Inference on biological mechanisms using an integrated phenotype prediction model

We propose a methodology for constructing an integrated phenotype prediction model that accounts for multiple pathways regulating a targeted phenotype. The method uses multiple prediction models, each expressing a particular pattern of gene-to-gene interrelationship, such as epistasis. We also propose a methodology using Gene Ontology annotations to infer a biological mechanism from the integrated phenotype prediction model. To construct the integrated models, we employed multiple logistic regression models using a two-step learning approach to examine a number of patterns of gene-to-gene interrelationships. We first selected individual prediction models with acceptable goodness of fit, and then combined the models. The resulting integrated model predicts phenotype as a logical sum of predicted results from the individual models. We used published microarray data on neuroblastoma from Ohira et al (2005) for illustration, constructing an integrated model to predict prognosis and infer the biological mechanisms controlling prognosis. Although the resulting integrated model comprised a small number of genes compared to a previously reported analysis of these data, the model demonstrated excellent performance, with an error rate of 0.12 in a validation analysis. Gene Ontology analysis suggested that prognosis of patients with neuroblastoma may be influenced by biological processes such as cell growth, G-protein signaling, phosphoinositide-mediated signaling, alcohol metabolism, glycolysis, neurophysiological processes, and catecholamine catabolism.