CausalDriver-GAT
Identifies which mutations are actually driving the cancer
Area under ROC curve for driver classification (v5.10 retrained)
Area under precision-recall curve (v5.10 retrained)
Known drivers in top 10 predictions
Attention weights show z_ctx influence
Overview
A graph neural network that identifies which mutations are actually driving the cancer versus which are passengers along for the ride. It uses the biological interaction network between genes — which proteins talk to which — combined with the specific tumor context to rank every mutated gene by how likely it is to be a true driver. It also classifies each driver as targetable, resistance-causing, or currently undruggable.
Inputs
3 inputsz_ctx_clean31Proliferation-free biological context from VAE
Gene Features4 per geneMutation status, expression, CNV, centrality
Gene Graph~3,000 genesCOSMIC+Reactome nodes, SIGNOR+STRING edges (conf>0.9)
Outputs
3 outputsdriver_prob[N]Per-gene driver probability scores
top_k_embedding[50, 64]Embeddings of top 50 predicted drivers
actionability[N, 3]Targetable, resistance, undruggable classification
Mathematical Formulation
Dynamic attention coefficients
Context modulation of node features
Neighborhood aggregation
Key Features
- GATv2 attention for dynamic edge weighting
- Context injection via FiLM conditioning on z_ctx_clean
- Multi-head attention (4 heads) with residual connections
- Hierarchical message passing (3 layers)
- Actionability classification head
Key Innovations
- 1Patient-specific driver ranking using tumor context
- 2Integration of biological context with graph structure
- 3Actionability-aware predictions for clinical utility
- 4Robust to graph noise via attention-based aggregation
Hyperparameters
64420.15e-450Training Details
Trained on 4,479 IntOGen entries (~568 drivers). Retrained on v5.10 latents. Accuracy 0.9591, best at epoch 82/100. Node features: [mutation, CNV, expression, essentiality] + z_ctx(31d).