A likelihood-free filtering method via approximate Bayesian computation in evaluating biological simulation models

详细信息    查看全文

• 作者：Takanori Hasegawa^a ; ^{t-hasegw@kuicr.kyoto-u.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Atsushi Niida^b ; ^{aniida@ims.u-tokyo.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Tomoya Mori^a ; ^{tmori@kuicr.kyoto-u.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Teppei Shimamura^c ; ^{shimamura@med.nagoya-u.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Rui Yamaguchi^b ; ^{ruiy@ims.u-tokyo.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Satoru Miyano^b ; ^{miyano@ims.u-tokyo.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Tatsuya Akutsu^a ; ^{takutsu@kuicr.kyoto-u.ac.jp" class="auth_mail" title="E-mail the corresponding author} ; Seiya Imoto^b ; ^{imoto@ims.u-tokyo.ac.jp" class="auth_mail" title="E-mail the corresponding author}
• 关键词：Approximate Bayesian computation ; Nonlinear state space model ; Biological simulation ; Gene expression
• 刊名：Computational Statistics & Data Analysis
• 出版年：2016
• 出版时间：February 2016
• 年：2016
• 卷：94
• 期：Complete
• 页码：63-74
• 全文大小：1369 K

文摘

For the evaluation of the dynamic behavior of biological processes, e.g., gene regulatory sequences, we typically utilize nonlinear differential equations within a state space model in the context of genomic data assimilation. For the estimation of the parameter values for such systems, the particle filter can be a strong approach in terms of obtaining their theoretically exact posterior distributions of the parameter values. However, it has some drawbacks for dealing with biological processes in practice: (i) the number of unique particles decreases rapidly since the dimension of the parameter vector and the number of observed time points are higher than its capability, (ii) it cannot be applied when the likelihood function is analytically intractable, and (iii) the prior distributions of the parameter values are often arbitrary determined. To address these problems, we propose a novel method that utilizes the approximate Bayesian computation in filtering the data and self-organizing ensemble Kalman filter in constructing the prior distributions of the parameter values. Simulation studies show that the proposed method can overcome these problems; thus, it can estimate the posterior distributions of the parameter values with automatically setting prior distributions even for the cases that the particle filter cannot perform good results. Finally, we apply the method to real observation data in rat circadian oscillation and demonstrate the usefulness in practical situations.