MoleculeFormer is a GCN-transformer architecture for molecular property prediction

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular modeling

Background:

• Artificial intelligence (AI) is crucial for drug discovery, especially in predicting molecular properties.
• Graph Neural Networks (GNNs) model molecular structures but face challenges in feature optimization and integration.
• Existing methods struggle to effectively integrate multi-scale molecular features and 3D structural information.

Purpose of the Study:

• To introduce MoleculeFormer, a novel multi-scale feature integration model for molecular property prediction.
• To address limitations in feature optimization and model integration in AI-driven drug discovery.
• To develop a generalizable and interpretable AI solution for diverse drug discovery tasks.

Main Methods:

• Developed MoleculeFormer, a hybrid model combining Graph Convolutional Networks (GCNs) and Transformers.
• Employed independent GCN and Transformer modules for atom/bond graph feature extraction.
• Incorporated rotational equivariance constraints and prior molecular fingerprints, including 3D structural information.

Main Results:

• MoleculeFormer demonstrated robust performance across 28 diverse drug discovery datasets.
• Achieved high accuracy in predicting molecular efficacy, toxicity, and ADME properties.
• Showcased strong noise resistance and enhanced interpretability via attention mechanisms.

Conclusions:

• MoleculeFormer offers an effective and generalizable solution for molecular prediction in drug discovery.
• The model successfully integrates local and global molecular features with 3D structural data.
• Its robust performance and interpretability make it a valuable tool for advancing pharmaceutical research.