Authors:

Jason Lin

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The global rise in aging populations poses a substantial healthcare and economic challenge. In response, "aging clocks" have been developed over the past decade to estimate biological age and aging rates independently of chronological age. These clocks aim to predict mortality risk, identify aging biomarkers, and uncover underlying mechanisms. However, current models largely rely on methylation, transcriptomic, and genomic data, often excluding metabolomics due to the limited availability of large-scale metabolomics datasets. In this project, we address this gap by leveraging metabolomics datasets from human studies to explore the intersection of metabolism and aging using machine learning (ML) and deep learning (DL) approaches. We utilized metabolomics datasets from the Wisconsin Registry for Alzheimer's Prevention (WRAP; PMID: 29322089) to build models that predict both chronological and biological age. To enhance model interpretability and biological relevance, we developed a biologically-informed neural network that integrates curated metabolic pathway data from resources such as the Relational Database of Metabolomics Pathways (RaMP) and the Human Metabolome Database (HMDB). We did this through directly integrating the pathway information into the architecture of the neural network itself by sparsely connecting metabolites nodes and pathway nodes based on biologically validated pathway data rather than relying on fully connected layers of multi-layer perceptrons. Our pathway-informed model (R2: 0.414, MAE: 4.06) outperformed conventional ML and DL models that rely solely on metabolites in predicting chronological age such as MLP, LASSO, Elastic Net, and XGBoost across 10-fold cross validation. We also found that this model performed well on longitudinal datapoints of succeeding visits of WRAP study individuals removed from the initial dataset (MAE: 2.25), also outperforming traditional MLPs (MAE: 2.43). We are currently extending this approach to predict aging-related phenotypes, including physical activity levels and cognitive function based on LIBRA and CHAMPs scores as well as laboratory based health measurements. Importantly, by applying SHAP, DeepLIFT, and perturbation analyses, we identified key metabolite biomarkers and associated pathways 66 predictive of aging. Among these were known aging-related pathways such as nicotinate and sulfur amino acid metabolism, highlighting the potential use of these models to uncover novel metabolic pathways relevant to aging biology. Overall, the integration of biological information such as pathway information with machine learning serves as an effective means to improve the predictive power and explainability of metabolomics-based models. In the future, we plan to test our model on an independent human metabolomics dataset and also experimentally validate the significance of metabolic pathways identified in model organisms. Jean Marseille:

Source:

Brown / Advanced Undergraduate Research Fellowships

Topics:

No topics listed

Co-authors:

Jason Lin