Virtual spatial transcriptomics from archived histology: An out-of-domain proof of concept for a rare paediatric disease

Very early onset inflammatory bowel disease (VEO-IBD) comprises severe and often treatment-refractory intestinal disorders in young children, for which molecular characterization can inform diagnosis and therapy, but remains unavailable for many patients. Histology-to-expression models could derive spatial molecular information directly from routine H\&E images, yet their evaluation is challenging because correlation against sparse, count-limited spatial transcriptomic lacks an interpretable denominator. Here we apply two published models, Phoenix and CellPin, without modification to $\sim$ 700K cells from 36 paediatric patients, a tissue outside the domain of the adult colon sections used to fit the released Phoenix model. Per-cell correlation was modest ($\rho_{\text{gene}}$ $= 0.064$), 18\,\% of the measurement-derived noise ceiling, whereas the same predictions preserved the relative spatial organisation of genes ($\rho_{\text{rank}}$ $= 0.849$). On a pathologist-annotated region task an image-only baseline matches the predicted expression pipeline, bounding what predicted expression contributes there. Virtual spatial transcriptomics therefore recovers biologically relevant spatial information in rare paediatric disease despite substantial domain shift, and its evaluation needs noise ceilings, permutation nulls, and morphology-only controls across spatial resolutions.