Development and Application of Numerical Methods in Biomolecular Solvation

Abstract

This work addresses the development of fast summation methods for long range particle interactions and their application to problems in biomolecular solvation, which describes the interaction of proteins or other biomolecules with their solvent environment. At the core of this work are treecodes, tree-based fast summation methods which, for N particles, reduce the cost of computing particle interactions from O(N^2) to O(N log N). Background on fast summation methods and treecodes in particular, as well as several treecode improvements developed in the early stages of this work, are presented.

Building on treecodes, dual tree traversal (DTT) methods are another class of tree-based fast summation methods which reduce the cost of computing particle interactions for N particles to O(N). The primary result of this work is the development of an O(N) dual tree traversal fast summation method based on barycentric Lagrange polynomial interpolation (BLDTT). This method is implemented to run across multiple GPU compute nodes in the software package BaryTree. Across different problem sizes, particle distributions, geometries, and interaction kernels, the BLDTT shows consistently better performance than the previously developed barycentric Lagrange treecode (BLTC).

The first major biomolecular solvation application of fast summation methods presented is to the Poisson–Boltzmann implicit solvent model, and in particular, the treecode-accelerated boundary integral Poisson–Boltzmann solver (TABI-PB). The work on TABI-PB consists of three primary projects and an application. The first project investigates the impact of various biomolecular surface meshing codes on TABI-PB, and integrated the NanoShaper software into the package, resulting in significantly better performance. Second, a node patch method for discretizing the system of integral equations is introduced to replace the previous centroid collocation scheme, resulting in faster convergence of solvation energies. Third, a new version of TABI-PB with GPU acceleration based on the BLDTT is developed, resulting in even more scalability. An application investigating the binding of biomolecular complexes is undertaken using the previous Taylor treecode-based version of TABI-PB. In addition to these projects, work performed over the course of this thesis integrated TABI-PB into the popular Adaptive Poisson–Boltzmann Solver (APBS) developed at Pacific Northwest National Laboratory.

The second major application of fast summation methods is to the 3D reference interaction site model (3D-RISM), a statistical-mechanics based continuum solvation model. This work applies cluster-particle Taylor expansion treecodes to treat long-range asymptotic Coulomb-like potentials in 3D-RISM, and results in significant speedups and improved scalability to the 3D-RISM package implemented in AmberTools. Additionally, preliminary work on specialized GPU-accelerated treecodes based on BaryTree for 3D-RISM long-range asymptotic functions is presented.