Dissecting the myeloid cell-driven ecosystems through multi-omics and heterogeneous graph learning in pancreatic ductal adenocarcinoma

Myeloid lineages exhibit prominent enrichment and profound heterogeneity within the pancreatic ductal adenocarcinoma (PDAC) microenvironment, significantly impacting clinical outcomes. Here, we analyzed 45,705 high-quality myeloid cells from patient-matched tumor and blood samples, with normal blood and tissues as controls. This revealed two tumor-promoting subsets: GBP5 ⁺ neutrophils sustaining a persistent pro-inflammatory circuit, presumably mediated by STAT3 , and THBS1 ⁺ monocytes putatively orchestrating immunosuppression via Amphiregulin-driven differentiation blockade. We developed moGT, a heterogeneous graph transformer integrating multi-omic profiles with myeloid cellular abundance. moGT deconvolutes PDAC into four molecular subtypes governed by specific myeloid ecosystems (ECs), architectures supported by spatial transcriptomics. Notably, Subtype 2 exhibits an unexpectedly poor prognosis despite a favorable ‘classical’ phenotype, characterized by a GBP5 ⁺ neutrophil-enriched EC and NAMPT inhibition vulnerability. Conversely, aggressive Subtype 4 is defined by a THBS1 ⁺ monocyte-dominated EC, showing predictive sensitivity to microtubule-targeting agents. Our cross-compartment analysis links myeloid architecture to clinical heterogeneity, providing a precision oncology framework.