Moment-based parameter inference with error guarantees for stochastic reaction networks

TL;DR

This paper introduces a moment-based parameter inference method for stochastic reaction networks that provides theoretical error guarantees and does not require likelihood computation or simulation.

Contribution

It offers a novel convex optimization approach to bound parameters with uncertainty quantification, applicable to complex biochemical models with steady-state and time-resolved data.

Findings

Provides bounds on parameters with error guarantees

Avoids likelihood computation and simulation

Demonstrates effectiveness on synthetic biochemical data

Abstract

Inferring parameters of models of biochemical kinetics from single-cell data remains challenging because of the uncertainty arising from the intractability of the likelihood function of stochastic reaction networks. Such uncertainty falls beyond current error quantification measures, which focus on the effects of finite sample size and identifiability but lack theoretical guarantees when likelihood approximations are needed. Here, we propose a method for the inference of parameters of stochastic reaction networks that works for both steady-state and time-resolved data and is applicable to networks with non-linear and rational propensities. Our approach provides bounds on the parameters via convex optimisation over sets constrained by moment equations and moment matrices by taking observations to form moment intervals, which are then used to constrain parameters through convex sets. The bounds on the parameters contain the true parameters under the condition that the moment intervals contain the true moments, thus providing uncertainty quantification and error guarantees. Our approach does not need to predict moments and distributions for given parameters (i.e., it avoids solving or simulating the forward problem), and hence circumvents intractable likelihood computations or computationally expensive simulations. We demonstrate its use for uncertainty quantification, data integration and prediction of latent species statistics through synthetic data from common non-linear biochemical models including the Schl\"ogl model and the toggle switch, a model of post-transcriptional regulation at steady state, and a birth-death model with time-dependent data.