AI-enabled single-cell dissection of the palmitoylation landscape identifies a multicellular prognostic program in gastric cancer

Gastric cancer remains highly lethal, yet how protein S-palmitoylation shapes tumour ecosystems and clinical outcome is unclear. We integrated single-cell RNA sequencing (119,931 cells from 25 gastric tumours) with spatial transcriptomics and bulk cohorts to delineate palmitoylation-linked states across malignant, immune, and stromal compartments. A palmitoylation-high malignant programme partitioned into three metastasis-enriched subclusters with increased fatty-acid metabolism and Ras-MAPK signalling and predicted worse survival. Spatial mapping and ligand-receptor inference revealed co-localised niches where palmitoylation-high tumour cells interacted with immunosuppressive myeloid cells and distinct CAF subsets, with strengthened pro-angiogenic and pro-fibrotic cues. We derived and validated an 87-gene multicellular palmitoylation signature for risk stratification, and higher scores were consistently associated with adverse outcomes in external cohorts. Drug-response modelling highlighted vulnerabilities involving the HSP90-PI3K/MAPK axis. Functional assays and xenografts confirmed SH3BGRL as a key driver within this poor-prognosis programme.