Rapid Computation of Thermodynamic Properties over Multidimensional Nonbonded Parameter Spaces Using Adaptive Multistate Reweighting
详细信息 查看全文
- 作者：Levi N. Naden ; Michael R. Shirts
- 刊名：Journal of Chemical Theory and Computation
- 出版年：2016
- 出版时间：April 12, 2016
- 年：2016
- 卷：12
- 期：4
- 页码：1806-1823
- 全文大小：802K
- 年卷期：0
- ISSN：1549-9626
文摘
We show how thermodynamic properties of molecular models can be computed over a large, multidimensional parameter space by combining multistate reweighting analysis with a linear basis function approach. This approach reduces the computational cost to estimate thermodynamic properties from molecular simulations for over 130,000 tested parameter combinations from over 1000 CPU years to tens of CPU days. This speed increase is achieved primarily by computing the potential energy as a linear combination of basis functions, computed from either modified simulation code or as the difference of energy between two reference states, which can be done without any simulation code modification. The thermodynamic properties are then estimated with the Multistate Bennett Acceptance Ratio (MBAR) as a function of multiple model parameters without the need to define a priori how the states are connected by a pathway. Instead, we adaptively sample a set of points in parameter space to create mutual configuration space overlap. The existence of regions of poor configuration space overlap are detected by analyzing the eigenvalues of the sampled states’ overlap matrix. The configuration space overlap to sampled states is monitored alongside the mean and maximum uncertainty to determine convergence, as neither the uncertainty or the configuration space overlap alone is a sufficient metric of convergence. This adaptive sampling scheme is demonstrated by estimating with high precision the solvation free energies of charged particles of Lennard-Jones plus Coulomb functional form with charges between −2 and +2 and generally physical values of σij and ϵij in TIP3P water. We also compute entropy, enthalpy, and radial distribution functions of arbitrary unsampled parameter combinations using only the data from these sampled states and use the estimates of free energies over the entire space to examine the deviation of atomistic simulations from the Born approximation to the solvation free energy.