Quality Control (QC)

The sum of all Unique Molecular Identifiers detected within a single cell barcode. This metric directly correlates with the cell's capture efficiency and cytoplasmic mRNA content.

• High counts: May indicate a doublet (two cells captured in one droplet) or a genuinely large, transcriptionally active cell like a hepatocyte.
• Low counts: Typically represent empty droplets containing only ambient RNA, or damaged cells with ruptured membranes that have leaked their cytoplasm.
• Filtering strategy: Set an upper and lower threshold based on the distribution. A typical lower bound removes the 5th percentile, while the upper bound removes outliers beyond 3 median absolute deviations (MADs).
• Example: In a PBMC dataset, viable cells might have 500–5,000 UMIs, while empty droplets cluster below 100 UMIs.

The count of unique genes with at least one UMI mapped. This metric identifies cells with low-complexity libraries, which are often stressed, dying, or stripped of their mRNA.

• Low gene count with normal UMI count: Suggests a cell dominated by a few highly-expressed transcripts, often a sign of cell damage or a specialized cell type like a red blood cell precursor.
• High gene count: Expected for large, complex cells like neurons or dendritic cells.
• Filtering strategy: Apply a lower threshold (e.g., >200 genes for human cells) to remove debris. An upper threshold can help catch doublets.
• Relationship to UMIs: These two metrics are strongly correlated. Cells that deviate from this correlation—high UMIs but low genes—are prime candidates for removal.

The percentage of UMIs mapping to the mitochondrial genome. A high fraction is the canonical "dead cell" signature.

• Biological basis: When the plasma membrane ruptures during cell death, cytoplasmic mRNA leaks out, but mitochondrial transcripts remain trapped within the organelle. The resulting library is enriched for mitochondrial genes.
• Thresholds: For most human tissues, cells with >15–20% mitochondrial reads are flagged. However, this threshold is tissue-specific—cardiomyocytes and renal proximal tubule cells naturally have higher mitochondrial content.
• Low fraction: An abnormally low fraction can indicate a loss of cytoplasm during library preparation, leaving only the nucleus.
• Example: A stressed hepatocyte sample might show a bimodal distribution, with a healthy peak at 5% and a dying population at 40%.

A computational prediction of whether a single cell barcode actually contains two or more cells. Doublets confound cell-type identification by creating artificial hybrid transcriptomes.

• Detection methods: Tools like Scrublet, DoubletFinder, and solo simulate doublets by averaging random cell pairs and then score each real barcode against this synthetic doublet distribution.
• Impact: Doublets between distinct cell types (e.g., a T cell and a monocyte) create a spurious "intermediate" cluster that can be mistaken for a novel cell state or transitional population.
• Homotypic doublets: Two cells of the same type are nearly impossible to detect computationally and are often tolerated.
• Expected rate: Doublet frequency scales roughly linearly with the number of cells loaded. A 10,000-cell 10x Genomics run typically yields 5–8% doublets.

The proportion of reads mapping to ribosomal protein genes (e.g., RPL, RPS families). This metric serves as a quality complement to mitochondrial fraction.

• High ribosomal fraction: Often indicates high translational activity in metabolically active cells, but can also spike in certain stress responses or technical artifacts.
• Low ribosomal fraction: A sudden drop in ribosomal reads, especially when paired with high mitochondrial reads, is a strong indicator of RNA degradation and cell death.
• Filtering strategy: This metric is typically used in conjunction with mitochondrial fraction. Cells with low ribosomal AND high mitochondrial content are removed.
• Tissue specificity: Stem cells and cancer cells often have naturally elevated ribosomal content, requiring adjusted thresholds.

Background counts from cell-free mRNA floating in the cell suspension that get encapsulated into droplets alongside—or instead of—a cell. This creates a low-level transcriptional "soup" that contaminates all barcodes.

• Source: Cells lyse during tissue dissociation, releasing their mRNA into the suspension medium.
• Detection: Tools like SoupX and CellBender model the ambient RNA profile from empty droplets and subtract it from cell-containing barcodes.
• Consequence: Without correction, ambient RNA can create false-positive signals for highly-expressed genes (e.g., hemoglobin in blood samples) in cell types that don't actually express them.
• Correction: This is a post-hoc computational step applied after initial QC filtering, distinct from the per-cell metric thresholds.