The problem. A Visium spot is a bag of several cells’ transcripts. To recover which cell types are where, you need to map a single-cell reference onto each spot — and do it in a way that survives technical noise and platform differences, not just curve-fit the counts.

The idea. Cell2location is a hierarchical Bayesian model. From an annotated single-cell reference it learns per-cell-type expression signatures, then models each spatial spot’s observed counts as a combination of those signatures scaled by (absolute) cell-type abundances, with explicit terms for technical effects — sensitivity differences, additive background — and a negative-binomial likelihood for the count noise. Inference is variational, and the payoff is calibrated absolute abundance estimates with uncertainty, not just proportions.

Why it matters. This is one of the methods the benchmark crowns, so reading the actual model — rather than treating it as a black-box winner — is what lets me explain why it works: sharing information through the reference, and modeling technical variation instead of pretending it away. That’s the same empirical-Bayes / hierarchical-modeling instinct that shows up in DESeq2 and scVI, which makes it a satisfying through-line across the reading list rather than an isolated tool.

Verdict. A genuinely principled method, and the Bayesian machinery is a feature (calibrated uncertainty) more than a cost. Honest limits: it needs a good matched single-cell reference and enough compute for variational inference, and like all deconvolution it’s for spot assays — imaging data at single-cell resolution wants segmentation instead. Read it back-to-back with RCTD to see two coherent takes on the same problem.