On Tuesday 8th August I-Chun Chou, Ph.D. Candidate at the Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University in Atlanta, will give a talk on Parameter Estimation with Alternating Regression: Applications to S-system Models and S-distributions
Time: 13:15 - 14:00
Place: UMB, Biotechnology Building 3rd floor, room BT3A-11
I-Chun Chou |
Novel high-throughput techniques are capable of producing dense time series data that are beginning to open up new modeling strategies. These data implicitly contain enormous information about the biological systems they describe, such as their functional connectivity and regulation. However, the identification of local features from “global data” is challenging, and parameter estimation continues to be the bottleneck of the computational analysis of biological systems and other applications. As more time series data are being generated and systems are becoming larger and more complex, it is therefore necessary to develop improved methods that are effective, fast, and scalable. In this seminar, I will present a new method of
alternating regression (AR) that provides a genuinely new tool for estimating S-system parameters from time series data [1]. The AR method is similarly beneficial for the identification of structure and regulation in S-system models of ill-characterized pathway systems, as well as for the estimation of S-distribution parameters from statistical data. The key feature of AR is the reduction of the nonlinear inverse problem of parameter estimation into iterative steps of linear regression. The method begins with the substitution of differentials with estimated slopes, which ultimately replaces the
n original differential equations with
n sets of
N algebraic equation. The iterative linear regression is initiated for each set of algebraic equations, corresponding to one differential equation, by first guessing all
ßi and
hij values of the equation and subsequently obtaining the parameters of the
ai-term through simple linear regression. The resulting estimates for
ai and all
gij are used for the next iteration, where now the parameters of the
ßi-term are estimated. The method thus switches back and forth, thereby rapidly improving estimates of all parameters. Because each step is linear, AR is extremely fast compared to nonlinear regression with differential equations. A remaining, challenging issue is convergence. While I will present insights into this issue, it is not yet fully understood. I will show with several artificial examples that the method works very well if it converges. In cases where the iterative process does not converge, the enormous speed of the method renders it feasible to select several initial guesses and search settings as an effective countermeasure. Furthermore, growing experience is beginning to suggest some heuristic rules for identifying initial settings.
[1] Chou IC, Martens H, and Voit EO (2006). Parameter Estimation in Biochemical Systems Models with Alternating Regression. BMC Theoretical Biology and Medical Modelling: accepted for publication.
