Optimal Design of Experiments for Simulation-Based Inference of Mechanistic Acyclic Biological Networks

Abstract

Biological signaling pathways based upon proteins binding to one another to relay a signal for genetic expression, such as the Bone Morphogenetic Protein (BMP) signaling pathway, can be modeled by mass action kinetics and conservation laws that result in non-closed form polynomial equations. Accurately determining parameters of biological pathways that represent physically relevant features, such as binding affinity of proteins and their associated uncertainty, presents a challenge for biological models lacking an explicit likelihood function. Additionally, parameterizing non-closed form biological models requires copious amounts of data from expensive perturbation-response experiments to fit model parameters. We present an algorithm (SBIDOEMAN) for determining optimal experiments and parameters of systems biology models with implicit likelihoods. We evaluate our algorithm using simulations of held-out true parameter values and demonstrate an improvement in the rate of accurate parameter inference over random and equidistant experimental designs when evaluated on two simple models of the BMP signaling pathway with an implicit likelihood function.

arXiv commentAccepted to Learning Meaningful Representations of Life workshop at NeurIPs 2021

References in corpus (12)

• The frontier of simulation-based inference
• Normalizing Flows for Probabilistic Modeling and Inference
• Fast $\epsilon$-free Inference of Simulation Models with Bayesian Conditional Density Estimation
• Automatic Posterior Transformation for Likelihood-Free Inference
• Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows
• Equivariant Hamiltonian Flows
• Neural Spline Flows
• Benchmarking Simulation-Based Inference
• A Unified Stochastic Gradient Approach to Designing Bayesian-Optimal Experiments
• Bayesian Experimental Design for Implicit Models by Mutual Information Neural Estimation
• E(n) Equivariant Normalizing Flows
• Gradient-based Bayesian Experimental Design for Implicit Models using Mutual Information Lower Bounds