Abstract
Abstract
Spatial transcriptomics platforms such as 10x Visium capture whole-transcriptome expression but at spot-level resolution where each measurement aggregates multiple cells. Existing computational deconvolution methods require external scRNA-seq references, and the growing set of morphology-guided methods (iStar, SpaHDmap, GHIST, Thor, PRTS) each requires per-dataset model training or paired subcellular spatial data. Here we present SpatialCell AI, the first framework to combine training-free operation, reference-free expression integration, and per-cell output granularity for spatial transcriptomics. We validated SpatialCell AI on a matched colorectal cancer sample analyzed across Visium (55 μm), Visium HD (8 μm and 16 μm), and Xenium (single-cell ground truth). On Visium HD 8 μm input, SpatialCell AI achieves an expression correlation of
r
= 0.791 against the Xenium reference — the highest of any method tested on this sample. Across six distribution-based validation metrics on the matched 408-gene panel, the SpatialCell AI HD variants lead on the majority of metrics, with no competing method winning on any. A strict matched-region head-to-head comparison reveals a clear architectural distinction: training-free integration improves monotonically as input bins approach single-cell scale, while the trained morphology-guided methods iStar and SpaHDmap track the raw baseline on every platform. SpatialCell AI achieves its strongest performance on Visium HD inputs, where it substantially outperforms both raw baselines and competing morphology-guided methods while transforming spot-level data into individual cell records with spatial coordinates and per-cell expression — an output no raw spot measurement can provide.