Integration of prior biological knowledge and epigenetic information enhances the prediction accuracy of the Bayesian Wnt pathway

Abstract

Computational modeling of the Wnt signaling pathway has gained prominence for its use as a diagnostic tool to develop therapeutic cancer target drugs and predict test samples as tumorous/normal. Diagnostic tools entail modeling of the biological phenomena behind the pathway while prediction requires inclusion of factors for discriminative classification. This manuscript develops simple static Bayesian network predictive models of varying complexity by encompassing prior partially available biological knowledge about intra/extracellular factors and incorporating information regarding epigenetic modification into a few genes that are known to have an inhibitory effect on the pathway. Incorporation of epigenetic information enhances the prediction accuracy of test samples in human colorectal cancer. In comparison to the Naive Bayes model where β-catenin transcription complex activation predictions are assumed to correspond to sample predictions, the new biologically inspired models shed light on differences in behavior of the transcription complex and the state of samples. Receiver operator curves and their respective area under the curve measurements obtained from predictions of the state of the test sample and the corresponding predictions of the state of activation of the β-catenin transcription complex of the pathway for the test sample indicate a significant difference between the transcription complex being on (off) and its association with the sample being tumorous (normal). The two-sample Kolmogorov–Smirnov test confirms the statistical deviation between the distributions of these predictions. Hitherto unknown relationship between factors like DKK2, DKK3-1 and SFRP-2/3/5 w.r.t. the β-catenin transcription complex has been inferred using these causal models.