Conditional independence relations among biomarkers under heterogeneity

Molecular biomarkers are currently used in cancer research for diagnostic and predictive purposes. The study of conditional independence relations among biomarkers allows to reduce costs of assessement by identifying highly informative subprofiles. Moreover it is a step towards the definition of cancer bioprofiles when a target variable has to be predicted on a molecular basis. The growing interests in Bayesian Networks (BNs) is partially due to the possibility of addressing probabilistic inference, classification and prediction in a unified framework which at one time provides efficient computational algorithms and semantic tools which are increasingly exploited in biomedicine and health care. BNs are also useful to reason about data, even if causality is involved. In this paper we perform structural learning of Bayesian Networks on a recently published case study. Conditional independence relations among five biomarkers and age are investigated under different structural hypotheses by scoring candidate structures with the BDe metric. Our results suggest that introducing a latent variable a better explanation of observed data is obtained. Moreover, by investigating triplo-negative configuration of biomarkers we found the marginal probability of the latent-grouping variable which defines groups of potentially heterogeneous individuals.