Simultaneous Structure Identification and Parameter Estimation of Gene Regulatory Networks

摘要

Much work has been done on the automated recovery of gene network structure. This paper presents a hybrid approach based on Genetic Programming (GP) and Particle Swarm Optimization (PSO). As a proof-of-concept, this hybrid algorithm attempts to discover the structure and dynamic model parameters of a gene regulatory network similar to that of flowering time control in Arabidopsis. To provide known ground truth, a network that mimics key features of the real Arabidopsis flowering time control genetic network is first synthesized and parameterized before hand. Synthetic bolting date data are then generated for 100 genetic lines based on multiple planting dates, daily temperatures, and day lengths (photoperiod) at 18 European sites that span the latitudinal range of the real species. Genetic lines (genotypes) are each described by 100 markers, assumed for prototyping purposes to equate to "genes". Using the dates, environmental data, and genotypes, the task is to identify (1) how many and (2) which of the genes comprise the network, (3) the functional type of each gene, (4) the network parameter values associated with the genes, and (5) the network structure. A GP -PSO hybrid algorithm is implemented for this purpose. Preliminary results are presented and discussed.