Efficient Integration of Molecular Representation and Message-Passing Neural Networks for Predicting Small Molecule Drug-like Properties

The physicochemical properties of a drug molecule determine its metabolism properties. There have been hybrid quantum mechanics approaches with computer-aided drug design and recent supervised machine-learning approaches to predict these properties of small-molecule drugs. However, these methods are low in accuracy and computationally expensive. To get around this problem and improve the performance of a model that predicts the properties of drug molecules, we came up with a novel architecture that uses a "bond order matrix" and structural information to improve molecular graph representations and information in the molecule. Message-passing neural networks (MPNNs) are a framework used to learn local and global features from irregularly formed data invariant to permutations. We take advantage of MPNN architecture and introduce a “semi-master node,” a unique way of representing the functional groups in a small molecule and aggregating features obtained from the functional groups, in anticipation of reverse engineering small molecules given the desired physicochemical properties. This novel architecture and molecule representation were evaluated on the QM9 dataset, which has 133,000 stable small organic molecules with nine heavy atoms (CONF) out of the GDB-17 chemical universe. The metric for evaluating the model's performance is DFT error, an estimated average error of the properties of each molecule. Our models have shown a performance gain of ~10%.