Large-scale simulation-based experiments with stochastic models using machine learning-assisted approaches : Applications in systems biology using Markov jump processes

Abstract: Discrete and stochastic models in systems biology, such as biochemical reaction networks, can be modeled as Markov jump processes. The chemical master equation describes how the probability distribution of a biochemical system's states evolves. Unfortunately, solutions to the chemical master equation only exist for trivial problems. However, the stochastic simulation algorithm (SSA) can generate exact sample paths. Large-scale simulation-based experiments involving variations to the model's parameters are computationally intensive and hinder modelers from exploring and inferring their models due to high-dimensional models.This thesis proposes methodologies and tools for model exploration and approximate parameter inference of high-dimensional stochastic models simulated via the SSA.  We propose a smart computational workflow using machine learning-assisted approaches to enable model exploration of gene regulatory networks where the objective is to assess different qualitative behaviors present in the model. An artificial neural network is proposed for learning summary statistics used in approximate parameter inference.  The neural network can find distinct local features from multivariate time series, enabling more complex models involving several biological species. By introducing epistemic uncertainty, we further explore Bayesian neural networks for approximate parameter inference. A classification approach is introduced, which learns the proposal posterior by an adaptive sampling scheme, ultimately reducing the number of simulations required for the inference task. We have also developed the software package Sciope to support modelers with machine learning-assisted techniques for model exploration and parameter inference. Sciope also comes with various features, such as experimental designs, traditional ABC algorithms, and a parallel backend to scale large simulation-based experiments from laptops to the cloud.Finally, to reduce the gap between modelers and biologists, StochSS Live! has been developed. StochSS Live! is a user-friendly web-based platform that enables any practitioners to build biochemical reaction models and perform simulation by ensemble analysis, model exploration, and approximate parameter inference.