Zero-annotation cell typing for spatial proteomics in the weak-linkage regime

Spatial proteomics measures a few dozen protein markers on every cell of a tissue section, and assigning each cell a type normally requires expert annotation of that tissue. Public single-cell RNA atlases carry expert labels, but share with a protein panel only the handful of genes that encode the measured proteins. Our zero-annotation method uses the atlas once to give every cell a rough label and then discards it. A classifier trained from these noisy labels on the tissue's own protein panel produces the final annotation inside a graph-regularized self-training loop. We evaluate three public benchmarks: tonsil CODEX, breast IMC, and colorectal cancer CODEX. Our mean accuracy over three seeds is highest on all three pairs: .93, .75, and .61, compared with .82, .69, and .43 for MaxFuse, the strongest competitor overall. In balanced accuracy, our method is higher than MaxFuse on tonsil and colorectal cancer. Two analyses separate the contributions of the reference and the loop. Registered tissue with known spatial correspondence shows that biological signal crosses the weak link even though corresponding spatial bins cannot be matched reliably. When initialization errors are scattered, self-training denoises the atlas-derived labels: on tonsil, accuracy rises from .78 to .93.