Droplet-based single-cell RNA-seq co-encapsulates a barcoded gel bead and (ideally) exactly one cell per oil droplet. Loading follows a Poisson process: at concentration λ (cells per droplet), the chance a droplet gets exactly k cells is fixed by nature, not by the machine.
P(k cells in droplet) = e^(-λ) · λ^k / k!
P(empty) = e^(-λ) P(singlet) = λe^(-λ)
doublet rate (of captured) = [1 - P(0) - P(1)] / [1 - P(0)]
- Loading concentration λ — push it up to capture more cells, but the same Poisson math forces more droplets to trap two cells (doublets) that later masquerade as a fake hybrid cell type.
- Sequencing depth — reads per cell; low depth undercounts genes (dropout) and blurs cell-type clusters together in the UMAP embedding.
- Cell types (K) — how many distinct transcriptional states are present in the tissue being profiled.
- Cells profiled — sample size passed through the pipeline / plotted in the atlas.
Real-world relevance: this exact Poisson trade-off is why labs report a "target doublet rate" (often ~1–8%) when choosing loading concentration for atlases like the Human Cell Atlas.