A Bayesian parameter estimation and model selection tool for molecular cascades: LetItB (Let It found by Bayes)

Abstract

Event Abstract Back to Event A Bayesian parameter estimation and model selection tool for molecular cascades: LetItB (Let It found by Bayes) Junichiro Yoshimoto1*, Tomofumi Inoue1 and Kenji Doya1 1 Okinawa Institute of Science and Technology, Japan Introduction Mathematical models of molecular cascade are essential for elucidation of dynamic mechanisms of cellular functions. In constructing such models, an important issue is selection of the model structure and setting of the model parameters, which are often hand-tuned to replicate the known system behaviors by trial-and-error. This is not only time-consuming but also compromise the objectiveness of the results. A theoretically sound and computationally efficient framework to identify them is crucial for making the model-based studies reliable and productive. We previously proposed a Bayesian estimation method to solve the issue (Yoshimoto et al., 2007). Here we present a multi-platform software tool with graphical user interface software to the Bayesian estimation engine, called LetItB (Let It found by Bayes). Method We model the chemical reactions within a compartment of interest by a set of ordinary differential equations (ODEs) specifying the concentration changes in the molecular species. The measurements are given by a subset of molecular species at certain time points subject to corruption by noise. In the Bayesian framework, we assume a prior distribution over the parameters of the ODEs and the observation, such as the reaction constants and the variance of measurement noise. Given a set of measurements, the likelihoods of different sets of parameters are evaluated by simulation and the posterior distribution of parameters is computed by integration with the prior distribution. Among a number of methods for Bayesian inference, we adopt Metropolis-Hasting sampling algorithm for our Bayesian estimation engine. Our software tool provides the functions of 1) import/export of cascade models in SBML, 2) graph representation of the model structure and editing the parameters, 3) simulation and visualization of the model dynamics, 4) import and visualization of experimental time series, 5) estimation and visualization of the posterior distribution of the model parameters, 6) evaluation of marginal likelihood for model selection. We used public software libraries: libSBML and SBML ODE Solver for SBML interface, Sundials and GNU Scientific Library for numerical simulation, and QT/QWT for graphic user interface (Top Panel in Fig. 1). LetItB can be compiled and run on multiple platforms including Linux, Mac OS X, and Windows. Results We tested the estimation engine and the user interface of LetItB with a number of benchmark problems. The bottom panel in Fig. 1 shows the screenshots for the multiple functions of LetItB. In its application to (an enzymatic reaction) model, the maximal a posterior estimate of the parameters by LetItB was more accurate than those found by conventional genetic algorithms, but we could also visualize the confidence intervals and the dependency between parameters by showing the posterior distribution. It was also verified that the marginal likelihoods computed for a number of candidate models can be used for selection of a model best suited to the given set of observation time courses. Conclusions As a solution of system identification problem in molecular cascade modeling, we developed a Bayesian estimation engine and combined it with a graphic user interface so that experimentalists without computational background can utilize the Bayesian system identification framework. LetItB is available from http://www.nc.irp.oist.jp/software/. To enhance the computational efficiency, we are incorporating parallelizable Bayesian sampling methods and implementing the LetItB engine for parallel computers.