Type
Text
Type
Dissertation
Advisor
Simmerling, Carlos L | Dill, Ken | Raineri, Fernando. | Green, David F
Date
2015-05-01
Keywords
AMBER, ff14SB, force field, molecular mechanics, optimization, parameter | Biophysics
Department
Department of Biochemistry and Structural Biology.
Language
en_US
Source
This work is sponsored by the Stony Brook University Graduate School in compliance with the requirements for completion of degree.
Identifier
http://hdl.handle.net/11401/76949
Publisher
The Graduate School, Stony Brook University: Stony Brook, NY.
Format
application/pdf
Abstract
Proteins carry out many diverse but important biological tasks, the understanding of which can be greatly augmented by theoretical methods that can generate microscopic insights. A popular method for simulating proteins is called molecular mechanics. Molecular mechanics drives the dynamics of molecules according to their potential energy surface as defined by a force field. Because force fields are simple, molecular mechanics can be fast; but force fields must simultaneously be accurate enough for the conformational ensembles they generate to be useful. One force field that has been widely adopted for its utility is AMBER force field 99 Stony Brook (ff99SB). The ff99SB protein backbone parameters were fit to quantum mechanics energies of glycine and alanine tetrapeptides, including a set of minimum energy conformations in the gas-phase. Although ff99SB rigorously reproduces many thermodynamic properties, it has shortcomings. Issues with backbone parameters may result from training against only energetic minima or from the energy calculations being in the gas phase. Problems with side chain parameters can stem from the protocol of ff99, where the amino acid side chain parameters were trained against energies of small molecules, while transferability from small molecules to amino acids may be problematic. Small updates to the backbone potential were applied by several groups, as well as the Simmerling group as part of ff14SB. Whereas ff99SB and ff14SB are fixed-charge, additive molecular mechanical models, there are also molecular mechanical models that include non-additive effects like charge polarization. Polarizable force fields, with their many additional degrees of freedom, promise enhanced accuracy relative to fixed charge force fields. But with so many degrees of freedom and thus parameters, polarizable force fields can be more difficult to train. Although this complexity may be overcome, it is unclear whether the utility of fixed-charge, additive force fields has been exhausted, warranting the great endeavors of developing a polarizable model. This dissertation seeks to answer how much more fixed charge force fields can be improved. Specifically, this work addresses two questions. Firstly, can the side chain parameters of ff99SB be improved by fitting to quantum mechanics energies? We investigated different options in the calculation of energies for parameter training, finding that how the structures were minimized can significantly affect transferability of parameters trained against them. Specifically, we found that loosely restraining the side chains, which were being refined, and tightly restraining the backbone, which was not, made the errors most similar between α and β backbone contexts. This transferability can be measured by improved agreement with the quantum mechanics training set as well as experimental scalar couplings. Secondly, can the backbone parameters of ff99SB be made more accurate, alternatively to empirical tweaks, by another, improved fitting to quantum mechanics energies? We found that better reproduction of NMR solution scalar couplings was possible, if energy calculations included solvation effects, full grids of structures were included, and, perhaps surprisingly, if parameters were extrapolated to those appropriate for a zero-length peptide. These results show that quantum mechanics can be effectively used to improve the accuracy of molecular mechanics force fields. These improvements have implications for protein structure prediction, aiding the successful folding of 16 of 17 proteins in GB-Neck2 implicit solvent. Beyond, the insights from the QM-based backbone training could be extended to develop residue-specific parameters that bolster the sequence-dependent structural preferences of proteins in simulation models. | Proteins carry out many diverse but important biological tasks, the understanding of which can be greatly augmented by theoretical methods that can generate microscopic insights. A popular method for simulating proteins is called molecular mechanics. Molecular mechanics drives the dynamics of molecules according to their potential energy surface as defined by a force field. Because force fields are simple, molecular mechanics can be fast; but force fields must simultaneously be accurate enough for the conformational ensembles they generate to be useful. One force field that has been widely adopted for its utility is AMBER force field 99 Stony Brook (ff99SB). The ff99SB protein backbone parameters were fit to quantum mechanics energies of glycine and alanine tetrapeptides, including a set of minimum energy conformations in the gas-phase. Although ff99SB rigorously reproduces many thermodynamic properties, it has shortcomings. Issues with backbone parameters may result from training against only energetic minima or from the energy calculations being in the gas phase. Problems with side chain parameters can stem from the protocol of ff99, where the amino acid side chain parameters were trained against energies of small molecules, while transferability from small molecules to amino acids may be problematic. Small updates to the backbone potential were applied by several groups, as well as the Simmerling group as part of ff14SB. Whereas ff99SB and ff14SB are fixed-charge, additive molecular mechanical models, there are also molecular mechanical models that include non-additive effects like charge polarization. Polarizable force fields, with their many additional degrees of freedom, promise enhanced accuracy relative to fixed charge force fields. But with so many degrees of freedom and thus parameters, polarizable force fields can be more difficult to train. Although this complexity may be overcome, it is unclear whether the utility of fixed-charge, additive force fields has been exhausted, warranting the great endeavors of developing a polarizable model. This dissertation seeks to answer how much more fixed charge force fields can be improved. Specifically, this work addresses two questions. Firstly, can the side chain parameters of ff99SB be improved by fitting to quantum mechanics energies? We investigated different options in the calculation of energies for parameter training, finding that how the structures were minimized can significantly affect transferability of parameters trained against them. Specifically, we found that loosely restraining the side chains, which were being refined, and tightly restraining the backbone, which was not, made the errors most similar between α and β backbone contexts. This transferability can be measured by improved agreement with the quantum mechanics training set as well as experimental scalar couplings. Secondly, can the backbone parameters of ff99SB be made more accurate, alternatively to empirical tweaks, by another, improved fitting to quantum mechanics energies? We found that better reproduction of NMR solution scalar couplings was possible, if energy calculations included solvation effects, full grids of structures were included, and, perhaps surprisingly, if parameters were extrapolated to those appropriate for a zero-length peptide. These results show that quantum mechanics can be effectively used to improve the accuracy of molecular mechanics force fields. These improvements have implications for protein structure prediction, aiding the successful folding of 16 of 17 proteins in GB-Neck2 implicit solvent. Beyond, the insights from the QM-based backbone training could be extended to develop residue-specific parameters that bolster the sequence-dependent structural preferences of proteins in simulation models. | 241 pages
Recommended Citation
Maier, James, "On the derivation of accurate force field parameters for molecular mechanics simulations" (2015). Stony Brook Theses and Dissertations Collection, 2006-2020 (closed to submissions). 2817.
https://commons.library.stonybrook.edu/stony-brook-theses-and-dissertations-collection/2817