In recent studies, several machine learning (ML) approaches have been explored for efficient data-driven prediction of different properties associated with molecules and materials; these include forces, molecular energies (e.g. enthalpy, atomization, binding and free energy) and other electronic properties. Although these methods have the potential to offer significant computational speed-up from the traditional quantum mechanical simulations such as density functional theory (DFT), many of them struggle to learn complex molecular representation leading to inaccurate predictions for many properties. The supposedly “optimal” set of features may still not lead to the desired chemical accuracy for any practical use cases. This may be due to a wrong selection of the ML algorithm or the inability of the network architecture to model the quantum interactions. To resolve some of these issues, the more recent graph neural networks (GNN) have been introduced. These frameworks model the molecules as graphs, subsequently applying a message passing scheme to learn accurate graph representations from the node embeddings and have demonstrated remarkable improvements in the molecular property predictions.<br/><br/>This work extends a somewhat simpler yet efficient multi-level GNN framework to accurately predict extensive and intensive properties of organic molecules. To this end, we utilize only the basic atomic features and corresponding structures to extract high-level molecular graph representations. Additionally, the network learns bond importances by leveraging graph attention message passing. These importances are modeled as bond attributes and are further modulated with the informative bond properties such as bond type and atom pair distances. To evaluate the performance of the method, a large-scale benchmark QM9 database is used for predicting the different properties. To further strengthen the network, molecular descriptors are combined with the graph neural network framework in an end-to-end manner. Results show that the predictive performance of the proposed model is better than most of the previous approaches and lies well within the chemical accuracy determined by the research community.<br/>    <br/>The proposed model is a general framework that can also be extended to large chemical spaces for efficient and accurate prediction of electronic properties which aids in efficient drug-discovery and autonomous high-throughput screening of chemical candidates.

Towards efficient and accurate predictive modelling of configurational properties of organic molecules using graph neural networks

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.