NEW COMPUTATIONAL METHODS FOR 3D STRUCTURE DETERMINATION OF MACROMOLECULAR COMPLEXES BY SINGLE PARTICLE CRYO-ELECTRON MICROSCOPY

Abstract

In cryo-electron microscopy (cryo-EM) single-particle reconstruction, projection images of many first order structural identical copies of the specimen of interest are combined to recover the underlying three-dimensional object. The analyzed particles are embedded in vitreous ice in a hydrated state at many random orientations and projection images are recorded using a transmission electron microscope (TEM). The heterogeneity of two-dimensional projection image data resulting from the co existence of different conformational or ligand binding states of a macromolecular complex remains a major obstacle as it impairs the validity of reconstructed density maps and limits the progress towards higher resolution. Nevertheless, single particle cryo-EM is the method of choice for structure determination of large macromolecular complexes that are difficult to purify in the amounts and quality needed for X-ray crystallization. In contrast to other structure determining techniques, single particle cryo-EM is not an ensemble technique. Images of individual macromolecular complexes are recorded that can potentially be used to detect structural variations within a single dataset which is highly important for structure determination of large and dynamic macromolecular assemblies. However, computational tools to sort large datasets of images into subpopulations representing images that belong to a unique 3D object are still missing. In structural analysis of macromolecular assemblies with a high degree of structural dynamics, separation of conformations is a challenging task because for each particle image there are six degrees of freedom (translations and rotations), an unknown number of conformations and a high level of noise that have to be considered simultaneously. Classical random conical tilt reconstruction, utilizing additional experimental information by recording of tilted image pairs generates single class sum starting models, limited by a missing cone of information and a low Signal to Noise ratio originated at the limited number of class members and thus projection views used for reconstruction. A new method is presented, based on acquisition of tilted image pairs under cryo conditions facilitating classification of both untilted and tilted raw images. In these double-classified tilted pairs in combination with their experimental tilt angular relationships of the classified particles, angular relationships relative to each other are conserved. This information can be utilized to compute multi class sum starting models without a missing cone. Additionally, the information obtained by classification and tilt relationships can be used to separate particles in different conformations into individual models. The novel strategy is intended to gain insight into the 3D structures and structural diversity of so far poorly characterized dynamic assemblies. Increased computational demands, originated at the increased number of projection images to process in order to analyze conformational differences, are addressed and strategies to overcome these using recent developments in information technology such as Multicore-, Parallel- and GPU- processing are presented. Each parallel processing paradigm alone can improve overall performance, while combining all paradigms, unleashes the full power of today's technology. The increased computational performance thus makes certain applications feasible that were virtually impossible before.