PRISM: probabilistic cell-type resolution with integrated spatial modelling for simultaneous ambient RNA decontamination and doublet correction in spatial transcriptomics

Spatial transcriptomics count matrices are confounded by ambient RNA contamination and cell-segmentation doublets, two artifacts that co-occur but are rarely modelled jointly. We present PRISM, an Expectation-Maximization framework that simultaneously (i) assigns each cell a probabilistic cell-type identity, (ii) detects doublets via log-likelihood ratio testing of two-component mixtures, and (iii) estimates a per-cell contamination fraction α anchored to a spatially resolved ambient profile. In an in silico benchmark of 40,800 simulated cells, PRISM recovers α with Pearson r=1.000 across all contamination levels (0.05 to 0.50) and UMI depths (500 to 5,000), with zero doublet false positives and 100% singlet accuracy. Applied to a Visium HD mouse cerebellum dataset (four samples, 125,289 cells, 17 reference types) from a metachromatic leukodystrophy (MLD) model, PRISM identifies a median α=0.60, flags 67.4% doublets (driven by the densely packed cerebellar granule layer), eliminates spurious myelin signal from microglia (0.82% to 0.00%), and doubles oligodendrocyte myelin purity (2.8% to 5.6%). PRISM runs on commodity CPUs in under 15 minutes and returns decontaminated integer count matrices for every measured gene.