BatmanNet: bi-branch masked graph transformer autoencoder for molecular representation.

Area of Science:

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

Background:

• Effective molecular representation learning is crucial for AI-driven drug discovery but remains challenging, especially with limited labeled data.
• Existing self-supervised learning methods for graph neural networks (GNNs) require complex tasks and large datasets, posing computational and time constraints.

Purpose of the Study:

• To develop a simple yet effective self-supervised strategy for learning molecular representations.
• To introduce a novel bi-branch masked graph transformer autoencoder, named BatmanNet, for enhanced molecular information capture.

Main Methods:

• Designed a self-supervised strategy to learn both local and global molecular information simultaneously.
• Proposed BatmanNet, a bi-branch masked graph transformer autoencoder with asymmetric branches for node and edge reconstruction.
• Utilized masked molecular graphs as input for the autoencoder architecture.

Main Results:

• BatmanNet effectively captures underlying molecular structure and semantic information.
• Achieved state-of-the-art results across multiple drug discovery tasks, including molecular property prediction, drug-drug interaction, and drug-target interaction.
• Demonstrated superior performance on 13 benchmark datasets, highlighting its potential for real-world applications.

Conclusions:

• BatmanNet offers a computationally efficient and effective approach to molecular representation learning.
• The proposed method significantly advances AI-driven drug discovery by improving model performance with limited data.
• BatmanNet shows great potential for accelerating the identification of novel therapeutics.