Empty Droplets and Ambient Barcodes

Why Empty Droplets Can Have Counts

In droplet-based single-cell assays, a barcode labels material captured in a droplet. Some droplets capture a cell; many do not. A droplet without a cell is not necessarily a barcode with zero recorded counts. RNA released from damaged or lysed cells can remain in the suspension, and fragments of this ambient RNA may enter droplets along with beads and reagents. Sequencing and UMI assignment can then produce a small number of counts for an otherwise empty droplet.

This creates two distinct situations that can look similar if attention is limited to a single barcode: a genuinely empty droplet containing ambient RNA, and a cell-containing droplet with low captured RNA. The observed count matrix records molecules assigned to barcodes and genes; by itself, it does not label the biological contents of each droplet. Count summaries help describe the evidence, but they do not make every barcode's status certain.

Here the counts are UMI counts, consistent with a UMI-based assay. A UMI count is an estimate of distinct captured molecules after deduplication, not a raw read count. In non-UMI assays, read counts can be used descriptively, but their scale and interpretation differ. The matrix structure and count units should be checked before comparing barcodes.

A Simple Model for Background Counts

A useful way to reason about an empty droplet is to imagine that it samples a small number of molecules from the ambient RNA pool. Let \(q_g\) be the proportion of ambient molecules assigned to gene \(g\), with \(q_g\geq 0\) and \(\sum_g q_g=1\). In a simple illustrative model, the count for gene \(g\) in an empty barcode follows a Poisson distribution with mean \(\lambda q_g\), where \(\lambda\) is the expected total ambient UMI count for that barcode. This model assumes a fixed ambient composition, independent Poisson counts across genes, and a stable mean for the barcode being described. Real data can depart from all three assumptions, so the model is an explanatory approximation rather than a universal description.

Under this model, the total count has expected value \(\lambda\), and the probability that gene \(g\) is detected is the probability of observing at least one count. For a Poisson random variable \(Y_{gb}\) with mean \(\lambda q_g\), that probability is \(1-P(Y_{gb}=0)=1-e^{-\lambda q_g}\). Summing the detection indicators across genes gives an expected number of detected genes:

This result explains why ambient barcodes may have multiple detected genes even when their total count is low. If ambient RNA is spread across many genes, a small number of molecules can produce a modest detected-gene count. Conversely, a concentrated ambient profile can yield several counts but only a few detected genes. Neither total counts nor detected genes alone is a definitive cell indicator.

For a cell-containing droplet, counts may include cell-derived RNA as well as ambient RNA. The cell-derived component often increases both \(L_b\) and \(G_b\), but low RNA content, cell damage, capture efficiency, and sequencing depth can keep both summaries small. Thus, the practical question is comparative: how does a barcode's count level and composition relate to the broader set of barcodes in the same library?

Using Count Size and Composition Together

A first descriptive summary is the total \(L_b\). Barcodes with very small totals are more compatible with background-only capture than barcodes with substantially larger totals, all else being equal. A second summary is \(G_b\), which describes how broadly counts are distributed across genes. It is useful to inspect these together: two barcodes can have the same total but differ in whether counts are concentrated in one or two genes or spread across many.

Composition adds another kind of evidence. Let \(A_g\) be an estimated ambient count for gene \(g\), and define the ambient proportion \(\hat q_g=A_g/\sum_h A_h\), provided the denominator is positive. A barcode whose counts are dominated by genes abundant in the ambient pool may be consistent with ambient capture. This comparison is descriptive: a cell can genuinely express common ambient genes, and an ambient profile can include genes that are also expressed by cells.

For a barcode with positive total, its observed gene proportion is \(Y_{gb}/L_b\). These proportions are unstable when \(L_b\) is tiny: one additional UMI can change a proportion substantially. Composition should therefore be interpreted alongside the total and detected-gene count, not as a stand-alone score. It is also important to estimate or compare ambient composition within a relevant library or sample, since cell lysis and sample handling can change which transcripts are abundant in the ambient pool.

Justifying the Detected-Gene Summary

The simple Poisson model also shows why detected-gene counts do not grow in direct proportion to total counts. Under the model, each gene contributes a detection probability \(1-e^{-\lambda q_g}\), which rises with \(\lambda\) but approaches 1 rather than increasing without bound. To see this, differentiate a single contribution with respect to \(\lambda\):

Thus, increasing the expected ambient count cannot decrease the expected number of detected genes under this model, but the marginal increase for a gene becomes small as that gene is almost always detected. This helps explain why the same total can correspond to different gene-detection patterns when the underlying composition differs. It also cautions against treating a fixed ratio of detected genes to UMIs as a universal separator between empty and cell-containing droplets.

Diagnostics, Alternatives, and Limitations

Before interpreting count summaries, check that the matrix is a gene-by-barcode count matrix with the expected count units, that barcode identifiers are not duplicated or mismatched, and that summaries are calculated across the intended genes. Mistakes in matrix structure or gene filtering can change both totals and detected-gene counts. When comparing samples, inspect them separately before pooling: different ambient profiles or sequencing depths can make a common numeric threshold misleading.

Warning signs include a supposed low-count group that contains barcodes with unexpectedly broad gene detection, an ambient profile that changes markedly across subsets of the library, and conclusions that reverse after excluding one or two highly abundant genes. Such patterns suggest that a single summary is inadequate, that the chosen ambient estimate may not represent the barcodes being assessed, or that biological and ambient expression overlap. Low counts can also arise from low-RNA cells or poor capture, while relatively high background can arise when the ambient RNA burden is substantial.

A fixed threshold on \(L_b\) is simple and transparent, but it discards information about composition and can fail when samples differ in depth or RNA content. Joint descriptive review of \(L_b\), \(G_b\), and composition uses more evidence but still requires judgment and does not provide a calibrated probability that a barcode contains a cell. Formal cell-calling methods provide a more explicit statistical comparison between background and cell-associated signal; that is a separate inferential task. These count-based summaries are useful for understanding the data and recognizing likely background, not for replacing a method designed to make and evaluate cell-calling decisions.

A common mistake is to say, “This barcode has only a few UMIs, so it is empty.” The correction is that few UMIs are evidence compatible with an empty droplet, not a definitive classification. Assess the total, detected genes, and composition together, compare them with barcodes from the same library, and retain uncertainty where low-RNA cells and ambient barcodes overlap.