How are cell-containing droplets told apart from empty droplets in droplet-based single-cell data?

Why

A droplet run produces far more barcodes than cells. Most barcodes come from droplets with no cell, but those droplets contain ambient RNA, cell-free transcripts released by lysed or stressed cells. Ranking barcodes by total UMI count gives a barcode rank plot.

A good sample shows a steep cliff between cells and background, and 10x Genomics describes a round curve with no steep cliff as a sign of low viability, premature lysis or a wetting failure. The old approach drew a line at the knee of that curve. The EmptyDrops authors point out that a single count threshold discards small cells with low RNA content, because their totals overlap with empty droplets.

EmptyDrops instead estimates the ambient profile from very low count barcodes and tests whether each barcode's profile deviates from it. The Bioconductor book calls cells at an FDR of 0.1 percent. Cell Ranger version 3 introduced automatic cell calling with an algorithm similar to EmptyDrops.

Cell calling does not judge quality. A called droplet can still hold a damaged or dying cell, so ordinary QC on mitochondrial fraction and library size still follows.