Property-Aware Relation Networks for Few-Shot Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Molecular property prediction is crucial for AI-aided drug discovery but faces challenges due to limited labeled data, characteristic of few-shot learning problems.
• Existing methods often struggle to capture the dynamic, property-specific relationships between molecules.

Purpose of the Study:

• To introduce Property-Aware Relation networks (PAR) designed to address the few-shot molecular property prediction challenge.
• To develop a transferable version (T-PAR) capable of handling distribution shifts common in drug discovery.

Main Methods:

• PAR utilizes a property-aware molecular encoder to generate property-specific molecular embeddings.
• A query-dependent relation graph learning module refines these embeddings by modeling inter-molecular relationships relevant to the target property.
• T-PAR employs joint sampling and relation graph learning schemes for simultaneous learning across source and target domains to mitigate distribution shifts.

Main Results:

• PAR significantly improves performance on few-shot molecular property prediction tasks compared to existing methods.
• T-PAR demonstrates superior results on transferable few-shot molecular property prediction, effectively handling domain shifts.
• Ablation and case studies confirm the effectiveness and design rationale of both PAR and T-PAR.

Conclusions:

• PAR and T-PAR offer effective solutions for few-shot and transferable few-shot molecular property prediction in drug discovery.
• The proposed methods leverage property-specific relationships and domain adaptation techniques to enhance molecular representation learning.
• These advancements hold promise for accelerating the identification of candidate molecules in AI-driven drug discovery pipelines.