Genome-wide Association Studies (GWAS) have resulted in many discovered risk variants for several obesity-related traits. However, before clinical relevance of these discoveries can be achieved, molecular or physiological mechanisms of these risk variants needs to be discovered. One strategy is to perform data mining of phenotypically-rich data sources such as those present in dbGAP (database of Genotypes and Phenotypes) for hypothesis generation. Here we propose a technique that combines the power of existing Bayesian Network (BN) learning algorithms with the statistical rigour of Structural Equation Modelling (SEM) to produce an overall phenotypic network discovery system with optimal properties. We illustrate our method using the analysis of a candidate SNP data set from the AMERICO sample, a multi-ethnic cross-sectional cohort of roughly 300 children with detailed obesity-related phenotypes. We demonstrate our approach by showing genetic mechanisms for three obesity-related SNPs.