A1464
Title: Bayesian variable selection for multi-layer biological data
Authors: Hao Cheng - University of California Davis (United States) [presenting]
Abstract: Advances in high-throughput sequencing technology have generated an increasing volume and diversity of multi-omics data that complement genomics. The effects of genetic variants (e.g., SNPs) on phenotypes can be mediated by multiple layers of molecular variations through mechanisms such as regulatory cascades from the genome to the transcriptome and proteome. These molecular variations serve as measurable intermediates between DNA and phenotype and are partially heritable across generations. In most models incorporating intermediate molecular variations for complex trait prediction, these variations are typically treated as independent variables alongside SNPs or as responses in addition to empirical complex traits in mixed models. This approach overlooks the sequential relationships between genetic variants, intermediate molecular variations, and complex traits. A Bayesian variable selection method is developed that extends mixed models into multilayer neural networks to capture the nonlinear relationships between genotypes and phenotypes, improving prediction and inference for complex traits. The neural network is further developed, named NNMM (a mixed-effect neural network), to incorporate intermediate omics features into the middle layers of the network. This enables mechanistic modeling of the regulatory pathways from genotypes, through intermediate omics features, to phenotypes.
