Abstract
Mass spectrometry-based metabolomics is a powerful analytical technique used for identifying small molecules in complex biological samples. However, current data acquisition methods have limitations in capturing all relevant molecules. To address this issue, we propose using artificial intelligence (AI) to optimize mass spectrometry data acquisition in real-time, maximizing the number and quality of identified metabolites. First, large amounts of publicly available mass spectrometry data will be used to develop a deep neural network that can predict the quality of generated fragmentation spectra based on instrument configurations. Second, we will use offline reinforcement learning to explore novel instrument configurations to enhance the data acquisition process. A critical focus will be placed on defining a suitable reward function that guides the AI agent's exploration, considering factors such as spectrum quality, novelty of acquired spectra, and resource utilization. Third, we will use a virtual mass spectrometry environment to simulate the fragmentation process and allow the AI agent to control data acquisition. This will enable thorough assessment and comparison against baseline approaches and alternative strategies. Once fully trained and validated, the AI agent will be deployed onto a mass spectrometer to autonomouslycontrol the data acquisition process in real time, evaluating its performance in detecting putative metabolites compared to traditional approaches. By utilizing AI to optimize molecular discovery from untargeted metabolomics experiments, we will enhance the identification of metabolites that were previously overlooked, unlocking valuable biological insights. These advances will have transformative implications for precision medicine, drug discovery, and many other areas of the life sciences.
Researcher(s)
Research team(s)
Project type(s)