Hybrid Model-Based and Data-Driven Frameworks for High-Resolution Tomographic Imaging - ABSTRACT
The ultimate goal of structural biology is to visualize biomolecules in action in their native environment and to
establish their structure-function relationship. Cellular cryo-electron and X-ray tomography have emerged as
powerful techniques for imaging complex biological samples such as intact cells, organelles, macromolecular
machines, and for quantifying the internal organization of biological objects in their native states in situ at
resolutions ranging from a few microns to tens of nanometers with X-rays to tens of angstroms with electrons.
However, compared to the mature techniques of X-ray crystallography and single-particle cryo-electron
microscopy, cellular tomography is yet to reach its potential due to a severe degradation in the resolution of
reconstructions because of the effects of mechanical misalignment and non-rigid sample deformation due to
radiation damage, missing-wedge artifacts, low signal-noise ratio in a crowded environment, and unresolved
conformational heterogeneity. To address these issues, we propose to leverage our new approach for automated
joint 3D alignment and regularized reconstruction that combines advances in iterative projection methods and
convex optimization to achieve better than state-of-the-art reconstruction resolution from severely misaligned
data. Infusing our framework with new advances in mathematical modeling and machine learning provides a
clear path to a host of new model-based and data-driven algorithms that could address current challenges and
bottlenecks in the analysis of cellular tomography data. In particular we propose to (1) develop techniques for
improved tilt-series alignment that account for rigid-body motion of the sample and recover the anisotropic
effects of radiation-induced warping by using optical flow alignment; (2) develop a decoder that leverages the
full frequency information contained in randomly oriented macromolecules in the cell volume to constrain the
effects of the missing-wedge; and (3) improve the resolution of subtomograms extracted from the refined volume
by developing a volume-encoder--deformation-decoder deep neural network to model conformational
heterogeneity. By developing new data-driven methods that constrain the missing-wedge information and treat
shape variability as a continuum of non-rigid deformations rather than discrete clusters, our algorithmic
framework will provide significant improvements in the resolution and quality of reconstructions over currently
existing methods for data analysis that neglect these effects. As the structural biology community is increasingly
focusing on cellular tomography, there is a growing need for easy to use, automated software amenable to both
experienced and novice users. (4) To this end, algorithms resulting from this proposal will be turned into GPU-
enabled open-source user-friendly software to accelerate the analysis of the growing pool of imaging data.
Ultimately, our algorithmic framework will be capable of yielding high-resolution structures from noisy,
incomplete and complex data, thereby enhancing the predictive power of cellular tomography to answer
important biological questions.
