B0282
Title: Gateaux differential-based variable selection for analysis of gene-gene interactions
Authors: Yi Li - University of Michigan (United States) [presenting]
Abstract: Identifying gene-gene interactions is fundamentally important to clarify genetic pathways. Three fundamental principles can be applied: 1) symmetric hierarchy; when an interaction is selected, the lower-order effect should also be present, 2) asymmetric hierarchy; in order to select an interaction, at least one of its parent lower-order effects should also be in the model, and 3) epistasis; interactions can be selected without the presence of their main effects. For complex diseases such as cancers, a variety of genetic alterations interact with each other, and important interactions may not pass lower-order examination. Therefore, asymmetric hierarchy or epistasis is often desirable. However, as the number of potential interactions is ultrahigh, most existing statistical methods lack sufficient computation power. To address this problem, we first proposed a grouped Gateaux differential-based boosting which can be applied to adapt asymmetric hierarchy. The proposed method has the advantage that the candidate set for variable selection is dynamic, which enforces hierarchy. To add more flexibility, we generalized our proposed method with Gateaux differential-based screening to consider epistasis. The proposed algorithm can be shown to take the group LARS as special cases. It is more flexible in that it can be applied to loss functions that are more complex than linear least squares and avoids some limitations inherited with the existing constrained optimization methods.
