Deep Learning-Based Predictive Model of Molecular Properties

With the rapid development of computer hardware and machine learning algorithms in recent decades, artificial intelligence has created a boom in various industries. Computer technology has been widely used in all stages of drug discovery, from machine learning to deep learning, and scientists have conducted numerous studies on the physicochemical properties of drugs. With the advancement of computer technology and deep learning algorithms, it has become possible to obtain large-scale molecular characterization data, extract different types of molecular representations, and construct more subtle models. Among them, molecular graph structure data is increasingly used in deep learning to predict molecular properties. It directly extracts the two-dimensional structural information of molecules, avoiding the defects of indirect structural information extraction in traditional methods. However, the performance of different algorithms based on molecular fingerprints, descriptors, molecular pictures and molecular graph structures is currently controversial. Some studies have found that descriptor-based models outperform graph-based models, while others do the opposite. This shows the difference in the impact of different tasks and datasets on the models. In this experiment, molecular graphs and molecular fingerprints are combined, and molecular graphs are fed into graph neural networks to extract features after the attention mechanism, molecular fingerprints are fed into a fully connected layer to extract features, and the features extracted from the two molecular representations are fused together and a specific activation function is adopted to see if the accuracy and stability of the model can be effectively improved.