<![CDATA[Graph convolutional neural networks (GCNs) have shown promising performance in modeling graph data, particularly for small-scale molecules. Message-passing neural networks (MPNNs) are an important form of GCN variant. They excel at gathering and integrating particular information about molecules via several repetitions of message transmission. This capability has resulted in major advances in molecular modeling and property prediction. By combining the self-attention mechanism with MPNNs, there is potential to improve molecular representation while using Transformers' proven efficacy in other artificial intelligence disciplines. This research introduces a transformer-based message-passing neural network (T-MPNN) that is intended to improve the process of embedding molecular representations for property prediction. Our technique incorporates attention processes into MPNNs' message-passing and readout phases, resulting in molecular representations that are seamlessly integrated. The experimental results from three datasets show that T-MPNN outperforms or equals cutting-edge baseline models in tasks involving quantitative structure-property connections. By studying case studies of SARS-COV growth inhibitors, we demonstrate our model's ability to graphically depict attention at the atomic level. This enables us to pinpoint individual chemical atoms or functional groups linked with desirable biological properties. The model we propose improves the interpretability of classic MPNNs and is a useful tool for investigating the impact of self-attention on chemical substructures and functional groups in molecular representation learning. This leads to a better understanding of medication modes of action.]]>
Copyrights © 2024
