With the genomes of several organisms sequenced, one goal is to use high throughput approaches to elucidate the function of genes in a global and systematic way. A typical procedure for inferring gene function is to identify sets of genes whose expression profiles are highly similar and then predict the function of uncharacterized genes based on the functions of the other genes in the set. I will discuss two different approaches for extracting reliable functional information about genes from diverse collections of microarray data. Both use a different form of prior knowledge to leverage the analysis.

The first approach, called the gene recommender, identifies new candidate genes related to a set of genes of interest. It was inspired by collaborative filtering methods that recommend books and movies. The method takes advantage of the idea that most genes of related function tend to be co-regulated in only a subset of the experiments represented in the database. By implementing a simple feature-selection method and then building a cluster around the input set using only the selected experiments, the method successfully identified new genes likely to be involved in a tumorogenesis pathway.

The second approach uses microarray data from multiple organisms to infer gene function. The most reliable gene-expression relationships may be those detectable in multiple organisms. Not only have such relationships been reproduced in independent experiments, they have also survived evolutionary selective pressures. I will present a method that identifies these relationships from DNA microarray data and then uses them to predict gene function. The results suggest combining data across multiple species can improve the functional annotation of conserved genes over using data from any single species alone.