Inference of Gene Pathways Using Gaussian Mixture Models

Identification of gene-gene interactions and complete characterization of gene pathways are critical in understanding the transcript processes underlying biological processes. Bayesian network is a powerful framework to infer gene pathways. We developed a novel Bayesian network, in which we use Gaussian mixture models to describe continuous gene expression data and learn gene pathways. Mixture parameters were estimated using an EM algorithm, while the optimal number of mixture component for each gene node and the optimal network topology best supported by the data were identified using the Bayesian Information criterion (BIC). We applied the proposed approach to a histone pathway in yeast and to a less explored circadian rhythm pathway in honeybee. The performance of the proposed approach was compared against alternative Bayesian network algorithms that either discretize the gene expression information or use single distribution instead of mixtures. Evaluation shows that our approach outperforms other approaches in terms of more accurate inference of the known network and can effectively predict gene pathways with different topology using continuous data. In addition, the estimated mixture model can facilitate an intuitive description of the gene node behavior, thus enhancing the interpretation of the inferred network.