OPEN LINK ↗
// 1h agoRESEARCH PAPER
NexuST Models Genes, Cells, Tissue Together
NexuST is a hierarchical spatial-transcriptomics foundation model that jointly models gene expression inside cells and cell organization within tissue. Its preprint reports training on 45.7 million cells across 72 datasets and competitive results across annotation, region prediction, gene recovery, and neighborhood composition tasks.
// ANALYSIS
NexuST targets the field’s central modeling tradeoff: preserving molecular detail without losing tissue context. The approach is promising, but its real test will be reproducibility and performance across technologies beyond the authors’ benchmark.
- –Interleaving gene-level and cell-level modeling lets each representation refine the other during pretraining
- –HumanST-46M provides unusually broad scale, spanning 11 organs and three imaging platforms
- –The model complements spatial specialists such as Novae and HEIST with tighter cross-level integration
- –Strong gains on context-dependent tasks matter more than headline performance on cell-intrinsic expression
- –As a preprint, NexuST still needs independent validation, released weights, and practical tooling before broad adoption
// TAGS
nexustresearchtrainingdatasetembedding
DISCOVERED
1h ago
2026-10-01
PUBLISHED
1h ago
2026-10-01
RELEVANCE
8/ 10
AUTHOR
changmyung1981