Mimosa: mixture model of co-expression to detect modulators of regulatory interaction

Functionally related genes tend to be correlated in their expression patterns across multiple conditions and/or tissue-types. Thus co-expression networks are often used to investigate functional groups of genes. In particular, when one of the genes is a transcription factor (TF), the co-expression-based interaction is interpreted, with caution, as a direct regulatory interaction. However, any particular TF, and more importantly, any particular regulatory interaction, is likely to be active only in a subset of experimental conditions. Moreover, the subset of expression samples where the regulatory interaction holds may be marked by presence or absence of a modifier gene, such as an enzyme that post-translationally modifies the TF. Such subtlety of regulatory interactions is overlooked when one computes an overall expression correlation. Here we present a novel mixture modeling approach where a TF-Gene pair is presumed to be significantly correlated (with unknown coefficient) in a (unknown) subset of expression samples. The parameters of the model are estimated using a Maximum Likelihood approach. The estimated mixture of expression samples is then mined to identify genes potentially modulating the TF-Gene interaction. We have validated our approach using synthetic data and on three biological cases in cow and in yeast. While limited in some ways, as discussed, the work represents a novel approach to mine expression data and detect potential modulators of regulatory interactions.