Attribute-guided prototype network for few-shot molecular property prediction.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Bioinformatics and computational biology

Background:

• Molecular property prediction (MPP) is vital for drug discovery, but deep learning methods require large labeled datasets.
• Few-shot MPP, predicting properties with limited data, presents a significant challenge in computational chemistry.
• Existing deep learning models struggle with data scarcity for accurate molecular property prediction.

Purpose of the Study:

• To develop a novel deep learning framework for few-shot molecular property prediction.
• To enhance the generalization capabilities of models in molecular property prediction tasks.
• To leverage molecular attributes for improved performance in data-scarce scenarios.

Main Methods:

• Proposed an attribute-guided prototype network (APN) incorporating a molecular attribute extractor.
• Extracted various fingerprint attributes (single, dual, triplet) and deep attributes via self-supervised learning.
• Designed an Attribute-Guided Dual-channel Attention module to refine molecular representations by integrating graph and attribute information.

Main Results:

• APN achieved state-of-the-art performance on benchmark datasets for few-shot MPP.
• Demonstrated the effectiveness of molecular attributes in improving few-shot MPP accuracy.
• Verified the strong generalization ability of APN across different data domains.

Conclusions:

• The proposed APN effectively addresses the challenges of few-shot molecular property prediction.
• Leveraging explicit molecular attributes enhances model generalization and predictive power.
• APN offers a promising approach for accelerating drug discovery through efficient molecule evaluation.