Accelerating the understanding of chemistry: with path-sampling and human understandable machine learning algorithms

Abstract

Understanding chemistry is essential for the optimization of reactions and the development of new reactions. Chemical reactions can be investigated by simulations without wasting any precious materials or be influenced by experimental artifacts and four of the six papers included in this thesis use simulations to investigate a wide range of chemistry. They investigate the effect of breaking assumptions of enzymatic assays for covalent inhibitors, the permeation of ions through a membrane, the effect of an oncogenic mutation on protein movements, and the deprotonation pathways for formic acid in atmospheric water droplets. One of the most efficient ways of simulating chemical reactions is replica exchange transition interface sampling (RETIS), where we focus the simulation on the reaction without wasting computational resources on simulating either the reactants or the products. In three of the six papers we further developed the RETIS algorithms and software, we interfaced with more molecular dynamics software, introduced more efficient Monte Carlo moves, and parallelized the RETIS algorithm. All of these increase the speed of RETIS simulations by orders of magnitude. Additionally, one of the papers specifically focuses on enhancing the analysis of RETIS simulations with machine learning (ML) algorithms. For this we introduce a new data representation that is translational, rotational and atom index invariant without any preselection of important variables or losing the ability to regenerate a 3D structure from it. This representation is then used with a human understandable ML algorithm, Decision Trees (DTs). The paper also introduces a way of investigating different initial splits of DTs with the help of random forests. This helps increasing the speed at which we can do analysis of RETIS simulations, while also reducing the risk of hypothesis bias.