The first fully differentiable tumor digital twin
A neuro-symbolic architecture that bridges abstract biology and continuous physics. One shared encoder, two paradigms: mechanistic interpretation (V1) and gradient-based treatment optimization (V2).
Why neither data source can work alone
Building a tumor digital twin requires both biological depth and temporal density. No single data source has both — each fills the other's fatal gap.
Human Clinical Data
6 complete modalities per patient
Measurements every 6-12 weeks. Only 3-4 data points per patient. Cannot learn continuous tumor dynamics from snapshots.
PDX Mouse Data
Measurements every 2-3 days. Dense longitudinal time-series that reveal growth dynamics, drug response curves, and resistance emergence.
- Mouse stroma replaces human stroma — contaminates signal
- Typically RNA-only (no methylation, no CNV)
- Immunodeficient host — no immune dynamics
Feature completeness across 6 modalities. The biological target for domain alignment.
Clinical predictions rely on PDX-learned physics. PDX relies on human-learned feature imputation.
Dense time-series reveal growth dynamics needed to train neural ODEs.
Dual-Path Architecture
Two paths, one foundation — hover to explore each path
Full multi-modal latent — direct input
UNI2-h histopathology via late gated fusion
ConditionedGatingModelV3 + FiLM + physics bottleneck
RNA-derived portion of VAE latent
DSNHypernetwork + FiLM + BatchNorm
Two paradigms, one foundation
Both paths share the same VAE encoder, serving different clinical and research needs
"Tell me what is happening"
"Tell me when progression occurs"
Multi-Modal VAE v5.10 “TME Boost”
The H-BDVAE compresses all available tumor data into a unified biological latent state (328d) that captures underlying disease biology while factoring out technical artifacts. Uses Product-of-Experts fusion for graceful handling of missing modalities.
Supported modalities
- RNA expression— 2,579 genes, log1p-transformed
- DNA mutations— 500 genes via DeepSomatic + CADD scoring
- Copy number variation— 1,886 genes, z-score standardized
- Methylation— 1,000 probes, beta values
- Histology (WSI)— UNI2-h embeddings — enters via Hypernet late fusion, not VAE
Key innovations
- Additive Decoder Architecture — interpretable reconstruction
- Pathway-guided factorization — 50 MSigDB Hallmark gene sets
- Solved latent collapse — z_meth variance 0.607 (vs ~0 baseline)
Latent Space Structure (328-dim)
Safety & Robustness
Missing Modality Imputation
Probabilistic Encoder Self-Distillation will train a student encoder to match a teacher that sees all modalities. Currently, missing modalities are handled via Product-of-Experts zero-masking.
Uncertainty Quantification
Physics Constraints (Blocker)
All ODE parameters satisfy biological constraints: ρ∈[0,0.3], β∈[0,1], ω>0. 0.00% violation rate. Violations fail regardless of C-index.
Triangulated Validation
Three independent checks — physics compliance, fidelity to data, and clinical utility — must all pass before any prediction is served.
The Model Pipeline
Seven interconnected models from data encoding to evolutionary simulation
H-BDVAE v5.10
Multi-modal encoder with Product-of-Experts fusion
Domain Separation Network
Strips mouse stroma, imputes missing epigenetics from RNA
CausalDriver-GAT
Context-aware driver gene identification via GATv2
TxResponse
Drug sensitivity with concept bottleneck architecture
Hypernetwork v3.2
Dual-path physics-informed parameter generation
Neural ODE
Continuous tumor dynamics with Dopri5 solver
EvoSim
Stochastic clonal evolution with Euler-Maruyama SDE
Glass Box, Not Black Box
Every DNAI prediction decomposes into a chain of inspectable, biologically-named computations. From raw gene expression through 328 named latent dimensions, six physics parameters with physiological units, to time-resolved trajectories with calibrated uncertainty — nothing is opaque.
Explore Prediction TraceabilitySee the evidence
Review validation metrics, benchmark comparisons, and model performance