Efficient Computational Methods for Cryo-Electron Microscopy Single Particle Reconstruction
Addressing the Problem
Cryo-electron microscopy (EM) single particle reconstruction (SPR) is an emerging technology for studying the 3D structures of macromolecules. Without the need of crystallization, cryo-EM SPR has great potential to unveil the protein structures at their native states and help biologists to study the functions of those macromolecular machines. In the experiments, the purified proteins are flash frozen at liquid ethane temperature into a very thin film of vitreous ice layer, capturing the native state at the moment of freezing. The micrographs that contain the 2D projections of the particles are then collected from the electron microscope. Individual particles are imaged only once at unknown random orientations. Since the projection images are extremely noisy, it requires collecting and analyzing a large amount of images to achieve a structure with near-atomic resolution. Despite the recent progress, there remain many challenges for the computational SPR methods. Open questions include how to efficiently extract information from the high-throughput noisy data and handle the compositional and conformational heterogeneity of the molecules.
The goal of this research is to develop a new computational framework for efficient cryo-EM SPR. We propose a new approach that circumvents the difficult orientation estimation problem for ab initio reconstruction in cryo-EM. This new approach has the potential to succeed below the critical signal-to-noise ratio for current methods and scales linearly with the number of images. Advanced machine learning and optimization techniques will be applied to reduce the overall computational complexity for denoising, initial model determination, and iterative refinement. We will apply the new algorithmic pipeline to high-throughput experimental data in cryo-EM to discover new macromolecular structures.