Inferring weighted gene annotations from expression data

Abstract Annotating genes with information describing their role in the cell is a fundamental goal in biology, and essential for interpreting data-rich assays such as microarray analysis and RNA-Seq. Gene annotation takes many forms, from Gene Ontology (GO) terms, to tissues or cell types of significant expression, to putative regulatory factors and DNA sequences. Almost invariably in gene databases, annotations are connected to genes by a Boolean relationship, e.g., a GO term either is or isn’t associated with a particular gene. While useful for many purposes, Boolean-type annotations fail to capture the varying degrees by which some annotations describe their associated genes and give no indication of the relevance of annotations to cellular logistical activities such as gene expression. We hypothesized that weighted annotations could prove useful for understanding gene function and for interpreting gene expression data, and developed a method to generate these from Boolean annotations and a large compendium of gene expression data. The method uses an independent component analysis-based approach to find gene modules in the compendium, and then assigns gene-specific weights to annotations proportional to the degree to which they are shared among members of the module, with the reasoning that the more an annotation is shared by genes in a module, the more likely it is to be relevant to their function and, therefore, the higher it should be weighted. In this paper, we show that analysis of expression data with module-weighted annotations appears to be more resistant to the confounding effect of gene-gene correlations than non-weighted annotation enrichment analysis, and show several examples in which module-weighted annotations provide biological insights not revealed by Boolean annotations. We also show that application of the method to a simple form of genetic regulatory annotation, namely, the presence or absence of putative regulatory words (oligonucleotides) in gene promoters, leads to module-weighted words that closely match known regulatory sequences, and that these can be used to quickly determine key regulatory sequences in differential expression data.