Towards a Universal Molecular Moiety Feature Set from the Topology of the Electron Density

点击翻译按钮获取中文摘要

The goal of chemistry is to develop new molecules and materials that improve our world. While we have seen significant improvements in our world, there have been some very troublesome issues that have resulted in a public that distrusts the chemical industry. Part of the reason for this is the limitations in our current models for developing new molecules. Many years ago, the approach was solely trial and error, which was very costly. With the advent of computers and computational chemistry, our models have improved and we continue to see dramatic improvements in our ability to develop materials that improve lives. Despite this, our models are still built on limited data sets. Machine learning techniques and algorithms show considerable promise for improving such models and allowing us to not only predict necessary properties important for a specific application, but also to account for secondary and tertiary effects that lead to problematic outcomes attributed to certain molecules. While this is an ongoing avenue of general research, one of the bottlenecks to its implementation is a valid, comprehensive, feature sets that describe molecules of interest. When developing new molecules, chemists generally think modularly, focusing on varying only specific fragments (functional groups, substrates, reaction centres). Over the years, descriptors for these fragments have been developed and employed to understand why such changes result in specific effects. Unfortunately, these descriptors do not truly describe the properties of the fragment, only their effect under specific conditions. This research aims to improve upon these models by first developing a set of fragment-based descriptors that are truly representative of its properties. Our focus is on developing such descriptors using the topological properties of the electron density as it has been shown that they are capable of demarcating any molecule into such chemically relevant fragments, and with the overall molecular property being a simple sum of the individual fragments. Three types of fragment properties are considered here: (1) those associated with a molecular graph, which help identify various bonding interactions; (2) those atomic properties obtained by integrating over the atomic basin as defined by the zero-flux surfaces of the molecular graph; and (3) the atomic critical points as obtained from a topological analysis of the laplacian. By combining these into an overall feature set for a large number of fragments, we will provide researchers with a data set for machine learning applications. Following the development of these feature sets, we will apply them in machine learning algorithms. The nature of our approach lends itself to using graph neural networks with vertices augmented by atomic properties and edges with molecular graph properties. This work will lead to an explanatory and predictive model for designing new molecules, hopefully resolving the secondary effect issue.