The problem. Deconvolution leans on a single-cell reference, but the reference and the spatial assay are measured on different platforms, with different biases. Ignore that gap and you get confident, wrong cell-type maps. RCTD’s central claim is that handling this platform effect is what separates a robust method from a fragile one.

The idea. RCTD (Robust Cell Type Decomposition, shipped as spacexr) fits a statistical model that learns cell-type profiles from a reference and estimates each spot’s mixture while explicitly estimating and correcting the systematic platform difference between reference and spatial data. It runs in modes suited to the assay’s resolution — from near-single-cell (one or two cell types per pixel) to lower-resolution spots — so it degrades gracefully rather than assuming one cells-per-spot regime.

Why it matters. Read alongside cell2location, this is the instructive contrast: cell2location leans on full hierarchical Bayesian modeling of technical variation; RCTD foregrounds the reference-vs-spatial platform mismatch as the thing to correct. Same problem, two philosophies — and understanding both is what lets me pick, and justify, one for a given dataset rather than defaulting to whatever’s popular. That “know why, not just which” is the standard I want for every pipeline choice.

Verdict. Robust and widely used for good reasons, and unusually clear about what it’s correcting for, which I appreciate in a method. Same honest ceiling as all deconvolution — it needs a decent reference and is aimed at spot/near-spot data, with segmentation the better frame for true single-cell imaging. This closes the deconvolution cluster; segmentation is where the reading goes next.