Implication Networks from Large Gene-expression Datasets

We present a new algorithm for building Boolean networks f rom very large amounts of gene expression data. The resulting networks include not only symmetric relationships between genes, such as co-expression, but also asymmetr ic relations that represent if-then rules. The approach is conceptually simple and fast eno ugh that it can build a complete gene network using 3 billion gene pairs with more than 9,500 e xpression values per genepair in less than 3 hours on an ordinary office computer. The algorithm was applied to publicly available data from thousands of microarrays fo r humans, mice, and fruit flies (for a total of 365 million Affymetrix probeset expression levels). The resulting network consists of hundreds of millions of relationships between genes, and contains biologically meaningful information about gender differences, tissue d iff rences, development, differentiation and co-expression. We also examine re lationships that are conserved between humans, mice, and fruit flies. The full Bool ean relationships are available for exploration at http://gourd.stanford.edu/~sahoo/recomb07/ .