Molecular Property Prediction by Contrastive Learning from Joint-Aggregator Neural Network

Prediction of drug molecule properties is a crucial step in the evaluation of drug candidates, as it aids researchers in identifying compounds with favorable pharmacokinetic and pharmacodynamic characteristics. However, existing methods have shortcomings in utilizing extensive drug molecule property features and 3D structural information. Therefore, how to use the 3D features of a molecule to guide the network in learning 2D features, allowing the model to incorporate 3D information into the 2D representation has become a significant issue.In this paper, we proposes a molecular property prediction method based on Graph Neural Network, which employs contrastive learning extracting the 3D features from the network, helping guide joint-aggregator graph neural network learning the 2D features. The method transfer specific molecular information from large compound dataset, showing a more comprehensive map for the molecule’s overall properties. Additionally, the approach introduces joint aggregation factors to extract local molecule feature from different levels. The experimental results on multiple properties datasets demonstrate that the method achieves state-of-the-art or comparable performance in predicting molecular quantum properties and physicochemical properties compared with existing methods.