MOSAIC (morphology-oriented spatial attention for inferring cell expression)

Spatial transcriptomics maps gene expression in tissue, but cost and throughput limit large studies of ulcerative colitis and related disorders. Here we present MOSAIC, a two-phase computational framework that predicts single-cell gene expression from routine H&E histology and then assigns cell lineages using only those predicted profiles. Phase A uses each cell’s image together with six spatially nearest neighbours; Phase B classifies nine curator-defined in-house lineages (integrative clustering of challenge expression; not supplied as challenge annotations) without further use of images. On inflamed colon tissue from the Autoimmune Disease Machine Learning Challenge, with training and test on different patient sections, mean agreement between predicted and measured expression was modest (Pearson 0.22 and Spearman 0.17 across 460 genes) with heterogeneous recovery across genes, and lineage accuracy reached 63%—well above chance (11%) and a majority-class rule (14%), though below a control trained on measured expression (95%). Predicted profiles therefore retain usable lineage structure on a held-out section. These findings support a proof-of-concept under conservative evaluation; they do not establish clinical utility or generalisation beyond this dataset.