Interactive molecular networks obtained by computer-aided conversion of microarray data from brains of alcohol-drinking rats.

ABSTRACT

Lists of differentially expressed genes in a disease have become increasingly more comprehensive with improvements on all technical levels. Despite statistical cutoffs of 99% or 95% confidence intervals, the number of genes can rise to several hundreds or even thousands, which is barely amenable to a researcher's understanding. This report describes some ways of processing those data by mathematical algorithms. Gene lists obtained from 53 microarrays (two brain regions (amygdala and caudate putamen), three rat strains drinking alcohol or being abstinent) have been used. They resulted from analyses on Affymetrix chips and encompassed approximately 6 000 genes that passed our quality filters. They have been subjected to four mathematical ways of processing (a) basic statistics, (b) principal component analysis, (c) hierarchical clustering, and (d) introduction into Bayesian networks. It turns out, by using the p-values or the log-ratios, that they best subdivide into brain areas, followed by a fairly good discrimination into the rat strains and the least good discrimination into alcohol-drinking vs. abstinent. Nevertheless, despite the fact that the relation to alcohol-drinking was the weakest signal, attempts have been made to integrate the genes related to alcohol-drinking into Bayesian networks to learn more about their inter-relationships. The study shows, that the tools employed here are extremely useful for (a) quality control of datasets, (b) for constructing interactive (molecular) networks, but (c) have limitations in integration of larger numbers into the networks. The study also shows that it is often pivotal to balance out the number of experimental conditions with the number of animals.