2022 Annual Meeting
Application of Geometric Algebra Neural Networks to Molecular Property Prediction
Accurate molecular property prediction has become more efficient and accessible due to advances in deep learning (DL) techniques. A fundamental consideration when designing geometric DL algorithms for chemistry is the choice of molecular input representation. Some common molecular representations include strings, graphs, and atomic coordinates. One challenge of using atomic coordinates as inputs is coordinate system dependence, which must be overcome in order to design a model whose output is equivariant to a molecule that is translated or rotated in 3D space. Recently, a DL approach called Equivariant Neural Networks (EQNNs) has emerged to address this challenge with high efficiency. One recent EQNN is based on geometric algebra, which uses bivectors and trivectors to represent 2- and 3-dimensional spatial relationships, respectively. In this framework, the geometric product is applied to pairs of atomic coordinates to obtain rotation-equivariant structural information with high efficiency. To demonstrate the application of the geometric algebra neural network and its ease of implementation, we apply it to Quantum Mechanics 9 (QM9)âa database consisting of 134,000 small organic molecules, along with their by-atom coordinates and element typesâto predict enthalpy of formation.