Source-linked AI summary
Data-driven Effective Modeling of Stochastic Chemical Reaction Networks
Yuan Chen, Weize Mao, Dongbin Xiu
TL;DR
SSA provides exact stochastic reaction-network trajectories but can be computationally costly because it resolves microscopic reaction events. This paper learns the finite-time transition kernel from short SSA bursts using a conditional normalizing flow, producing a coarse stochastic propagator. Across numerical examples, the model reduces simulation steps while accurately reproducing the true dynamics’ statistical properties.
Problem
SSA suffers from high computational cost, motivating a finite-time modeling approach that avoids resolving every microscopic reaction event.
Method
The method learns the finite-time transition kernel of the SSA-induced continuous-time Markov chain from short SSA data using a conditional normalizing flow.
Results
The learned propagator generates statistically consistent long-time trajectories at a constant coarse time step while significantly reducing simulation steps compared with SSA.
Takeaways & Limitations
The approach provides a data-driven effective model that reproduces statistical properties of stochastic reaction dynamics at a user-defined coarse time step.
Takeaways & Limitations
Conservation of molecule numbers is enforced by construction, while other invariants, including signal temporal logic requirements, are left for future work.
Abstract
from arXiv · showhide
The Stochastic Simulation Algorithm (SSA), widely considered an exact algorithm for stochastic chemical reaction networks, suffers from high computational cost. In this work, we propose a data-driven effective model that operates on a user-defined coarse time step independent of the underlying microscopic reaction-event scale. This is accomplished by directly approximating the finite-time transition kernel of the continuous-time Markov chain induced by SSA, using a generative machine learning model trained on short bursts of SSA simulation data. The trained model constructs a stochastic propagator that recursively generates statistically consistent trajectories at the constant coarse time step, with significantly reduced computational cost. In this paper, we employ conditional normalizing flow as the stochastic propagator. A comprehensive set of numerical examples is presented to demonstrate the accuracy and efficiency of the proposed method.
1. Introduction.
SSA generates exact trajectories for stochastic reaction networks but can be computationally expensive because it resolves microscopic reaction events. The paper instead learns finite-time transition behavior from SSA data to build a coarse stochastic propagator using conditional normalizing flows.
- SSA exactly generates sample trajectories of the continuous-time Markov jump process underlying stochastic chemical kinetics.
- SSA can require extremely many reaction events when populations are large or reaction rates are stiff, increasing long-time simulation cost.
- The proposed approach directly approximates the finite-time transition kernel of the SSA-induced Markov process instead of approximating individual reaction events.
- The learned stochastic propagator operates at a user-defined coarse time step that can substantially exceed the average SSA firing time.
- Conditional normalizing flows model the highly state-dependent, non-Gaussian transition distributions of stochastic chemical kinetics.
- Numerical examples show statistically consistent trajectories while significantly reducing simulation steps relative to SSA.
2. Preliminary.
The paper models well-mixed stochastic reaction networks as continuous-time Markov chains whose exact trajectories can be generated by SSA. Its central modeling target is the finite-time transition kernel, approximated from SSA dynamics at a prescribed coarse time step.
- 2.1. Problem Setup.: A stochastic chemical reaction network contains N species and M reaction channels, with molecule counts forming the state vector X(t).
- 2.1. Problem Setup.: Under the well-mixed assumption, the reaction network evolves as a continuous-time Markov chain governed by stoichiometric state changes and propensity functions.
- 2.1. Problem Setup.: SSA generates exact sample trajectories by simulating the reaction events of this continuous-time Markov chain.
- 2.1. Problem Setup.: For a finite lag ∆, P∆ gives the conditional distribution of X(t + ∆) given X(t) = x and defines the associated Markov semigroup.
- 2.1. Problem Setup.: The paper aims to approximate P∆ from data using a prescribed coarse step ∆ chosen independently of microscopic SSA firing times, with ∆≫E[τ] possible.
- Existing model-based approaches are less capable for SSA-generated processes because their transition distributions are non-Gaussian and lack a standard parametric form.
3. Main Method: Effective Stochastic Propagator.
The method learns a stochastic flow map that approximates the finite-time SSA transition kernel and recursively propagates states at a constant, user-defined coarse time step. It trains a conditional normalizing flow on short SSA trajectory bursts, with enforcement operators preserving physical state constraints.
