An Approach Of Epistasis Mining Based On K-tree Optimizing Bayesian Network

Analyzing genetic mechanism of complex diseases has always been an pressure and challenging topic.In recent years,finding epistatic interaction among loci has been more imperative.In terms of missing heritability problem,poor robustness,high false positive rate,low computational efficiency in existing methods,Meanwhile,consider that the k-tree algorithm can cover the whole genome data by uniformly sampling data and Bayesian network fully reflect the causal relationship among loci,perform well for high-order interactions.This research proposed precisely and approximately epistasis detection methods which sample k-tree to optimize Bayesian network.Thus,large-scale of epistatic loci for whole-genome data be detected efficiently.The article includes the following three aspects of research:(1)The uniform sampling dandelion code better reflect the bijective relationship between graph and coding.It transforms the network construction including large-scale of SNPs and phenotype trait into the problem of specific coding generation.Uniformly sampling dandelion codes could ensure better coverage of the whole genome SNP loci.We construct the k-tree including a large number of SNPs and phenotype trait through uniformly sampling dandelion codes which can better cover the SNP loci of the whole genome.Then the decomposition algorithm based on neighbor nodes used to decompose the graph corresponding to k-tree into different k-cliques.(2)According the sub-networks corresponding to different k-cliques.Our research propose approximately(Ktree BN)and precisely(KIBEpi)epistatic detection methods.The Ktree BN method utilizes the improved Fast-IAMB(omb-Fast)to learn the sub-Bayesian network structure corresponding to k-cliques quickly.The KIBEpi method regards the construction of Bayesian network through different k-cliques as integer linear programming problems.Cutting plane method,sub-IP and other methods utilized to obtain the global optimal sub-Bayesian network.Finally,we merge the sub-networks to obtain the whole network.Successfully finding the epistatic loci which affects the phenotypic traits.(3)Epistatic data be generated by GAMETES software which commonly used in epistatic research.Ktree BN and KIBEpi methods proposed are compared with the commonly mining methods of epistatic.The experiment results show that Ktree BN and KIBEpi methods have higher accuracy,F1-score and lower false positive rate compared to other existing algorithms.In addition,the Age-related Macular Degeneration data used to verify the perform of Ktree BN and KIBEpi methods.The results show that Ktree BN and KIBEpi methods mining more effective epistasis loci compare to other epistatic methods.