Hypernetwork
Generates personalized tumor parameters for each patient
Phase 0: gate clamp + aux survival + physics bottleneck
18 reliable / 33 cancer types
DSN Hypernet via 128-PDX DSN, v5.10
0.00% violation rate across both paths
Overview
Takes a patient's compressed biological fingerprint and generates the specific growth rate, drug sensitivity, and immune response parameters for that individual's tumor simulation. Every parameter is bounded by biological law — growth rates can't be negative or impossibly fast, drug sensitivity is between 0% and 100%. Two separate paths: one for human clinical data, one for mouse model data. The model automatically learns how much to weight molecular data versus histology images for each cancer type.
Inputs
4 inputsz_full (Path A)328Full multi-modal VAE latent for human/clinical data
z_dsn (Path B)281z_shared(201) + z_meth_imputed(48) + z_cnv_imputed(32) via DSN
WSI embeddings1,536UNI2-h histopathology embeddings (late gated fusion)
Cancer type64 (embedding)Cancer type embedding for FiLM conditioning
Outputs
6 outputsrho[K]Growth rates, constrained to [0, 0.3]/day via sigmoid
beta[K, D]Drug sensitivity, constrained to [0, 1] via sigmoid
omega[1]Immune killing coefficient, constrained > 0 via softplus
N0[K]Initial clone populations (sum = 1)
sigma[1]Stochastic noise scale, constrained > 0
risk[1]Calibrated survival risk score
Mathematical Formulation
Sigmoid-constrained to [0, 0.3]/day
Learned gate with per-cancer bias
Cox partial log-likelihood
Key Features
- Two separate paths: Specialist (328d) and Translator (281d via DSN)
- Late gated fusion: omics 77% / WSI 23% (Path A)
- FiLM conditioning on cancer type embedding (64d)
- Physics bottleneck: ODE params → dynamics projection → survival head
- Anti-gate-starvation: gate clamping + auxiliary per-modality survival heads
- MC Dropout uncertainty quantification
Key Innovations
- 1Dual-path architecture for human and PDX data
- 2Physics bottleneck creates gradient path: L_survival → rho/omega
- 3Triangulated validation (physics + fidelity + utility)
- 4Conditioned gating with per-cancer bias (zero-initialized)
Hyperparameters
328→128 (Path A), 281→128 (Path B)1536→256→12864128200 (early stopping ~61)3e-4 (AdamW)Training Details
Path A: Trained on v5.10 latents (9,393 samples) + WSI + survival. Phase 0 flags: gate_clamp, aux_survival, cancer_moddrop, physics_bottleneck. Early stopping at epoch 61. Path B: Trained on DSN-processed latents (281d) with same architecture. Stratified K=4 cancer type cross-validation.