When is a cell atlas complete? A new study asks where to stop

The idea in the new paper

Miihkinen and colleagues adapt species-richness estimation from ecology to single-cell sequencing. They aim to estimate cellular cluster richness and identify when extra sequencing is unlikely to uncover new clusters. The available HTML describes this purpose, but supplies no benchmark results we can independently evaluate.

Write the target before the word complete

Imagine a map of a city. Recording every street does not establish that every building, opening hour and resident has been recorded. A cellular atlas faces a similar problem: the inventory has to specify what counts as an item.

Our suggested completeness statement would name the tissue, the preparation, the sampled population and the rule used to define a cluster. A reader could then ask whether a later claim concerns that same inventory. Changing the target changes the meaning of completeness, even if the dataset itself stays fixed.

More observations and broader coverage are different purchases

Consider two proposed next batches: many additional cells from already sampled material, or fewer cells from a previously unrepresented donor or region. They might have different value even if the total cell counts match. This is a hypothetical design comparison, not a measured result from the paper.

Our proposed reporting improvement is to show discovery curves beside a coverage table. The curve addresses additional observations under the current strategy. The table asks which donor groups, locations and preparation conditions that strategy leaves out.

Make the estimate face a held-out batch

A useful future evaluation would freeze the initial map and its stopping decision, then sequence a prespecified held-out batch. Record whether genuinely new groups appear and whether they change the intended research conclusion. Repeat across different starting sizes and sampling conditions.

We would score missed discovery and unnecessary sequencing separately. An estimator could save resources while missing a rare group important to one application. Another might protect discovery while recommending much more sampling. The right balance depends on the question the atlas is meant to answer.

Do not let a clustering choice decide the whole story

As a proposed robustness check, repeat the analysis under justified alternative cluster definitions and technical processing choices. If the stopping decision changes sharply, that sensitivity belongs next to the result.

The goal is not to manufacture more clusters. It is to reveal whether the estimate depends on a biological distinction, a technical artifact or an arbitrary grouping rule. We have not inspected the paper’s implementation and cannot say which of these checks its authors performed. They are questions for the full methods and an independent evaluation.

A useful future atlas would publish its unfinished edges

Our proposed atlas dashboard has three entries: the defined discovery target, the uncertainty about unseen groups, and known gaps in sampling. A new batch would update those entries rather than simply enlarging the headline cell count.

Read this with our brain-atlas and map-to-mechanism guides. The connection is a question: after locating a group, what experiment would establish its function? An inventory can guide that experiment while leaving its answer open. The opportunity is more transparent resource allocation and clearer maps, not a certificate that the biology is finished.

What would a completeness claim need to specify?

Original evaluation framework. These are proposed checks, not performance results reported in the accessible abstract.