Spatial Analysis of cDNA Microarray Experiments

Some normalization methods assume that the two channels of intensities are related by a constant so that Microarray experiments allow RNA level measurements for many the center of the distribution of the log of the ratio is genes in multiple samples. However, mining the biological information shifted to zero. Chen et al. (1997) proposed an iterative from the large sets of data generated by microarrays requires the use method for estimating that constant. Kerr et al. (2000) of appropriate statistical methods to adjust the observed values for experimentally introduced variability (normalization process) before and Wolfinger et al. (2001) performed the normalization testing differences among samples. Normalization of microarray exwithin the ANOVA. These methods do not consider that periments is a critical step for reducing false positives and false negathe relationship between channels (Cy3 and Cy5) change tives. This paper explores the normalization of cDNA microarray exwith the intensity of the signal. The lowess smoothing periments by a method that uses the blank spot intensity values to normalization method (Yang et al., 2002) performs a make spatial adjustment (SA) of both foreground and background local linear fit and is effective for removing the dye DNA spot intensity values, by fitting an autoregressive mixed linear effect. If the lowess correction is based on all the genes, model through the residual maximum likelihood (REML) methodolthen it should be assumed that most genes are not differogy in the direction of the rows and the columns of the microarray. entially expressed. Application of this spatial normalization to three cDNA array experiThe dye effect is not the only source of systematic ments serves as a case study to validate the SA. Results show that error in microarray experiments. Other spatial effects the spatial analysis allows selection of candidate genes with lesser can cause systematic error and are not considered in numbers of false positive and false negative genes. some standard normalization methods. Several authors have found spatial variation in microarray experiments and considered different spatial models to normalize M are fast becoming an essential tool gene expression in microarray data (Balázsi et al., 2003; in modern biology and the life sciences. Because Wilson et al., 2003; Yang et al., 2003; Wernisch et al., of the complicated steps involved in generating microar2003; Baird et al., 2004). Wernisch et al. (2003) used a ray data and intrinsic experimental variability, microarspatial lowess surface after lowess normalization. Wilray experiments are usually very noisy and have no son et al. (2003) proposed an intensity-based color norfixed scale. Thus, statistical methods for appropriate malization by fitting a single lowess curve to the transexperimental designs and data standardization and norformed data and smoothing the residuals with a median malization are essential for identifying spots with true filter to estimate spatial trend. They found that a 3-by-3 differential expression when genes are applied to the block of spots is appropriate, but this may not be the array. case in other arrays, which introduces the problem of In the process of printing microarrays, of obtaining, defining the block size. Hoffmann et al. (2002) compared labeling, and hybridizing RNA, and of reading and different normalization procedures and statistical analyquantifying the results, several experimental factors can ses for detecting differentially expressed genes and concause systematic spatial variability within and among cluded that normalization had a severe influence on the arrays and slides, thus distorting the estimation of gene detection of genes with different expressions. expression patterns and contrasts. Some of these factors Mixed linear models with a specific variance–covariinclude unequal amounts of DNA deposited at each ance structure can be used to control the spatial variabilspot, the intensity-dependent dye effect, reverse tranity within and among slides of cDNA microarray experiscription efficiency, and settings and calibration of the ments. In this study, we model the spatial variability laser scanner (Schuchhardt et al., 2000). A good experiof microarray experiments by considering the spatial mental design, if taken into account when designing the allocation of the spots within the slide. The separable autoregressive correlation (AR) structure proposed by slides, can reduce the impact of some of these disturCullis and Gleeson (1991) and Gilmour et al. (1997), bance factors. Unless these factors are appropriately extensively used in agriculture field experiments, was accounted for in the statistical model, they will signifiused for studying spatial variability in the slide of cDNA cantly reduce the precision of estimates of gene expresmicroarray experiments. Baird et al. (2004) used this sion and gene expression contrasts. SA for normalizing microarray experiments including all the spots on the array. We applied this SA to three J. Burgueño and J. Crossa, Biometrics and Statistics Unit, Internadifferent sets of cDNA microarray experiments, but tional Maize and Wheat Improvement Center (CIMMYT), Apdo. used only the blank spots present on the array. Mixed Postal 6-641, 06600 México DF, México; D. Grimanelli, O. Leblanc, linear models based on SA adjusted data were used as and D. Autran, Institut de Recherche pour le Développement (IRD), an a priori procedure before selecting the candidate Apdo. Postal 57297, 06501 México DF, México. Received 24 June 2004. Genomics, Molecular Genetics & Biotechnology. *Correspondgenes with significant expressions. Also, we used the ing author ([email protected]). Abbreviations: AR, separable autoregressive correlation; GA, global Published in Crop Sci. 45:748–757 (2005). © Crop Science Society of America adjustment; LA, lowess adjustment; REML, residual maximum likelihood; SA, spatial adjustment. 677 S. Segoe Rd., Madison, WI 53711 USA 748 Published online February 23, 2005