Bayesian Network Structure Learning For Prediction Of Human Disease Associated To Non-synonymous Variants

Uncovering complex architecture of human genome is an important task in studying human disease. Facilitated by the decades of high-throughput bio-technology, causal variants were accumulated for various types of diseases. The ability to discriminate between pathogenic and benign variants computationally, which select and prioritize likely candidates from a pool of data, could significantly help to target disease-associated variants, then to promote the following study of molecular mechanism and finally to achieve personalized medicine.In current, the major strategy to detect effect of non-synonymous variants is still data-driven based, such as improvement on machine learning algorithms, but rare in biological feature linkages and mechanisms. Here, we focused on correlation and coverage of feature candidates, and choose prediction endpoints by considering both clinical disease classification and underlying similarly mechanism.We perform interpretable damaging effect prediction on the levels of disease-omics and genomics by Bayesian Network Classifier, including178diseases and34890mutations. These classifiers for different diseases were constructed by cross-linked consensus related features such as dScore and PHAT, and also gene function and biological pathways, which were important for complex diseases. The mean prediction accuracy for all diseases is0.92, and all the accuracies range from0.77to0.99. An new feature were test for damaging perdition, termed as PPI-preference, which links to a new type of hub gene that may be important in the disease casual mechanisms.In conclusion, this study build a Bayesian probability based framework to link clinical and biological information, and to combine data driven and hypothesis driven study strategies. This framework may provide a transition but enlightenment method on the journey to personalized medicine.