How would you validate that a computationally inferred cell state is real biology?

Why

A clustering or trajectory algorithm will always return structure, so finding a state computationally is a hypothesis, not a result. Validation proceeds in two layers. Inside the data, the state should appear in several donors and batches, keep its markers when the number of principal components or the clustering resolution changes, and remain after removing doublets, damaged cells and stressed cells.

The Heumos review stresses that results depend on capturing the major sources of variation, and that ignoring sources of biological variability inflates false discoveries. Outside the data, the state needs independent evidence. Its marker proteins can be measured with flow cytometry or CITE-seq antibodies, its location seen with in situ hybridization or spatial transcriptomics, and its function tested by sorting the cells and assaying them.

A state defined by a gradient, such as activation, may be real but continuous, and then the claim should describe a gradient rather than a separate population. Milo's neighborhood testing is one way to show a state changes with condition without forcing discrete clusters.