A Transformer-Based Foundation Model for Human Microbiome Analysis

Abstract

The human microbiome plays a crucial role in maintaining our health. Alterations in the microbiome have been linked to various chronic conditions like autoimmune disorders, metabolic diseases, and cancer. While various tools have been developed to study the microbiome, each tool tends to be specialized for a specific task. To overcome this limitation, we report on the development of a foundation model pretrained on 13,524 human microbiome metagenomic samples. The model was then fine-tuned to predict the clinical status of the host. Our model was able to differentiate between healthy and diseased samples in 10-fold cross-validation on the training dataset with an accuracy of 83.7%. On an external validation dataset of 927 samples, our model had an accuracy of 74.9%. Notably, our model performed even better at differentiating diseases from one another. On the diseased samples in the training dataset, it classified samples with an accuracy of 93.3% in 10-fold cross-validation. Together, our results show that generative AI has the potential to transform microbiome research and advance personalized medicine.