Empowering Biomolecular Modeling with Neural Networks and Empirical Evidence: An Integrative Approach to Hypothesis Generation

Computational biophysics is a powerful tool for observing and understanding the microscopic machinery that underpins the biological world. Molecular modeling and simulations can help scientists understand a cell’s behavior and the mechanisms that drive it. Empirical evidence can provide information on the structure and organization of biomolecular machines, which serve as the backbone of biomolecular modeling. Experimental data from probing the cell’s inner workings can provide modelers with an initial structure from which they can hypothesize and independently verify function, complex formation, and response. Additionally, molecular data can be used to drive simulations toward less probable but equally interesting states. With the advent of machine learning, researchers now have an unprecedented opportunity to take advantage of the wealth of data collected in a biomolecular experiment. This dissertation presents a comprehensive review of atomistic modeling with cryo-electron microscopy and the development of new simulation strategies to maximize insights gained from experiments. The review covers the integration of cryo-EM and molecular dynamics, highlighting the evolution of their relationship and the recent history of MD innovations in cryo-EM modeling. It also covers the discoveries made possible by the integration of molecular modeling with cryo-EM. Next, this work presents a method for fitting small molecules into cryo-electron microscopy maps, which uses neural network potentials to parameterize a diverse set of ligands. The method obtained fitted structures commensurate with, if not better than, the structures submitted to the Protein Data Bank. Additionally, the work describes the data-guided Multi- Map methodology for ensemble refinement of molecular movies. The method shows that cryo-electron microscopy maps can be used to bias simulations along a specially constructed reaction coordinate and capture conformational transitions between known intermediates. The simulated pathways appear reversible with minimal hysteresis and require only low-resolution density information to guide the transition. Finally, the study analyzes the SARS-CoV-2 spike protein and the conformational heterogeneity of its receptor binding domain. The simulation was guided along an experimentally determined free energy landscape. The resulting motions from following a pathway of low-energy states show a degree of openness not observed in the static models. This sheds light on the mechanism by which the spike protein is utilized for host infection and provides a rational explanation for the effectiveness of certain therapeutics. This work contributes to the understanding of biomolecular modeling and the development of new strategies to provide valuable insights into the workings of cellular machinery.