A Novel Descriptor and Molecular Graph-Based Bimodal Contrastive Learning Framework for Drug Molecular Property Prediction

In AI drug discovery, molecular property prediction is critical. Two main molecular representation methods in molecular property prediction models, descriptor-based and molecular graph-based, offer complementary information, but face challenges like representation conflicts and training imbalances when combined. To counter these issues, we propose a two-stage training process. The first stage employs a self-supervised contrastive learning scheme based on descriptors and graph representations, which pre-trains the encoders for the two modal representations, reducing bimodal feature conflicts and promoting representational consistency. In the second stage, supervised learning using target attribute labels is applied. Here, we design a multi-branch predictor architecture to address training imbalances and facilitate decision fusion. Our method, compatible with various graph neural network modules, has shown superior performance on most of the six tested datasets.