Genome-wide association studies have had great success in identifying human genetic variants associated with disease, disease risk factors, and other biomedical phenotypes. Many variants are associated with multiple traits, even after correction for trait-trait correlation. Discovering subsets of variants associated with a shared subset of phenotypes could help reveal disease mechanisms, suggest new therapeutic options, and increase the power to detect additional variants with similar pattern of associations. Here we introduce two methods based on a Bayesian framework, SNP And Pleiotropic PHenotype Organization (SAPPHO), one modeling independent phenotypes (SAPPHO-I) and the other incorporating a full phenotype covariance structure (SAPPHO-C). These two methods learn patterns of pleiotropy from genotype and phenotype data, using identified associations to discover additional associations with shared patterns. The SAPPHO methods, along with other recent approaches for pleiotropic association tests, were assessed using data from the Atherosclerotic Risk in Communities (ARIC) study of 8,000 individuals, whose gold-standard associations were provided by meta-analysis of 40,000 to 100,000 individuals from the CHARGE consortium. Using power to detect gold-standard associations at genome-wide significance (0.05 family-wise error rate) as a metric, SAPPHO performed best. The SAPPHO methods were also uniquely able to select the most significant variants in a parsimonious model, excluding other less likely variants within a linkage disequilibrium block. For meta-analysis, the SAPPHO methods implement summary modes that use sufficient statistics rather than full phenotype and genotype data. Meta-analysis applied to CHARGE detected 16 additional associations to the gold-standard loci, as well as 124 novel loci, at 0.05 false discovery rate. Reasons for the superior performance were explored by performing simulations over a range of scenarios describing different genetic architectures. With SAPPHO we were able to learn genetic structures that were hidden using the traditional univariate tests.