Computational Biology Department
School of Computer Science, Carnegie Mellon University
Inferring Viral Capsid Self-Assembly Pathway from
Self-assembly is a common phenomenon in the macro-molecular environment inside the cell and is critical for many cellular functions. Viral capsid assembly has been studied as a key model for self-assembly systems by researchers from different fields. There nonetheless remains a substaintial gap between experimental observations and current models, as the direct measurement of the assembly dynamics is currently intractible. Simulation-based methods can help bridge the gap, but the validity of such methods relies on the accuracy of a variety of physical parameters needed to instantiate the models, which also currently cannot be aquired by direct measurement.
The work of this thesis is focused on developing a parameter-learning framework that can infer kinetic parameters of viral assembly models by fitting the models to indirect bulk experimental measurements. The underlying rationale is based on the assumption that the set of parameters that minimizes the difference between simulated and experimental results would be the most plausible candidate. The framework extends existing stochastic self-assembly simulation methods, viral capsid models, and a prior heuristic optimization method to a flexible architecture that is adaptive to multiple data sources and alternative optimization methods.
The thesis specifically explores prospects for greater efficiency and accuracy through the use of more advanced algorithms or data sources for simulation-based model fitting. The framework has been tested on three in vitro viral assembly systems: human papillomavirus (HPV), heptatitis B virus (HBV) and cowpea chlorotic mottle virus (CCMV). The best fitting results from static light scattering (SLS) experiments suggest distinct in vitro assembly pathways for the three icosahedral viruses. Simulation experiments introducing synthetic non-covalent mass spectrometry(NCMS) data suggest that richer data sources can lead to substantial improvement in fitting accuracy. Complementary experiments on alternative optimization algorithms based on derivative free optimization (DFO) suggest that algorithmic advances can also substantially improve accuracy of model fits. Together, these results suggest that the methods can effectively reconstruct model parameters and assembly pathways given currently feasible algorithms and data sources, but that there is room for further advancement in improving both experimental and computational technologies underlying the approach.