Abstract:
|
Gene-environment interactions have important implications for complex disease such as cancer and diabetes. It is challenging to analyze interactions due to the high dimensionality and low signal levels. Given the lack of information, incorporating additional information is desired. However, most of existing methods ignore such information and treat genetic factors equally a priori. We propose a penalized approach that is customized to incorporate additional information for identifying important genetic main effects and hierarchical interactions. Under a marginal analysis framework, the proposed method adopts minimax concave penalty for regularized estimation and Laplacian quadratic penalty based on additional information. Here, the additional information can be the adjacency structure in certain chromosome, correlation structure among gene expressions, and data extracted from existing literature. Extensive simulation shows our proposed approach outperforms multiple alternatives in marker identification. In the analysis of TCGA melanoma and GENEVA diabetes data, the proposed method demonstrates practical applicability and provides sensible findings.
|