Meta-MolNet: A Cross-Domain Benchmark for Few Examples Drug Discovery.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Artificial intelligence for pharmaceuticals

Background:

• Predicting molecular properties is crucial for drug discovery but current machine learning models struggle with new molecular scaffolds.
• Existing methods lack standardized benchmarks, leading to unreliable evaluations and fragile predictions for novel chemical structures.

Purpose of the Study:

• Introduce Meta-MolNet, a novel benchmark platform for assessing model generalization and uncertainty quantification in molecular property prediction.
• Propose Meta-GAT, a cross-domain meta-learning graph attention network designed for robust predictions on new molecular scaffolds with limited data.

Main Methods:

• Developed Meta-MolNet, a comprehensive benchmark suite with diverse molecular datasets presenting significant domain shift challenges.
• Proposed Meta-GAT, a graph attention network employing bilevel optimization for meta-knowledge acquisition from source domains.
• Utilized meta-knowledge to enable few-shot learning for reliable predictions on unseen molecular scaffolds in target domains.

Main Results:

• Meta-MolNet effectively evaluates model performance in domain generalization and uncertainty quantification.
• Meta-GAT demonstrated state-of-the-art domain generalization capabilities.
• Meta-GAT robustly estimated prediction uncertainty even with minimal training examples for new scaffolds.

Conclusions:

• Meta-MolNet serves as a vital resource for the AI-assisted drug discovery community, standardizing model evaluation.
• Meta-GAT offers a powerful solution for reliable molecular property prediction in data-scarce scenarios.
• The study highlights the potential of meta-learning for advancing generalization in cheminformatics.