Accelerating Computational Chemistry Algorithms: towards Accurate Binding Free Energies

Abstract

Advances in high performance computing (HPC) are pivotal for accelerating drug discovery by offering the capacity for large scale high accuracy virtual screening. There has been considerable interest in the large scale application of the Quantum Mechanical / Poisson-Boltzmann Surface Area (QM/PBSA) end point method towards predicting binding affinities of protein-ligand systems. However, its application is severely limited by the high computational costs of traditional QM methods as well as the slow performance of existing PBE solvers for PBSA calculations.

This thesis presents methods and algorithms to overcome such bottlenecks. The first half of this thesis is dedicated to accelerating QM calculations. An automated accurate molecular fragmentation scheme is presented that divides large systems into smaller, computationally feasible partitions, whilst enhancing algorithmic parallelisability. Its applicability on protein and lipoglycans/glycolipids is demonstrated. On the other hand, improved initial guess methods for self-consistent field (SCF) calculations are presented (basis set projection and fragmentation) and their performance is systematically analysed against the traditional superposition of atomic density (SAD) scheme. Results consistently indicate the improved performance of SCF calculations with non-SAD schemes.

To address the lack of fast PBE solvers to model solvation, a high-performance GPU-accelerated solver is presented. The algorithm exploits the sparsity pattern exposed in its application on molecular systems to accelerate matrix-vector contractions prevalent in conjugate gradient solvers, outperforming existing multi-core CPU and GPU-based PBE solvers.

These tools are integrated into a QM/PBSA workflow and applied to large biologically relevant protein-ligand complexes to predict binding affinities. The influence of various factors---fragmentation level, protonation states, and solvation methods---on the performance of the proposed QM/PBSA workflow is systematically analysed and compared to other computational approaches including alchemical free energy and scoring function methods.

This thesis seeks to improve upon the accuracy, computational efficiency and feasibility of large scale QM/PBSA workflows by leveraging chemical concepts and HPC optimisations. Beyond drug discovery, the methodologies presented have broad applicability to molecular molecular systems beyond proteins, supporting research in materials science, chemistry and energy applications. Such advancements become increasingly important as HPC systems continue to evolve, offering the potential to study molecular systems at even larger scales and and higher accuracy.