Multimodal Modules and Self-Attention for Graph Neural Network Molecular Properties Prediction Model

This paper proposes a multimodal Graph Neural Network (GNN) model for predicting molecular properties. Our model combines molecular graph topology information from a baseline GNN with an additional module, such as graph signal processing or text-based SMILES, to improve the accuracy of molecular property prediction. We utilized the CMPNN model as our baseline GNN and combined it with a Bidirectional LSTM module for text sequence in SMILES format or a spectral graph convolution module. Additionally, we also experimented with integrating self-attention into the CMPNN model through the use of the alpha coefficient method from GATConv. Our results demonstrate the effectiveness of our multimodal GNN model for molecular property prediction, outperforming the baseline on a range of datasets. This approach has potential applications in drug discovery and other areas of chemistry.