Treatment Optimizer
Ranks treatment regimens using counterfactual reasoning
Cell-holdout rank correlation on GDSC drug sensitivity
Factual concordance of TARNet counterfactual model
All end-to-end pipeline tests passing
Not deployed for clinical use — needs OPE validation + external treatment data
Overview
A counterfactual treatment ranking engine that combines cell-line drug sensitivity data with patient-level survival modeling to rank treatment regimens. Phase 1 learns drug-cell interactions from GDSC (685 cell lines, 295 drugs). Phase 2 uses TARNet with inverse propensity weighting to estimate conditional average treatment effects from TCGA survival data. Includes abstention gating, Pareto optimization across efficacy/toxicity/resistance, and sequential treatment planning with beam search. Currently in shadow mode — not deployed for clinical recommendations.
Inputs
3 inputsz_cell / z_patient328Full VAE latent representing cell line or patient biology
drug_emb128Drug molecular embeddings
cancer_type32 (embedding)Cancer type embedding for treatment-specific heads
Outputs
3 outputsIC50 predictions[295]Predicted drug sensitivity for 295 compounds
Treatment ranking[49 regimens]Pareto-ranked regimens with abstention flags
Sequential plan[3 lines]Beam-searched 3-line treatment sequence with resistance penalties
Mathematical Formulation
Bilinear drug-cell sensitivity prediction
Conditional average treatment effect from factual/counterfactual heads
Inverse propensity weighting for observational confounding
Key Features
- Phase 1: Bilinear drug-cell interaction model pretrained on GDSC (242K pairs)
- Phase 2: TARNet with per-treatment FiLM heads and IPW deconfounding
- Phase 3: 49 regimen registry with DrugMapper (GDSC↔TCGA↔clinical)
- Phase 4: Pareto optimization with abstention gate and propensity safety
- Phase 5: Sequential beam search over 3 treatment lines with resistance penalty
Key Innovations
- 1Cross-domain drug sensitivity transfer (cell line → patient)
- 2Counterfactual treatment effect estimation with TARNet
- 3Abstention gating for uncertain treatment recommendations
- 4Sequential treatment planning with pathway-based resistance modeling
Hyperparameters
1e-35e-412825649 (10 types + default)5 (sequential planner)Training Details
Phase 1: GDSC pretrained on 242K drug-cell pairs (685 cell lines with VAE latents, 295 drugs). Phase 2: TARNet trained on TCGA survival with IPW. Phase 4: Pareto ranking across efficacy, toxicity, resistance scores. Expert consensus: beta is under-identified, propensity uncalibrated for rare classes. Needs OPE validation.