Spatial transcriptomics measures gene expression while preserving tissue context, but learning useful representations from that data is tricky. Standard gene selection can miss signals that only appear in a spatial pattern, and methods that encode cell neighborhoods often throw away direction or rely on absolute coordinates tied to the tissue sample. That means cells with similar neighbor distances but different spatial layouts can end up with nearly identical representations, even though their biological contexts differ.
Tessera, introduced in a paper accepted at the NeurIPS 2026 Workshop on AI for Drug Discovery, tackles both problems at once. It combines spatially informed gene selection with a geometric inductive bias that places each cell in its local tissue environment. By modeling relative cell-cell relationships, the framework is translation-invariant yet direction-sensitive, so neighboring cells contribute differently depending on their spatial arrangement and state.
Across held-out tissues and technologies, the authors report that Tessera outperforms existing spatial foundation models in representation quality. It also generalizes zero-shot to contexts up to 16 times larger than those seen during training. The paper is a single-source report, so no independent comparisons or conflicting results are noted here.