05. Filtering Empty Droplets and Low-Quality Cells

Single-cell quality control (QC) is not simply deleting rows with small numbers. First ask whether a cell barcode, the address of a candidate cell, represents a captured cell, then ask whether that cell’s expression profile is usable.

Question for this lesson: Which barcodes can be used as expression profiles of individual captured cells?

A unique molecular identifier count (UMI count) is the number of original RNA molecules remaining after duplicates from polymerase chain reaction (PCR) amplification are collapsed. Mitochondrial fraction is the fraction of all counts assigned to mitochondrial genes. A doublet is two cells captured in one droplet, and ambient RNA is RNA released from broken cells into the surrounding solution.

Separate barcode, cell, and sample checks

unit	question	representative signals
barcode	Does the droplet contain a cell?	total UMI and difference from the empty-droplet RNA pattern
cell	Is this one usable cell profile?	detected genes, mitochondrial fraction, doublet likelihood
sample	Did the experiment work as expected?	recovered cells, genes detected per cell, valid-barcode rate

These levels point to different causes. A run with too few cells is not the same problem as a run with enough cells but few detected genes per cell.

1. Separate empty droplets from cells

Droplets without cells can contain RNA released into the solution. Their barcodes have nonzero counts but do not represent one cell.

A curve that sorts barcodes by total UMI is called a barcode-rank plot. The high-count beginning is enriched for cells, and the low-count tail for empty droplets. The boundary is not always a clean cliff.

Normal low-RNA cells may sit near the background. EmptyDrops estimates the ambient expression profile from low-count droplets and identifies barcodes that differ significantly from it.

2. Read three per-cell metrics together

metric	very low may suggest	very high may suggest
total UMI	empty droplet, damage, poor capture	doublet or a naturally RNA-rich cell
detected genes	little usable information	doublet or a complex normal cell
mitochondrial fraction	not a problem by itself	loss of cytoplasmic RNA from damage

Absolute thresholds vary by tissue and protocol. Applying one UMI upper bound to lymphocytes and large tumour cells can remove large valid cells as doublets.

Plot sample-specific distributions and inspect cell-type characteristic genes, or markers, near the extremes. Record thresholds chosen from the observed data and experiment context.

3. Doublets create synthetic rows

Two cells in one droplet share a cell barcode. Their RNA becomes one row and may resemble a nonexistent intermediate state.

Scrublet adds pairs of observed cells to simulate doublets, then scores how much each real row resembles those synthetic profiles. The score is supporting evidence, not a self-contained verdict. Check total UMI, incompatible marker combinations, and position among other cell groups together.

4. Ambient RNA also contaminates occupied droplets

RNA released from broken cells can enter droplets that contain real cells. A gene strongly expressed by one cell type may then appear weakly across many cell groups, or clusters.

SoupX estimates the background profile from empty droplets and models contamination in cell rows. Preserve matrices before and after correction and verify whether marker specificity improves. Weak true expression should not be assumed to be contamination automatically.

Record every filtering transition

raw barcodes
  -> barcodes selected as real-cell candidates
  -> cells after quality filtering
  -> cells after doublet filtering

Record the tool, version, metric, threshold, and removed count at every transition. A final Uniform Manifold Approximation and Projection (UMAP), a two-dimensional map of cells, cannot show whether a rare population was absent or removed during QC.

See Single-Cell QC Metrics for field-level definitions and a reporting checklist.

Summary

Separate empty droplets from cell-associated barcodes first. Then inspect per-cell UMI, gene count, mitochondrial fraction, doublets, and ambient RNA. Thresholds must follow sample distributions and tissue biology rather than one universal recipe.

Next, Grouping Similar Cells into Clusters transforms the QC-approved matrix and builds a neighbour graph.

---

Sources

• Cell calling from ambient profiles: Lun et al., Genome Biology (2019)
• SoupX ambient correction: Young & Behjati, GigaScience (2020)
• Scrublet doublet detection: Wolock et al., Cell Systems (2019)
• Single-cell QC best practices: Heumos et al., Nature Reviews Genetics (2023)