CRCNS: Repurposing Transformers to Characterize Gut Microbiome & Aging Brain Axis

For debilitating neurodegenerative diseases such as Alzheimer’s disease (AD), some drugs have been approved but only offer a modest attenuation of cognitive decline. There is a need to better understand the role of modifiable risk factors which could decrease the risk of AD and other types of dementia. The human gut microbiome–comprising of microbes and their associated genes–shows differences in composition between healthy individuals and those with AD dementia, and based on recent studies appears to be a promising modifiable risk factor. Despite encouraging findings, large gaps remain regarding our knowledge of the gut microbiome’s relationship with the aging brain and dementia pathology measured via neuroimaging, fluid biomarkers, and cognitive assessments. Fortunately, datasets that can enable this analysis are growing in terms of sample sizes, making a comprehensive machine-learning based investigation of these associations timely and potentially highly rewarding. To this end, our team proposes to make use of and extend contemporary developments in Transformer-based models that power advancements in large language model (LLM) technologies, and repurpose them into specialized models for these analyses for multi-modal datasets. The key components of this project include the following aims. Aim 1: We will design special Transformer-based models that will be trained using microbiome data via attaching novel trainable modules. These repurposed models will enable analysis of associations between gut microbiome composition and brain health in dementia. Aim 2: Our models will be extended to handle 3D neuroimaging data with an eye on computational efficiency. The models will be rigorously evaluated on large imaging datasets as well as on multi-modal datasets. Their performance/generalization behavior will be profiled. Aim 3: The models will be used to study the relationship between the gut microbiome and the brain, anchored by eight specific scientific hypotheses. Pretrained models and software tools will be disseminated to the community. Significance: This project will yield sophisticated methods for multimodal analysis of microbiome, brain imaging and clinical/cognitive data for aging and Alzheimer’s disease. These tools will be disseminated and can be used to investigate a variety of scientific hypotheses to determine novel relationships between the microbiome and the brain. This in turn is expected to open new avenues for preventing and treating neurodegenerative disease and improve the health and well-being of aging Americans.