TxResponse
Predicts how sensitive a tumor is to different drugs
Cell line pretraining on GDSC2 dataset
Domain adaptation to patient samples
Uncertainty calibration on clinical data
Correlation with clinical response
Overview
Predicts how sensitive a tumor is to different drug classes based on its molecular profile, using interpretable biological pathway scores. Instead of a black-box prediction, it shows which biological pathways are driving the sensitivity or resistance — so a clinician can understand why the model recommends one drug over another. Trained first on cell-line drug response data, then adapted to match real patient biology.
Inputs
3 inputsz_bio248Biological state (z_prolif(1) + z_pathway(200) + z_ctx_clean(31) + z_residual(16))
driver_emb[50, 64]Top driver gene embeddings
drug_emb[D, 128]Drug molecular embeddings (ChemBERTa/Morgan fingerprints)
Outputs
3 outputsbeta[K, D]Clone-specific drug sensitivity for K clones, D drugs
ic50_pred[D]Predicted IC50 values per drug
concepts[50]Interpretable pathway concept activations
Mathematical Formulation
Pathway concept extraction
Drug-specific feature modulation
Final sensitivity prediction
Key Features
- 50 interpretable pathway concepts for explainability
- FiLM-based drug conditioning for drug-specific predictions
- 3-phase training: GDSC → CORAL → calibration
- Cross-attention fusion of driver and drug embeddings
- Uncertainty quantification via MC Dropout
Key Innovations
- 1Interpretable predictions through pathway concepts
- 2Domain adaptation from cell lines to patients
- 3Drug-context interaction modeling
- 4Clinically calibrated uncertainty estimates
Hyperparameters
502561280.210050 eachTraining Details
Phase 1: GDSC2 cell line IC50 regression. Phase 2: CORAL domain adaptation to TCGA expression profiles. Phase 3: Temperature scaling on TCGA clinical response labels.