Deep learning for modeling protein atomic structures from cryo-EM density maps
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Proteins are the dynamic molecular machines that perform essential biological processes important for life. While cryo-electron microscopy (cryo-EM) has revolutionized our ability to visualize large macromolecular complexes at near-atomic resolution, accurately interpreting these density maps to build atomic structures remains challenging. This dissertation presents computational methods that address critical challenges in atomic structure modeling from cryo-EM density maps. The primary contribution of this work is the development of Cryo2Struct, a method for atomic structure modeling using a de novo approach that combines 3D transformers with Hidden Markov Models to build atomic structures directly from cryo-EM density maps. This method is further extended in Cryo2Struct2, which incorporates evolutionary information from protein language models and utilizes cryo-EM-based predicted atomic structures as templates for AlphaFold3 to refine and correct protein structures. To support these method development, this dissertation introduces Cryo2StructData, a large labeled cryo-EM dataset for artificial intelligence (AI)-based structure modeling, now publicly available to the scientific community. Additionally, to overcome the inherent noise in cryo-EM data, this work develops and releases datasets specifically designed for AI-based density map enhancement, for improving map interpretability which helps in both manual and automatic atomic structure modeling tasks. This work also contributes to the development of DeepProLigand, a framework for studying protein-ligand interactions using cryo-EM density maps data. The computational methods developed in this dissertation demonstrate substantial improvements in both accuracy and completeness of atomic structure modeling from experimental cryo-EM density map. These contributions advance our ability to determine protein structures from cryo-EM density map, ultimately increasing our understanding of protein functions. All tools and datasets developed through this research are publicly available to promote further scientific advances in this rapidly evolving field.
Table of Contents
PubMed ID
Degree
Ph. D.