Constructing cancer patient-specific and group-specific gene networks with multi-omics data.

Cancer is a complex and heterogeneous disease with many possible genetic and environmental causes. The same treatment for patients of the same cancer type often results in different outcomes in terms of efficacy and side effects of the treatment. Thus, the molecular characterization of individual cancer patients is increasingly important to find an effective treatment. Recently a few methods have been developed to construct cancer sample-specific gene networks based on the difference in the mRNA expression levels between the cancer sample and reference samples. We constructed a patient-specific network with multi-omics data based on the difference between a reference network and a perturbed reference network by the patient. A network specific to a group of patients was obtained using the average change in correlation coefficients and node degree of patient-specific networks of the group. In this paper, we present a new method for constructing cancer patient-specific and group-specific gene networks with multi-omics data. The main differences of our method from previous ones are as follows: (1) networks are constructed with multi-omics (mRNA expression, copy number variation, DNA methylation and microRNA expression) data rather than with mRNA expression data alone, (2) background networks are constructed with both normal samples and cancer samples of the specified type to extract cancer-specific gene correlations, and (3) both patient individual-specific networks and patient group-specific networks can be constructed. The results of evaluating our method with several types of cancer show that it constructs more informative and accurate gene networks than previous methods. The results of evaluating our method with extensive data of seven cancer types show that the difference of gene correlations between the reference samples and a patient sample is a more predictive feature than mRNA expression levels and that gene networks constructed with multi-omics data show a better performance than those with single omics data in predicting cancer for most cancer types. Our approach will be useful for finding genes and gene pairs to tailor treatments to individual characteristics.