Development of the Lennard-Jones Parameters for the Polarizable Classical Drude Oscillator Force Field
Abstract
The quality of Force Fields (FF) determines accuracy of observations made through molecular simulations. Accuracy of such simulations may be achieved by explicit inclusion of electronic polarization, such as via the implementation of the Drude harmonic oscillator, as in Drude Polarizable FF. Although the Drude Polarizable FF spans a large range of biomolecules including proteins, nucleic acids, lipids and carbohydrates, an expansion of its small molecule FF is essential, given the vastness of chemical space. Such an expansion must be accompanied by the optimization of van der Waals (vdW) interactions, in the context of the Lennard-Jones (LJ) formalism. Optimization of the LJ parameters is a multivariate and multi-objective problem and is one of the most challenging aspects of FF optimization. Through this thesis, we have developed a method that utilizes the sampling power of Latin Hypercube Design (LHD) and learning abilities of Deep Neural Network (DNN) to overcome some of these challenges. The model is trained on empirical pure solvent/crystal properties of a selected set of “training set” compounds, where the final selection is based on the quality of both gas phase and condensed phase properties. The optimized LJ parameters are validated for transferability on “validation set” compounds, while their ability to reproduce other experimental thermodynamic properties such as hydration free energy and dielectric constants, are also verified. Chapter 1 of this thesis presents an introduction to underlying concepts of FFs, with a major focus on polarizable FFs. Chapter 2 details development of the method, using four different chemical classes, i.e., alkenes, 3 and 4 membered rings and nitriles. Chapter 3 updates the method developed in chapter 2, addressing the challenge of parameter correlation. Chapter 4 applies the updated method to another chemical class (alkynes), while Chapter 5 concludes the thesis and is a discussion of the challenges associated with empirical FF development with a focus on LJ parameters. Overall, the method developed through this thesis addresses the most challenging aspect of FF development, i.e., LJ parameter development, implemented in a manner that could be utilized in context of development of both additive and polarizable FF.Description
University of Maryland, Baltimore. Pharmaceutical Sciences. Ph.D. 2022Keyword
polarizable force fieldslennard-jones parameters
latin hypercube design
drude
condensed phase data
Deep Learning