Dissertation
Accelerating Molecular Dynamics Simulations and Predicting Reaction Pathways with Free Energy Surface Gradients Derived From Machine Learning Models
Doctor of Philosophy (PhD), Washington State University
05/2024
Abstract
Molecular Dynamics (MD) simulations are a computational tool to investigate atomic interactions in proteins and molecular systems. However, the slow rate of simulation renders many biological processes inaccessible to study. To combat this, machine learning techniques are under development to identify reaction coordinates as a function of molecular descriptors (a.k.a. molecular coordinates). Once relevant reaction coordinates are known, the energy of the molecular system can be biased to reduce energy barriers that inhibit conformational sampling.
In this work, a machine learning approach is developed to discover reaction pathways by training a machine learning model on trajectories of MD simulations and extract gradients of the free energy surface (FES). With these gradients, trends or patterns in the reaction pathways can be identified – sometimes even before a reaction has been fully sampled.
The developed algorithm log-probability estimation via invertible neural networks for enhanced sampling (LINES) is validated on several small systems that showcase the approach’s ability to learn the FES and produce reaction coordinates that significantly accelerate sampling during MD simulations. The method is later applied to discover protein-peptide binding sites, explore novel protein conformations, and evaluate the strength of inter-residue interactions that stabilize protein complexes.
Metrics
3 File views/ downloads
17 Record Views
Details
- Title
- Accelerating Molecular Dynamics Simulations and Predicting Reaction Pathways with Free Energy Surface Gradients Derived From Machine Learning Models
- Creators
- Ryan E. Odstrcil
- Contributors
- Jin Liu (Co-Chair)Prashanta Dutta (Co-Chair)Soumik Banerjee (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Mechanical and Materials Engineering, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Number of pages
- 150
- Identifiers
- 99901121439901842
- Language
- English
- Resource Type
- Dissertation