Large scale genetic studies for numerous traits have implicated genetic variants across the genome. This is well illustrated through the recent genome wide association study (GWAS) of 1.2 million individuals leading to the discovery of 406 loci associated with tobacco and alcohol use (1). Yet, there still lies a knowledge gap of the functional mechanisms and biological etiology behind the variant-phenotype associations found. With the advances in modern genomic technologies, datasets often gather a multitude of biological measures, including DNA genotypes, RNA expression and epigenetic information (2). Association studies using these integrative datasets will not only implicate associated genes, but also reveal underlying mechanisms for diseases (3). Here, we look to understand the mechanisms behind gene regulation that influence complex traits. This will be accomplished by evaluating current transcriptome wide association study (TWAS) methods and developing a new multi-omic TWAS framework. Transcriptome wide association studies (TWAS) are a popular approach to multi- omic integrative methods. To date, TWAS methods rely on statistical models to predict gene expression due to a lack of individual level expression data available. These predicted expression values are further correlated to the phenotype of interest to see which genes (and their expression) effect the trait. Mainstream TWAS methods use genetic variants that are within a 1Mb distance from the gene as predictors for the gene's expression. This heuristic definition of cis-regulatory region is often not optimal and fails to pinpoint the true set of eQTLs. Alternatively, we propose using chromatin conformation data and enhancer activity marks to select the genetic variants used for transcriptome prediction. This will use molecular and spatial knowledge to have a biologically informed method in selecting genetic variants for expression prediction model training. Furthermore, we apply our approach to a cross-tissue framework for gene expression prediction, as well as an ensemble approach to produce optimal models from either cross-tissue or single tissue measures based on gene context. We implemented this suite of frameworks to 13 human tissues in the Genotype-Tissue Expression project and compare our findings to previous methods. Our approach resulted in an average gain of 52% and 14% more significant imputed models and an average of 44% and 5% improvement in prediction accuracy when compared to two widely-used methods: PrediXcan and UTMOST, respectively. Finally, we apply our expression prediction models to the genome-wide association results of the largest smoking and drinking use cohort to highlight our methods advantages for analyzing complex traits. We present that the improved expression prediction accuracy from multi-omic integration leads to increased power to detect gene-trait associations. Ultimately, this dissertation highlights the use of integrative approaches for genomic association studies. This dissertation provides a foundation for future epigenetic integration for association studies, and emphasizes that multi-omic approaches will substantially improve our understanding of the mechanisms behind complex traits.