Posters
« Back
Practical Identifiability Analysis of QSP Models with LikelihoodProfiler
EP34611
Poster Title: Practical Identifiability Analysis of QSP Models with LikelihoodProfiler
Submitted on 24 Dec 2020
Author(s): Ivan Borisov, Evgeny Metelkin
Affiliations: InSysBio
This poster was presented at ACoP11
Poster Views: 79
View poster »


Poster Information
Abstract: Practical Identifiability Analysis of QSP Models with LikelihoodProfiler
Ivan Borisov1
, Evgeny Metelkin1
1
InSysBio, Moscow, Russia
Background: Identifiability analysis is a crucial step in improving reliability and predictability of QSP models.
Various methods of practical identification have been proposed including Monte Carlo simulations, Fisher information
matrix, Profiler Likelihood, etc. Profile Likelihood (PL) is a reliable though computationally expensive approach.
Confidence Intervals by Constraint Optimization (CICO) algorithm based on Profile Likelihood was designed to
reduce computational requirements and increase the accuracy of the estimated parameters confidence intervals.
Objectives: The software package LikelihoodProfiler implementing CICO algorithm was tested on real QSP models
and compared to other approaches to identifiability analysis. The results of these analyses will be presented.
Methods: CICO algorithm has several advantages over original stepwise PL implementations. Stepwise approach
implies visualizing parameters profiles and deriving confidence intervals endpoints as profiles intersections with the
significance level threshold. This approach often produces inaccurate endpoint estimates and is computationally
expensive. CICO algorithm tests parameters identifiability by estimating their confidence interval endpoints with some
preset tolerance. The algorithm does not recover the whole profile to obtain confidence intervals endpoints (or to state
non-identifiability), which results in less likelihood function evaluations and increased performance in comparison
with stepwise PL approach. CICO translates confidence intervals estimation problem into an optimization problem
and solves it by converging to confidence interval endpoint. The algorithm does not require likelihood function
derivatives and enables modeler to choose derivative-free or gradient-based optimization algorithm.
Results: CICO algorithm is proposed for complex QSP models where each likelihood estimation can be
computationally expensive and some parameters are non-identifiable. The algorithm was tested on SB (Systems
Biology) and QSP models described in published studies, including Cancer taxol treatment model [1], TGF- β pathway
model [2], etc. Confidence intervals endpoints estimated with CICO correspond with the values reported in the articles.
It was shown, that on average the algorithm overperforms stepwise PL approach especially in case of non-identifiable
parameters. Detailed analysis of each model can be found on github: https://github.com/insysbio/likelihoodprofilercases. The CICO algorithm is available in a free software package LikelihoodProfiler based on Julia
(https://github.com/insysbio/LikelihoodProfiler.jl) and Python (https://github.com/insysbio/LikelihoodProfiler.py).
Conclusions: The proposed algorithm and software package can be used as a means of identifiability analysis and
confidence intervals evaluation for SB and QSP models of different complexity.
Summary: The proposed algorithm and software package can be used as a means of identifiability analysis and
confidence intervals evaluation for SB and QSP models of different complexity
References: [1] M. C. Eisenberg and H. V. Jain, “A confidence building exercise in data and identifiability: Modeling cancer
chemotherapy as a case study,” J. Theor. Biol., vol. 431, pp. 63–78, Oct. 2017, doi:
10.1016/j.jtbi.2017.07.018.
[2] A. Gábor, A. F. Villaverde, and J. R. Banga, “Parameter identifiability analysis and visualization in large-scale
kinetic models of biosystems,” BMC Syst. Biol., vol. 11, no. 1, p. 54, Dec. 2017, doi: 10.1186/s12918-017-
0428-y.
Report abuse »
Questions
Ask the author a question about this poster.
Ask a Question »

Creative Commons