methods.md

_run_tier2.py

Gene expression (9,660 features)
        ↓  GCN with Reactome pruning masks
        ↓  layer0 → layer1 → layer2 → layer3
Cell embedding  (embed_dim=32)       ← pathway-structured
        ↓  AttentionPool
Donor embedding (embed_dim=32)       ← still pathway-structured content
        ↓  fusion MLP
Donor prediction

Instead of Attention poolig cell to donor could use:

1. pseudobulk (just as a benchmark)
2. Multihead-attention
3. Hierarchical GNN