Forma, Vol. 12 (No. 2), pp. 123-142, 1997

A Neural Network Modeling for Concentration-Dependent Pair-Rule Gene Expression

Yasuhiko Takeda and Yoh Iwasa

Department of Biology, Faculty of Science, Kyushu University, Fukuoka 812-81, Japan

(Received April 25, 1996; Accepted June 28, 1996)

Keywords: Gene Regulation, Generalization, Fuzzy-Neuro Network

Abstract. Four stripes of gene expression of pair-rule genes, controlling Drosophila segmentation in the developmental process, are modeled by neural networks. In the learning procedures, the expression patterns of gap genes (giant, hunchback, Krüppel and knirps) at the beginning of cellularization of syncitial blastoderm embryo for various mutants and wild type are used as the input of the network, and the disturbed and wild type expression patterns of even-skipped 2-nd and 3-rd stripes and hairy 5-th and 6-th stripes are used as the output of the network. The performance of the learned networks is examined by [1] rotationplot (a density plot of output intensity on a phase plane of input factors) and by [2] mutagenesis (perturbed spatial patterns when some of the response units for the input genes are deleted). To model the gene regulatory region, we study a type of fuzzy-neural network and compare it with ordinary Parallelly Distributed Processing (PDP) two-layered type network. In the fuzzy-neuro type network, the convergence of learning procedure was faster than in the ordinary network, and concentration-dependent rules of gene expression, characterized from molecular experimental results, are fairly accurately reproduced.