Variance-based sensitivity analysis for stochastic chemical kinetics

TL;DR

This paper introduces new unbiased schemes for variance-based sensitivity analysis of stochastic chemical kinetics models, demonstrating significant improvements in estimator accuracy and variance reduction through novel methods and simulation techniques.

Contribution

It proposes new estimation schemes for sensitivity indices in stochastic models and analyzes their variance properties, improving accuracy over previous methods.

Findings

New schemes reduce estimation error significantly.

RTC algorithm yields over 30 times lower variance than Gillespie's method.

Variance of estimators depends on reaction order in simulation methods.

Abstract

Sensitivity analysis is a process of computing sensitivity indices, which are certain measures of importance of parameters in influencing the outputs of mathematical models. Sensitivity indices computed in variance-based sensitivity analysis yield quantitative answers to questions like how much on average the variance of model output, measuring its uncertainty, decreases, if exact values of certain unknown parameters are determined, e. g. in an experiment. We propose new schemes for estimation of variance-based sensitivity indices of outputs of stochastic models, their conditional expectations, and histograms given the parameters. Unbiased estimators obtained in these schemes can be used in a Monte Carlo (MC) procedure approximating sensitivity indices. We derive relations between variances of final estimators of MC procedures making the same number of evaluations of given function, but using different schemes, both for the newly introduced schemes and for some used before in the literature. Numerical experiment for a discrete state stochastic Markov model of a chemical reaction network (DM) shows that our method can lead to much lower error than method analogous to the one offered by Degasperi et al. Further numerical experiments demonstrate that the application of random time change (RTC) algorithm due to Rathinam et al. for simulation of DM can lead to over 30 times lower variance of estimators of certain sensitivity indices than when Gillespie's direct (GD) method is used, and that this variance may significantly depend on the order of reactions in GD method. We provide some intuitions explaining these effects. We generalize measures used for comparing dispersion of different distributions, such as coefficient of variation and Fano factor to the random parameters case, in a way that they can be computed along with variance-based sensitivity indices.