GMPP-NN: a deep learning architecture for graph molecular property prediction

Abstract

The pharmacy industry is highly focused on drug discovery and development for the identification and optimization of potential drug candidates. One of the key aspects of this process is the prediction of various molecular properties that justify their potential effectiveness in treating specific diseases. Recently, graph neural networks have gained significant attention, primarily due to their strong suitability for predicting complex relationships that exist between atoms and other molecular structures. GNNs require significant depth to capture global features and to allow the network to iteratively aggregate and propagate information across the entire graph structure. In this research study, we present a deep learning architecture known as a graph molecular property prediction neural network. which combines MPNN feature extraction with a multilayer perceptron classifier. The deep learning architecture was evaluated on four benchmark datasets, and its performance was compared to the smiles transformer, fingerprint to vector, deeper graph convolutional networks, geometry-enhanced molecular, and atom-bond transformer-based message-passing neural network. The results showed that the architecture outperformed the other models using the receiver operating characteristic area under the curve metric. These findings offer an exciting opportunity to enhance and improve molecular property prediction in drug discovery and development.