Omnicell Learns Tissue-Contextual Gene Function from Single-Cell and Spatial Transcriptomes

Authors:

Abstract
Transcriptional programmes are interpreted through the tissue neighbourhoods in which they occur, yet most transcriptomic foundation models learn from dissociated proffles or add spatial information after representation learning. Here we introduce OmniCell, a tissue-contextual foundation model pretrained on 67 million dissociated and spatially resolved proffles. OmniCell encodes each cell or spot as a sequence of gene tokens that combine gene identity, expression magnitude and observed tissue coordinates, allowing spatially aware attention and a symmetric cell–gene objective to learn coupled cell and contextual gene representations. Across cellular, molecular and tissue scales, OmniCell preserved transferable cell-state structure across species, batches and rare populations, recovered marker genes and tissuealigned gene programmes, and improved spatial identity mapping, deconvolution and anatomical domain reconstruction. In human liver cancer Stereo-seq data, OmniCell resolved a tumour-margin transition zone and showed that gene-neighbourhood relationships shift between tumour core, transition zone and adjacent non-malignant tissue, consistent with niche-dependent gene function. In silico spatial perturbations further nominated region-speciffc regulatory hypotheses in developing mouse brain and marmoset cortex. OmniCell therefore provides a framework for modelling gene function as a property of cells embedded in intact tissue, rather than as a global expression pattern alone.