Phenome-wide association studies (PheWAS) are instrumental in investigating genotype-phenotype correlations in large-scale biobanks. Recently, several PheWAS resources have become available, utilizing electronic health records and genotypes predominantly from European populations. However, current resources have certain limitations, such as the lack of clinically relevant phenotype definitions, underrepresentation of non-European populations and a limited scope of laboratory measurements. Also, none of these resources have distinguished between the phenotypic consequence of loss-of function (LoF) and gain-of-function (GoF) variants. To bridge these gaps, we developed a comprehensive PheWAS resource containing ancestry-specific and panancestral analysis results. We utilized The Mount Sinai BioMe BioBank comprising African, European and Admixed American participants, and predicted LoF and GoF missense variants obtained from LoGoFunc, a machine learning classifier tailored to predict LoF, GoF and neutral variants. Utilizing two whole-exome sequencing cohorts of BioMe, including 27,739 and 14,186 participants, we performed association testing of genome-wide predicted LoF, GoF and neutral variants across nearly 2,000 binary and quantitative phenotypes. Our analyses revealed population-specific associations for numerous variants and genes, underscoring the importance of considering variant effect and genetic diversity in disease genomics to advance precision medicine.