How can the completeness of large-scale single-cell sequencing projects be estimated?

Counting cells does not measure completeness

Most single-cell atlases report their size as the number of cells stored. The paper argues this count cannot answer the question that matters for a cataloguing project: how many cell types in the tissue remain undetected. Its own illustration makes the gap concrete. CELLxGENE stores 96.1 million cells, which against roughly 36 trillion cells in the human body would correspond to about 0.0002% completeness, yet because those cells come from organs in a non-uniform way, the authors say such a figure is not an appropriate measure of atlas completeness.

The alternative the paper proposes is to estimate cellular richness, the total number of cell types (treated as cellular species) present in a tissue, and compare it with the number already observed. Because there is no settled definition of a cell type, the work refers to cell classes as cellular species and relies on unsupervised clustering of transcriptomic profiles to define the units being counted. Completeness therefore means the fraction of the estimated richness that current data have recovered, not the fraction of all cells sequenced.

Why ecology estimators fit cell atlases

Species-richness estimators were built for the same practical problem ecology faces: observers can never see every species in a habitat, so the count of observed species underestimates the true total. The paper transfers this logic to single-cell data by building a species-by-sample matrix, in which each entry records whether a given cell species was detected in a given sample. The sparsity structure of that matrix, meaning which species appear in only one or a few samples, carries the information the estimators use.

The transfer is plausible because cell abundances in human organs follow heavy-tailed distributions. Fitting log-normal, negative binomial, log-series, Pareto and Poisson models to cell type abundances across 20 organs in Tabula Sapiens, the authors found log-normal and negative binomial distributions to be the plausible generators of the observed data. They also checked that this pattern was not a droplet-sequencing artefact by reproducing heavy-tailed distributions with Split-seq data and in blood, a liquid tissue.

Chao2 reached the true richness fastest

Because no consensus estimator exists for cellular ecosystems, the authors benchmarked four non-parametric estimators against a simulated human body with a known ground-truth richness. The simulation was scaled up from Tabula Sapiens parameters across 20 organs, and samples were drawn at coverage levels from 10⁻² down to 10⁻⁸. All tested estimators began to stabilise toward the ground-truth richness after sufficient sample coverage, even in large organs with many rare species, and using more than three samples improved accuracy further.

Among them, Chao2 plateaued fastest toward the true population value without tending to overestimate. In the error analysis, Chao2 produced a median 12% reduction in error compared with estimates based solely on observed cellular species. The authors also tested robustness to cell loss by using binomial thinning to raise loss probabilities in individual samples, and found minimal impact on species discovery or on Chao2 performance, which they attribute to the estimator operating on sparsity across all samples rather than on within-sample detection.

The simulation also set a practical expectation: richness could be estimated from samples covering 0.001% or more of the whole cellular population, and the profiling effort needed depended strongly on organ complexity, with more complex organs requiring more sampling.

Real atlas analyses flagged incomplete tissues

The framework was then applied to real data. In the human dorsolateral prefrontal cortex dataset of Emani et al., 79,586 cells passed quality control, and the authors estimate this covers about 0.0011% of the roughly 6.9 billion cells in that brain area, above the threshold their simulations required for the most complex organs. Chao2 recovered the full-dataset richness on average from three or more samples, with more samples narrowing the confidence intervals.

A finer subsampling and re-clustering analysis asked how deep each sample must be sequenced. For the seven high-coverage samples, species recovery rates consistently exceeded 95% and median cluster-agreement scores were above 0.95 once 3,000–4,000 cells per sample were analysed; beyond that, gains in agreement and recall were minimal. The authors describe this as practical saturation and combine it with Chao2 predictions as a strategy for guiding atlas-scale projects.

When the same approach was extended to adult eye, developing eye and developing prefrontal cortex datasets profiled with single-nucleus or single-cell technologies, all were flagged as incomplete, implying that further profiling could still reveal unseen cellular species. The large mouse isocortex dataset of Yao et al. was also assessed, where massively sized samples had already captured the full transcriptional complexity.

What the estimate depends on

The authors state that completeness estimation is agnostic to specific cell type definitions and can be performed with any single-cell technology combined with unsupervised clustering. That flexibility comes with a dependence on clustering quality. They advise testing different cluster numbers and resolutions, avoiding artificially small clusters that introduce spurious sparsity and bias the estimates, and not restricting dataset variability through feature selection such as highly variable gene selection before running the analysis. Because the method relies on detecting sample-unique cell species, each such observation needs checking as a true event or a clustering artefact, and the quality of data integration after batch correction needs careful assessment.

The authors also note that Chao2 has been widely used for clustered data, and they call for future work on manifold-aware estimators that do not depend on clustering. Their simulations were generated from a multivariate log-normal distribution, though as a non-parametric estimator Chao2 is expected to perform adequately even if the underlying population deviates from log-normality.

Where stoppage-points should guide sequencing

The paper's practical proposal is to treat richness estimates as stoppage-points: points at which additional sequencing no longer reveals new cellular clusters, allowing projects to be completed more cost-efficiently. The evidence supporting this differs by tissue. In the simulated organs, three or more samples were enough for the estimators to stabilise, and in the prefrontal cortex data Chao2 recovered full-dataset richness on average from three samples, with narrower intervals as samples increased. The subsampling analysis adds a per-sample depth benchmark of roughly 3,000–4,000 cells for the high-coverage samples studied.

The authors recommend running richness estimates at regular intervals on every profiled organ, and they argue that future effort should shift toward profiling tissues with several complementary technologies, such as single-cell, single-nucleus and spatial methods, which can record different and often mutually exclusive cell types. The priority the paper sets is therefore not simply to sequence more cells, but to check each organ's estimated richness against what has been observed and to direct remaining effort where the gap is largest.

Sources used in this answer