Workflow to Mine Frequent DNA Co-methylation Clusters in DNA Methylome Data

The advances in high-throughput nucleotide sequencing technology revolutionized biomedical research. Vast amount of genomic data rapidly accumulates in a daily basis, which in turn calls for the development of powerful bioinformatics tools and efficient workflows to analyze them. One of the approaches to address the "big data" issue is to mine highly correlated clusters/networks of biological molecules, which may provide rich yet hidden information about the underlying functional, regulatory, or structural relationships among genes, proteins, genomic loci or various types of biological molecules or events. A network mining algorithm lmQCM has recently been developed, which can be applied to mine tightly connected correlation clusters (networks) in large biological data with big sample size, and it guarantees a lower bound of the cluster density. This algorithm has been used in a variety of cancer transcriptomic data to mine gene co-expression networks (GCNs), but it can be applied to any correlational matrix. lmQCM is available through R package lmQCM as well as the online tool package TSUNAMI ( https://biolearns.medicine.iu.edu ). In this study, the purpose is to establish an analytical workflow to apply lmQCM for frequent (consensus) cluster mining in multiple DNA methylation datasets in different cancers and extract the underlying common co-methylation networks for genes.Specifically, the workflow is applied to analyze DNA methylome data across different cancer types using lmQCM. It mines frequent DNA methylation clusters based on individual clustering mining results, identifying common as well as distinctive DNA methylation patterns among different cancer types. This workflow has successfully identified frequent GCNs in 33 types of cancers, thus proven to be a powerful tool to analyze large biological data. It helps to identify common features as well as distinctions among different diseases, disease subtypes, or among different biological processes. The resulted frequent clusters may provide new insights on the pathway/function networks. In the case of disease study, the results lead to new directions for biomarker and drug target discovery. The advantages of this workflow include the highly efficient processing of large biological data generated from high-throughput experiments, quick identification of highly correlated interaction networks, substantial reduction of the data dimensionality to a manageable number of variables for downstream comparative analysis, and consequently increased statistical power for detecting differences between conditions.