Unmasking conditional dependencies of proteins influencing islet biology using machine learning

PROJECT SUMMARY/ABSTRACT Abstract: High-risk polymorphisms for diabetes account for only a fraction of the overall disease heritability, implying significant conditional dependencies regulate the influence genes exert on islet function. Machine learning (ML) methods can identify such dependencies from omics data. I will apply machine learning (ML)- based methods to identify conditional dependencies regulating islet function. I will derive and validate ML models predicting islet function from proteins measured in islets of Diversity Outbred (DO) mice (AIM 1). For models with high predictive accuracies in both mouse and human data, I will use the proteins’ model weights as meta-traits to identify genomic regions altering those proteins’ influence on insulin secretion. One protein my ML models and prior studies predict to alter insulin secretion is transketolase (TKT), an enzyme in the pentose shunt. Knockdown of TKT in mouse and human islets increases insulin secretion. Since the pentose shunt is not thought to be a key pathway in β-cells, I will characterize a novel function of TKT and will use ML to determine its conditional dependence on other proteins. Using constructs restricting TKT activity and localization, I will establish whether TKT’s enzymatic activity and nuclear localization are required for its suppression of insulin secretion (AIM 2). Using these constructs, I will identify TKT’s binding partners in the cytosol and nucleus by immuno-precipitating the compartment-restricted TKT followed by mass spectrometry (IP/MS), filtering for non-specific proteins using ML-based likelihood scoring of the identified proteins. Collectively these studies will provide novel models for interrogating protein conditional dependencies in islets and identify partners for TKT that could offer potential new drug targets. Training plan/career goals: My career goal is to become an independent NIH-funded investigator at a top-tier research institution. Identifying conditional dependencies in complex data is key for understanding metabolic disorders. My mentors, Drs. Craven and Smith are experts in ML and proteomic analyses, respectively. Other advisors have specific areas of focus tailored to sub-components of the grant’s respective aims. Implementing ML in biological data is potentially powerful but not straightforward. The proposed courses and workshops and working directly with my mentors’ labs will provide conceptual and programmatic background for implementing these methods. UW-Madison is an ideal location for this because of the strong support for learning and applying new computational methods. In particular, the Center for High-Throughput Computing offers free 24/7 access to clusters of high-end processors and the necessary memory and professional support for working with these complex data. They also offer training for computing languages in which many of these newer methods are implemented. Collectively, the proposed project training will enable me to identify conditional dependencies for metabolic regulators beyond the specifics of these studies and provide an essential foundation for me to transition to independence.