Spatial multimodal framework for prostate cancer bone metastasis characterization using Xenium 5K spatial transcriptomics.
OncoGrad integrates three parallel encoders to characterize tumor microenvironments at neighborhood resolution:
- BioKGEncoder (RGCN): encodes gene regulatory state through a biological knowledge graph built from TRRUST TF-target relationships, GO biological processes, and KEGG pathways over 2,000 highly variable genes
- XeniumTransformer: encodes spatial cellular composition from Xenium single-cell annotations within 50µm-radius neighborhoods
- ViT: encodes tissue morphology from aligned H&E patches (224×224px, 0.5µm/pixel)
Applied to prostate cancer (primary PRAD vs bone metastasis) across 20 spatially profiled samples.
- Test AUC: 0.86 on held-out patient neighborhoods
- Key findings: AR, GATA2, SOX9, HIF1A/ARNT, SMAD3/SMAD7, MEF2A/NFKB2
See requirements.txt for full version specifications.
- Install dependencies: pip install -r requirements.txt
- Update paths in the Configuration cell of OncoGrad.ipynb
- Run cells sequentially
Xenium 5K prostate cancer dataset (primary PRAD + bone metastasis, 56 samples total, 20 with aligned H&E). Data access through WashU Ding Lab.