Multi-Task Learning to Map Proteins Inside Cells at Near-Atomic Resolution

Understanding how proteins interact within the cell to perform specific functions is a major goal of modern biology, and vital for understanding the diverse roles these molecules play in biomedicine. Cryo-electron tomography (cryo-ET) combined with sub-volume averaging (SVA) is currently the most promising imaging technology for visualizing macromolecules within their native environment at high resolution. However, the densely packed cytoplasm makes it difficult to identify the location of proteins of interest within tomograms, and technical challenges in image analysis have limited the applicability of this technique. This project will design computational tools to accurately detect molecules in cellular tomograms and determine their high-resolution structures. To achieve these goals, the team will take a holistic approach to identifying proteins that uses a multi-task, semi-supervised learning strategy to simultaneously de-noise, segment, and detect targets of interest within crowded cellular environments. They will also develop a constrained projection alignment and classification strategy to determine the near-atomic resolution structure of heterogeneous complexes of biomedical importance studied in-situ by cryo-ET. Ultimately, these advances will allow the visualization of molecules in their native environment at unprecedented levels of detail, helping to realize the promise of visual proteomics.