Relevance Networks

Relevance networks allow one to build networks of features, whether they represent genes, phenotypic or clinical measurements. The technique works by first comparing all features with each other in a pairwise manner, similar to the initial steps of dendrogram construction. Typically, two genes are compared with each other by plotting all the samples on a scatterplot, using expression levels of the two genes as coordinates. A correlation coefficient is then calculated, though any dissimilarity measure can be used. A threshold value is then chosen, and only those pairs of features with a measure greater than the threshold are kept and graphically displayed.

We applied relevance networks, a method that we developed to find networks of correlated variables, to a pharmacogenomic data set, generating hypotheses of putative functional relationships between pairs of genes and pharmaceuticals. In collaboration with the Whitehead Institute and the National Cancer Institute, we used baseline RNA expression levels of 6,701 genes measured from the NCI60, a set of 60 human cancer cell lines.46 To this data, we joined a database of measures of cancer susceptibility to 4,991 anti-cancer agents, to see how the baseline RNA expression levels in the cell lines correlated with the inhibition of growth of these same cell lines to thousands of anti-cancer agents. We found only one significant association between a gene’s expression and measures of an anti-cancer agent’s susceptibility: increased expression of LCP1 is associated with increased susceptibility to the anti-cancer agent NSC 624044, a thiazolidine carboxylic acid derivative. LCP1 has been found in many types of tumors, and other thiazolidine carboxylic acid derivatives are known to inhibit tumor cell growth, but there was previously no known relationship between this drug and gene in the biomedical literature. We published this work in the Proceedings of the National Academy of Science in 2000.