EvoSim
ODE-coupled stochastic ensemble for clonal evolution under treatment
Shannon entropy matches clinical samples
Matches multi-region sequencing patterns
No explosions in 10,000 runs
Phylogeny reconstruction accuracy
Overview
An ODE-coupled stochastic ensemble that models how tumor subpopulations compete, evolve, and respond to treatment. The Neural ODE first learns patient-specific growth, drug-response, and competition parameters from data — then EvoSim runs 100 stochastic simulations using those parameters to produce a distribution of possible outcomes. The result is not a single trajectory but a full uncertainty picture: median trajectory with 5th-95th percentile bands, clone extinction and dominance probabilities, and regimen-aware resistance onset timing. Clones can be initialized from multi-region sequencing or phylogenetic data for maximum biological fidelity.
Inputs
4 inputstrajectory[T, K]Deterministic clone trajectories from Neural ODE
sigma[1]Stochastic noise scale from Hypernet
mutation_rate[1]Per-cell mutation probability
fitness_landscape[K, M]Fitness effects of M potential mutations
Outputs
6 outputsensemble_trajectory[N, T, K']100-run ensemble of clone trajectories with median and 5th/95th percentile bands
clone_extinction_prob[K]Per-clone probability of extinction across ensemble runs
clone_dominance_prob[K]Per-clone probability of becoming the dominant subpopulation
resistance_onset[K]Predicted time to resistance per clone (net growth > 0 under active treatment)
clone_treeGraphPhylogenetic tree of clone relationships
tumor_burden[T]Total tumor burden distribution (median + percentile bands)
Mathematical Formulation
Euler-Maruyama discretization with confidence cones (5th-95th percentile)
Multiplicative fitness from mutations
Poisson mutation process
Key Features
- Neural ODE-learned patient-specific parameters (rho, beta, omega)
- 100-run stochastic ensemble with Euler-Maruyama integration
- Gillespie SSA for exact stochastic simulation in low-population regimes
- Hybrid simulator auto-switches SDE↔SSA at population threshold
- Median trajectory with 5th/95th percentile uncertainty bands
- Clone extinction and dominance probability estimation
- Regimen-aware resistance onset detection
- Phylogenetic initialization from multi-region sampling
Key Innovations
- 1Coupling learned ODE parameters with stochastic simulation
- 2Ensemble-based outcome distributions instead of single trajectories
- 3Regimen-aware resistance: net growth under active treatment > 0
- 4Phylogenetic-initialized clone definitions from real sequencing data
Hyperparameters
0.01 (days)From Hypernet1e-6 per cell per day100 cells5010 cellsTraining Details
EvoSim is a simulation module, not independently trained. Patient-specific growth, drug-response, and competition parameters are learned by the Neural ODE from omics data. EvoSim uses those parameters to run a 100-member stochastic ensemble, producing outcome distributions. Validated against multi-region sequencing data.