Combination Discovery Engine
Predicts synergistic drug combinations from monotherapy data alone
Predicted vs actual synergy on 1,209 drug pairs (p < 0.0001)
Prediction survives when entire drug classes hidden (5,302 pairs)
Dose reduction vs concurrent while preventing resistance
GDSC drugs with pathway sensitivity profiles
Overview
Discovers synergistic drug combinations without training on any combination data. The MechanismOperator maps 286 GDSC drugs to pathway sensitivity profiles via the structured VAE latent space, then the CombinationsEngine scores drug pairs by orthogonal clonal targeting: finding combinations where drug A suppresses dominant clones while drug B targets the Resistance Sentinel. Uses Factorized β (bulk pathway sensitivity × clone-specific mutation modifiers from 130+ curated drug-gene associations) to compute per-clone drug sensitivity. Includes PK/PD-constrained Schedule Optimizer that finds time-varying dosing strategies minimizing tumor burden while respecting toxicity budgets.
Inputs
4 inputsz_pathway200 (50×4)Patient pathway activations from VAE latent space
Clone genotypes4 slotsPer-clone mutation lists from CloneMapper (driver mutations + resistance markers)
Clone fractions4Tumor mass fraction per clone (simplex, sums to 1.0)
GDSC drug-cell pairs242K pairsDrug sensitivity (LN_IC50) for 286 drugs across 685 cell lines with VAE latents
Outputs
3 outputsRanked combinationsTop N pairsDrug pairs scored by synergy (orthogonality × coverage × sentinel targeting)
Per-clone β(4, N_drugs)Factorized drug sensitivity per clone: β_bulk × Δ_clone
Optimized scheduleT × 2Time-varying doses for drug A and B with PK/PD constraints
Mathematical Formulation
Per-clone drug sensitivity = bulk pathway score × mutation modifier
Weighted orthogonality, coverage, and sentinel targeting
Clonal dynamics with drug-modulated death term
Drug concentration decay with half-life
Key Features
- Zero-shot combination prediction from monotherapy data only
- Factorized β: bulk pathway sensitivity × clone-specific mutation modifiers
- 130+ curated mutation→drug associations for per-clone sensitivity
- Adaptive threshold and diversity filter for robust scoring
- PK/PD-constrained schedule optimization (half-life, toxicity budgets)
- Resistance Sentinel targeting as explicit combination criterion
Key Innovations
- 1Predicts drug combinations without training on any combination data
- 2Orthogonal clonal targeting: drug A suppresses dominant clones, drug B targets resistant sentinel
- 3Factorized β resolves the levels-of-variation mismatch between bulk RNA and per-clone mutations
- 4Differentiable schedule optimization through Lotka-Volterra ODE with PK constraints
Hyperparameters
286 profiled685 with 328d VAE latents130+ mutation-specific scores12 monthly (L-BFGS-B)4 classes (cytotoxic, targeted, immunotherapy, default)orth=0.35, cov=0.35, sentinel=0.30Training Details
MechanismOperator fitted on 242K GDSC drug-cell pairs (286 drugs, 685 cell lines with 328d VAE latents). Per-drug pathway profiles computed via Spearman correlation of z_pathway activations with LN_IC50. Holdout prediction: mean Spearman ρ = 0.432. Combination validation: 1,209 pairs with ≥30 shared cell lines, predicted vs actual ρ = 0.800. Leave-target-family-out: 5,302 pairs, ρ = 0.689 (14% degradation, proves genuine mechanistic understanding). Schedule optimizer: L-BFGS-B with 12 monthly blocks, PK/PD constraints (half-life decay, max cumulative dose). 42% dose reduction vs concurrent in test scenario.