Heterotypic and Homotypic Doublets

Why Doublet Type Matters

A doublet barcode contains RNA from two cells captured together. The two cells need not be transcriptionally alike. If they represent distinct cell types, the barcode may contain evidence of both identities; if they represent the same type, its profile may look much like an ordinary cell of that type. This difference matters because the chance of forming a doublet and the chance of recognizing one from its transcriptome are not the same thing.

The distinction is relative to a classification. Two cells assigned to the same broad category may belong to different subtypes, while two cells assigned to different labels may have very similar expression profiles. It is therefore useful to keep separate the labels used to describe cells and the actual transcriptional difference between the cells contributing RNA.

Expected Composition of Doublets

Suppose the cell suspension has \(K\) types, with type \(k\) making up proportion \(\pi_k\), where \(\sum_{k=1}^{K}\pi_k=1\). Assume the two cells in a doublet are sampled independently from this composition. This is a useful approximation when the pool is large and cells mix without type-specific pairing. It does not describe situations in which capture or cell-cell aggregation depends on cell type.

The probability that both sampled cells are type \(k\) is \(\pi_k^2\). Summing over types gives the homotypic probability. The heterotypic probability is its complement:

These are proportions among doublets under the stated pairing model, not the overall probability that a barcode is a doublet. The total doublet rate is a separate quantity. Also, these formulas depend on the cell-type proportions: a rare type contributes relatively few pairings, whereas a dominant type can contribute many same-type pairings.

If a finite pool contains \(N\) cells, of which \(n_k\) are type \(k\), and the two cells must be distinct, the exact probability of a homotypic pair is \(\sum_k n_k(n_k-1)/[N(N-1)]\). The independent-sampling expression is an approximation to this quantity when the pool is large. Either calculation assumes random pairing; preferential aggregation or differences in capture can change the actual composition.

Cell-Type Resolution Changes the Labels

The homotypic-versus-heterotypic calculation depends on how finely cell types are defined. If two subtypes are grouped into one broad type, a pair drawn from those subtypes is called homotypic at the broad level. If the subtypes are separated, the same pair is called heterotypic. The RNA in the barcode has not changed; only the classification has.

Why Transcriptomes Often Reveal Heterotypic Events More Readily

A useful model for the expression profile of a barcode makes the distinction precise. Let \(\boldsymbol{\theta}_a\) and \(\boldsymbol{\theta}_b\) be the vectors of relative transcript frequencies for two cell types. Each vector has nonnegative entries that sum to one. If the cells contribute \(U_a\) and \(U_b\) captured molecules, respectively, the expected relative profile of their pooled RNA is a weighted average:

This expression assumes that captured counts add across cells and that the relative profile is considered after accounting for total library size. It does not require equal RNA contribution. If the cells contribute equally, \(w=0.5\). If the two profiles are identical, the pooled profile equals either one, whatever the value of \(w\). If the profiles differ, the pooled profile lies between them and can contain transcripts characteristic of both.

Sequencing counts are a sample from this underlying profile, so a barcode does not necessarily show every transcript that the mixture could contain. Finite depth, low expression, and sampling variation can hide informative genes. Still, when the two cells have sufficiently distinct markers and enough of each marker is captured, the joint presence of both signatures can be evidence of a heterotypic event. A homotypic pair lacks this contrast: after normalizing for library size, pooling two cells with the same profile does not create a new cell-type combination.

Interpretation, Diagnostics, and Limits

Heterotypic doublets are often easier to recognize because a mixed profile can combine lineage or cell-state signals that would not ordinarily occur together in one cell. The advantage is strongest when the contributing types are well separated, both contribute appreciable RNA, and their distinguishing genes are expressed and captured. It weakens when one cell contributes little, sequencing depth is low, or the two profiles are similar.

Homotypic doublets are not necessarily identical to singlets in every respect. Their total captured RNA or detected-gene count might be unusually high. But these quantities also vary among genuine cells because of cell size, RNA content, sequencing depth, and technical capture. Consequently, elevated counts alone do not establish a homotypic doublet, and normalization can remove much of the total-count difference while leaving little type-specific evidence. Conversely, a heterotypic barcode may not look obviously mixed if one cell contributes few molecules or the cell types share most of their expression profile.

A practical diagnostic is to ask whether a barcode’s apparent mixed identity is supported by multiple informative genes from each putative type, rather than by one isolated transcript. Then consider whether the types are biologically and transcriptionally distinct at the resolution being used, whether depth is sufficient to observe the markers, and whether other explanations—such as ambient RNA or uncertain cell labels—could produce the same pattern. These checks guide interpretation; they do not by themselves provide a calibrated detection rule.

The main alternatives are to treat all doublets as equally detectable, or to distinguish them only by an increase in total counts. The first ignores cell composition and expression similarity. The second may flag some doublets but confounds them with genuine high-RNA cells. A compositional view explains why neither approach captures the full problem: the expected heterotypic fraction depends on type frequencies, while transcriptomic evidence depends on profile separation and sampling. Neither the composition formula nor the pooled-profile argument is a method for assigning an individual barcode to a doublet class.

The key limitation is that a cell-type label is not itself a measure of expression distance. Two distinct labels may have nearly indistinguishable profiles, and a broad label may conceal substantial subtype differences. Therefore, heterotypic status predicts a possible source of evidence, not a fixed detection outcome. Any estimate of doublet composition should state the classification resolution and pairing assumptions; any claim about detectability should account for the transcriptomic and sampling conditions that make the mixed signal observable.