Extracting the Structure and Conformations of Biological Entities from Large Datasets

Abstract

In biology, structure determines function, which often proceeds via changes in conformation. Efficient means for determining structure exist, but mapping conformations continue to present a serious challenge. Single-particles approaches, such as cryogenic electron microscopy (cryo-EM) and emerging "diffract & destroy" X-ray techniques are, in principle, ideally positioned to overcome these challenges. But the algorithmic ability to extract information from large heterogeneous datasets consisting of "unsorted" snapshots - each emanating from an unknown orientation of an object in an unknown conformation - remains elusive. It is the objective of this thesis to describe and validate a powerful suite of manifold-based algorithms able to extract structural and conformational information from large datasets. These computationally efficient algorithms offer a new approach to determining the structure and conformations of viruses and macromolecules. After an introduction, we demonstrate a distributed, exact k-Nearest Neighbor Graph (k-NNG) construction method, in order to establish a firm algorithmic basis for manifold-based analysis. The proposed algorithm uses Graphics Processing Units (GPUs) and exploits multiple levels of parallelism in distributed computational environment and it is scalable for different cluster sizes, with each compute node in the cluster containing multiple GPUs. Next, we present applications of manifold-based analysis in determining structure and conformational variability. Using the Diffusion Map algorithm, a new approach is presented, which is capable of determining structure of symmetric objects, such as viruses, to 1/100th of the object diameter, using low-signal diffraction snapshots. This is demonstrated by means of a successful 3D reconstruction of the Satellite Tobacco Necrosis Virus (STNV) to atomic resolution from simulated diffraction snapshots with and without noise. We next present a new approach for determining discrete conformational changes of the enzyme Adenylate kinase (ADK) from very large datasets of up to 20 million snapshots, each with ~104 pixels. This exceeds by an order of magnitude the largest dataset previously analyzed. Finally, we present a theoretical framework and an algorithmic pipeline for capturing continuous conformational changes of the ribosome from ultralow-signal (-12dB) experimental cryo-EM. Our analysis shows a smooth, concerted change in molecular structure in two-dimensional projection, which might be indicative of the way the ribosome functions as a molecular machine. The thesis ends with a summary and future prospects.