TERRA: A tissue world model, with boundaries
— by Vivax
TERRA learns contextual representations across genes, cells and tissue neighborhoods from spatial transcriptomics. It is a preprint, not a clinical simulator.
Spatial transcriptomics shows gene activity and cell location, but many analyses summarize cells alone. TERRA models an index cell with up to ten nearest neighbors. It creates contextual embeddings for genes, cells and neighborhoods, a precise—not universal—sense of a tissue world model.
Its tokens combine gene symbol, expression value, within-cell rank, and cell-rank or proximity; an optional batch token accounts for assay variation. The vocabulary has 23,407 Ensembl-keyed genes. Local arrangement is part of the representation, not metadata outside it.
TERRA uses JEPA: a context encoder sees masked input, an EMA target encoder supplies latent targets, and a predictor matches them. It predicts masked gene-token representations rather than observed counts. It is not a pixel generator, causal proof engine, or complete tissue simulator.
HST-Corpus-112M contains 112,578,039 cells from 636 sections, five imaging assays, 20 tissues, and health plus 26 disease conditions. Across four held-out niche datasets, authors report highest NMI among comparators. That is a representation benchmark, not clinical efficacy. This bioRxiv work is not peer reviewed.
In unseen fetal pancreas, 12 sections across seven stages produced 1,586,341 cells, 13 niches and 53 gene programs; a capillary precursor is suggested, not lineage-proven. In kidney, CTLA4/PDCD1 edits yielded divergence near 0.08 versus 0.13 unperturbed. This is an in-silico alignment signal, not treatment or toxicity prediction.
Across organs, 16 macrophage archetypes explained 76.2% of inter-niche compositional variance; tumour-boundary and TCGA survival findings remain observational. Vivax sees TERRA as evidence that clinical world-model research should connect molecular state, local context and explicit validation—not as evidence that an in-silico perturbation is a clinical recommendation.