PlantDeconv: A structure-guided probabilistic deconvolution framework for plant spatial transcriptomics

Plant tissues have spatial organization that differs from many animal tissues. Because plant cells are immobilized by cell walls, cell types often occupy stable domains, form clear anatomical boundaries, or vary gradually along developmental axes. These features are important for interpreting plant spatial transcriptomic data, but most deconvolution methods rely mainly on expression similarity or generic spatial smoothing and do not explicitly encode plant tissue structure. Here, we present PlantDeconv, a structure-guided probabilistic deconvolution framework for plant spatial transcriptomics. PlantDeconv links reference single-cell expression profiles to spatial transcriptomic spots through cluster-to-spot probabilistic mapping and incorporates four plant-relevant constraints: a cell-type-to-region prior, expression-aware spatial continuity, a differentiation-gradient constraint, and spot-wise adaptive regularization. Across multi-resolution pseudo-spot benchmarks, PlantDeconv achieved low composition error. In two real plant datasets, it produced spatial predictions with stronger regional enrichment, clearer anatomical boundaries, and less ectopic spread, especially in tissues with complex boundaries or continuous developmental transitions. These results support the use of plant tissue structure to improve anatomical consistency and biological interpretability in spatial deconvolution.