Goal: Integrate spatial transcriptomics with single‑cell RNA‑seq (scRNA‑seq) from pancreatic ductal adenocarcinoma (PDAC) to reveal spatially‑resolved tumor–stroma interactions using a Graph Neural Network–enhanced Multi‑Modal Variational Autoencoder (VAE).
- Graph Neural Networks (GNNs) preserve spatial proximity between cells/spots.
- Modality‑specific priors in the VAE respect the distinct data distributions of spatial and RNA modalities.
- Joint latent space fuses modalities while retaining biological signals.
- Explainable AI (e.g., SHAP) interprets latent features.
- Source: Public dataset from Ateeq M. Khaliq et al., Nat Genet 2024 (30 matched primary & metastatic PDAC samples).
- Article link: https://doi.org/10.1038/s41588-024-01914-4
- Original Format: RDS files provided by authors.
- Processed Format: Converted to
h5adfor compatibility with Scanpy-based tools. - Access: https://doi.org/10.6084/m9.figshare.28835765.v1
- Status: Already pre‑processed, quality‑controlled, doublet‑filtered (Seurat/Scanpy).
-
Download & QC
- Pull spatial and scRNA‑seq matrices.
- Confirm QC metrics; filter low‑quality cells/spots.
-
Graph Construction
- Nodes: cells / spots.
- Edges: spatial distance + expression similarity.
-
Encoder Setup
- Spatial encoder → GNN layers.
- RNA encoder → dense layers.
- Learn modality‑specific priors.
-
Latent Fusion
- Project each modality into a shared latent space.
- Optimize reconstruction loss.
-
Decoder & Reconstruction
- Reconstruct original modalities from latent vector.
- Evaluate with held‑out data.
-
Interpretation
- Apply SHAP to latent features.
- Map important dimensions back to tissue coordinates.
-
Biological Analysis
- Identify spatial gene modules.
- Detect tumor microenvironment niches.
- Nominate biomarkers & therapeutic targets.
- Method: Open‑source framework for spatial multi‑omics integration.
- Insights: Spatially coherent signatures of tumor, stroma, and immune niches.
- Applications: Diagnostics, prognostics, and therapy guidance in PDAC and beyond.