Gut microbial composition is dynamic in healthy mammalian hosts, however seasonal patterns of compositional changes remain unidentified and uncharacterized on a global level due to the limitations imposed upon traditional non-parametric tests and wave regression methods by the complexity, variability, and sparsity of large human 16s sequencing data. To detect seasonal microbial signatures, a random forest classifier and a transformer neural network (TNN) are developed to input 16s sequencing data from human stool microbiome samples belonging to the American Gut Project and output a prediction for the month of the year that the sample was collected. The TNN’s prediction MAE of 0.92 months for stool samples represented a major improvement over the random forest model’s MAE of 1.93 months. These results indicate the presence of predictable annual variability in population-level microbiome composition and the value of multi-head attention to analyzing microbiome data.