Experimental Design for Time-Series Microarray Analysis

Microarray experiments designed to analyze time-series gene expression levels are important to interpret the bacterial gene functions. For this purpose, two-color microarray is indispensable tool for viewing the gene expression changes as snapshot on a large scale. In two-color microarray experiments, differently labeled cDNAs derived from control and target samples are hybridized on an identical slide at the same time so that the differences between mRNA quantity of control and target samples can be compared. Generally, variances of intensity value obtained from microarray experiments increase when the intensity is large [1]. In microarray data, log-transformation of the intensity is performed for the elimination of the inconsistency of variance [2], and the difference of expression level in a gene is estimated by the log-ratios of intensities between a target and a control samples. Therefore for reliable time-series analysis using multiple microarrays, it is necessary to control the variance so that the log-ratio displays expression changes more precisely. It is considered that improvement of quality of data by choosing appropriate experimental design is crucial for accurate time-series microarray analysis. Previously, Yang and Speed [3] proposed the experimental design for time-series microarray analysis and they evaluated synthetically the practicability at the view point of variability, mRNA resources and so on. However the evaluation of experimental design based on real data has not been performed. The goal of our study is to propose a reliable experimental design for consecutive time-series microarray analysis. Initially, we examine reference design, that is, the control sample is collected at one representative time point. This design is widely used for multivariate analysis using multiple microarrays because of their simplicity that enables us to compare each variable, but there are two difficulties. First, changes of neighboring time-point are measured indirectly via control sample. So the consequential changes could not be detected obviously. Secondly, hybridization has to be performed between two time-points whose gene expression states are largely different. Bacterial gene expression is drastically altered when it happened to encounter the environmental stresses including nutrient limitation, low temperature and growth to stationary phase. In this case, variances of intensity values derived from target sample differ from that of control sample, which would produce larger variance in log-ratio values. To estimate log-ratio values, normalization process, which removes the bias of different labels, is performed under assumption that most of all gene expression levels are equivalent [4]. To compensate these disadvantages, we adopted the sequential design, that is, two samples are collected at the closest two time points. Based on this design, gene expression changes between time-point ti and ti+1 can be measured directly. So it is expected that more precise data will be obtained which indicates the consequential changes.