Genomic computing. Explanatory analysis of plant expression profiling data using machine learning.

As with every other organism whose genome has been sequenced (Hinton, 1997; Bork et al., 1998), a chief finding in plants (Bevan et al., 1999; Somerville and Somerville, 1999) is the presence of a vast number of genes (many with no relatives in the databases) whose existence, let alone function, had previously gone unrecorded. The importance of finding the function of these genes has led to what amounts to a complete reversal of conventional scientific strategies (Brent, 1999, 2000; Kell and Mendes, 2000), in which one would start with a phenotype (e.g. flower color) and devise experiments that would lead one to the genes whose products were responsible for producing that phenotype. Now, the dawn of the postgenomic era has (consequently) spawned major commercial and academic programs in which plants with more or less defined genotypes (e.g. knockouts; Martienssen, 1998) are being subjected to parallel and high-throughput analyses at the level of the transcriptome (Ruan et al., 1998; Schaffer et al., 2000; Schenk et al., 2000), the proteome (Santoni et al., 1998; Jacobs et al., 2000; Prime et al., 2000; van Wijk, 2000), the metabolome (Oliver et al., 1998; Trethewey et al., 1999; Fiehn et al., 2000; Johnson et al., 2000; Kell and Mendes, 2000; Raamsdonk et al., 2001; Trethewey, 2001), and the phenotype (Rieger et al., 1999), which will provide the wherewithal to assess the contribution of different genes through the activities of their products to the overall functioning of cells and organisms. The problem at hand is then how best to exploit the high-dimensional data floods so generated (e.g. with thousands of gene products or metabolites) for providing the comparatively lowdimensional explanations that we require at higher levels of organization (this gene is or is not important, for example, in cold tolerance). MULTIVARIATE DATA ANALYSIS AND MACHINE LEARNING