Structure-Based Robust Fractal Graph Neural Network with Molecular Fingerprint BERT for Molecular Property Prediction

Accurate molecular property prediction is of great importance for AI-based drug design and bioinformatics. Despite recent promising progress, existing methods face challenges due to insufficient data labeling and imbalanced label distribution. Moreover, many current molecular property prediction models are based on 2D molecular graphs, neglecting the 3D characteristics of molecules. Therefore, in this paper, we propose the structure-based robust fractal graph neural network with Molecular Fingerprint Bidirectional Encoder Representations from Transformers (MFBERT) model for molecular property prediction. We construct the pre-trained MFBERT model to explore the use of unlabeled data and obtain representations from molecular extended connectivity fingerprints as the node features of molecular graphs. Then we design the fractal graph neural network that incorporates molecular 3D conformations while ensuring equivariance to molecular rotations and translations. A 3D feature extraction module and a coordinate update module are proposed to enhance the model's ability to capture molecular structural features. Additionally, weighted-gradient distributionally robust optimization is combined with the gradient descent algorithm to update model parameters, which can mitigate issues stemming from imbalanced label distribution and improve the robustness of our model. We conducted experiments on seven datasets from MoleculeNet, comprising three regression tasks and four classification tasks. Our method achieves the state-of-the-art performance compared to the existing models on all these datasets, which demonstrates its effectiveness for molecular property prediction. Additionally, we conducted ablation experiments on our designed modules, further validating the necessity of each component in our method.