Is theoretical chemistry entering a new era of big data learning and reasoning: deep electron cloud-activity and field-activity relationships

Abstract

Chemists have been pursuing general mathematical laws to explain the properties of chemical molecules for a long time. However, the traditional quantitative structure-activity relationship (QSAR) models only provide small pieces of knowledge due to their poor generalization performance and limited application domain. This paper attempts to find a path to realize aunified QSAR that can theoretically predict ANY properties of ANY molecules. A framework combining deep learning, accurate quantum chemistry data and a large number of molecularentities is proposed, and its feasibility is studied. The proposed methods are called deep electron cloud-activity relationships (DECAR) and deep field-activity relationships (DFAR), which consist of three essentials: (1) thousands of molecule entities and activities (even reaching the limit of the current computing resources) as input objects and activity responses; (2) three-dimensional electron cloud density or related field data obtained by accurate density functional theory (DFT) methods as input features; and (3) a deep learning model with three-dimensional convolution layers that can learn these large data. The DECAR and DFAR methods are used to distinguish 977 sweet and 1965 nonsweet molecules, and the classification performance of deep learning models is demonstrated to be significantly better than that oftraditional least squares support vector machine (LS-SVM) models. This work will enable the establishment of an interactive international platform for collecting and sharing the accurate electronic cloud and field data of millions of molecules with annotated activities. We envision the development of hundreds of networks trained for various molecular activities using a large number of molecules, which will become open and shared learning and inference tools for chemists all over the world. These networks will make the best use of our existing knowledge of molecular activity to infer the properties of an increasing number ofunknown molecules and will promote our prediction and understanding of molecular properties with unprecedented strength.