing and quality metrics. Use it to understand UMI based counting and droplet technologies.

2. EMBL-EBI Training offers practical modules on scRNA seq preprocessing including ambient RNA removal and QC visualization.

3. Galaxy Training Network has hands on tutorials for filtering cells and genes. The material includes per sample QC guidance.

4. Bioconductor hosts software packages such as scater, scran, and DropletUtils. The documentation includes vignettes on QC.

5. NCBI Sequence Read Archive is the public repository where raw scRNA seq data can be accessed. Use it to find datasets from your tissue of interest.

6. Source [6] discusses mitochondrial activity in stromal and immune cells within keloid tissue. It shows how cell type influences mito read fractions.

7. Source [7] examines mitochondrial gene expression in ovarian granulosa cells. It demonstrates that high mito reads can be biologically meaningful.

8. Source [8] provides methods for cell interaction analysis. It highlights the importance of removing doublets to avoid false interaction signals.

9. Source [9] covers plant single cell transcriptomics. It reports cell type specific heat stress responses including changes in mitochondrial gene expression.

10. Source [10] describes a protocol for minimized bias profiling of liver and adipose tissue. It includes detailed QC steps for single nucleus RNA seq.

11. Source [11] uses bioinformatics to predict mitochondrial mechanisms in COPD. It shows that lung tissue can have elevated mitochondrial reads due to disease.

Related Articles

• Single-Cell RNA-seq Workflow: A Practical Analysis Roadmap
• Cell Type Annotation in Single-Cell RNA-seq: A Layered Evidence Framework
• Batch Effects in RNA-seq: How to Recognize and Plan Around Them
• Pseudobulk Analysis for Single-Cell RNA-seq
• Gene Set Enrichment Analysis for RNA-seq: Choosing and Interpreting Results