- Finite-Time Stochastic Flow Map: The model approximates the finite-time transition kernel by mapping each coarse-step state and an auxiliary Gaussian sample to a stochastic successor.The learned flow map is intended to reproduce the conditional distribution of the SSA process over interval ∆.
- Recursive System Prediction: Recursive application generates discrete-time trajectories whose finite-time statistics approximate those of SSA while using a constant coarse step independent of microscopic firing times.This reduces the number of simulation steps required for long-time simulations.
- Short-Burst Data Generation: Training uses independent state pairs generated from short SSA trajectory bursts at the prescribed coarse time step.Short bursts avoid extremely long SSA simulations and are naturally parallelizable.
- Conditional Normalizing Flow Approximation: A conditional normalizing flow parameterized by a neural network represents the state-dependent stochastic map and is trained by maximizing conditional likelihood.The invertible transformation enables tractable likelihood evaluation through the change-of-variable formula.
- Computational Details for Enforcing Operator: Post-processing projects real-valued flow outputs onto non-negative integer molecule counts to keep generated states physically admissible.The operator applies componentwise nonnegativity enforcement followed by nearest-integer rounding.
4. Numerical Examples.
Numerical experiments evaluate sFML’s computational efficiency, one-step transition-kernel accuracy, and long-time trajectory statistics across several reaction networks. The learned propagator uses substantially fewer simulation steps than SSA and adaptive tau-leaping while producing close agreement with SSA reference behavior and statistics.
- Computational efficiency: sFML takes substantially fewer simulation steps than SSA and adaptive tau-leaping across the tested reaction-network examples.sFML uses exactly T/∆ steps, whereas SSA and adaptive tau-leaping step counts are averaged over simulated trajectories.
- Distributional accuracy: One-step distributions from the learned stochastic flow map are compared with SSA transition kernels using MMD at 30 conditioning states per example.For each selected state, 10,000 conditional samples are generated from both SSA and sFML; minimum, median, and maximum MMD values are reported.
- Global trajectory accuracy: Global trajectory errors compare sFML with fixed-step tau-leaping at the same coarse step size τ = ∆ using mean and standard-deviation curves over 10,000 trajectories.These errors quantify discrepancies in ensemble mean and standard-deviation trajectories over the full discrete rollout.
- Transfer Process: For the transfer process, predicted trajectories and ensemble statistics closely agree with SSA reference statistics through T = 10.0.The test uses initial condition (83, 26, 69) and statistics averaged over 10,000 trajectories.
- Lotka-Volterra Model: For Lotka-Volterra tests, learned trajectories and mean and standard deviation are in close agreement with reference results across two parameter regimes.The second regime also shows consistent qualitative behavior and accurate learned marginal and conditional distributions relative to SSA.
- Periodic reaction networks: For the Brusselator, Autocatalysis, and Oregonator, dominant Fourier modes of learned and SSA trajectories agree closely, including the Oregonator comparison.The Brusselator and Oregonator examples exhibit periodic or nearly periodic behavior, and the learned model captures the corresponding oscillation frequencies.
5. Conclusion.
The proposed effective model learns a stochastic propagator from SSA data to generate long-time trajectories at a constant coarse step while reproducing the true dynamics’ statistical properties. Numerical examples demonstrate its effectiveness across trajectory and Fourier-mode comparisons.
- The model directly approximates the SSA-induced transition kernel from short bursts of SSA data using a generative model.
- Example 4.4: For Example 4.4, the paper compares two independent SSA trajectories with two trajectories generated by the trained model through T = 10.
- Example 4.4: For Example 4.4, Fourier modes are compared between an SSA trajectory and the corresponding model trajectory.
- A conditional normalizing flow realizes the stochastic propagator for recursively generating coarse-step trajectories.
- The propagator uses a user-defined constant coarse time step independent of microscopic firing times, significantly reducing simulation steps while accurately reproducing statistical properties.
- Example 4.5: For Example 4.5, state-space, phase-space, and Fourier-mode comparisons are made between SSA and trained-model trajectories up to T = 6.