Dysbiosis, departure of the gut microbiome from a healthy state, has been suggested to be a powerful biomarker of disease incidence and progression(1-3). Diagnostic applications have been proposed for inflammatory bowel disease diagnosis and prognosis(4), colorectal cancer prescreening(5) and therapeutic choices in melanoma(6). Noninvasive sampling could facilitate large-scale public health applications, including early diagnosis and risk assessment in metabolic(7) and cardiovascular diseases(8). To understand the generalizability of microbiota-based diagnostic models of metabolic disease, we characterized the gut microbiota of 7,009 individuals from 14 districts within 1 province in China. Among phenotypes, host location showed the strongest associations with microbiota variations. Microbiota-based metabolic disease models developed in one location failed when used elsewhere, suggesting that such models cannot be extrapolated. Interpolated models performed much better, especially in diseases with obvious microbiota-related characteristics. Interpolation efficiency decreased as geographic scale increased, indicating a need to build localized baseline and disease models to predict metabolic risks.