Meta-MolNet

Meta-MolNet manages a wide range of molecular datasets and provides a benchmark platform and suite of algorithms for measuring model generalization and uncertainty quantification in drug discovery.

Key Features:

• Cross-domain benchmark datasets: Manages molecular datasets with high ratio of molecules/scaffolds, which leads to more difficult data shift and generalization problems.
• Meta-GAT algorithm: Implements a graph attention network based on cross-domain meta-learning that uses bilevel optimization to learn meta-knowledge from scaffold family molecular datasets in the source domain.
• Few-examples prediction: Meta-GAT reduces sample complexity requirements to enable reliable predictions of new scaffold molecules in the target domain through internal iteration of a few examples.
• Uncertainty quantification: The benchmark measures and the Meta-GAT model robustly estimates uncertainty under few examples constraints.
• Baseline evaluations: Evaluates existing methods as baselines for the community with AI-ready data, evaluation frameworks, and baseline results.

Scientific Applications:

• Pharmacological activity prediction: Predicts pharmacological activity of molecules across different scaffold families.
• Toxicity prediction: Predicts toxicity properties of molecules in data-scarce target domains.
• Pharmacokinetic property prediction: Predicts pharmacokinetic properties of molecules with limited examples.
• Domain generalization assessment: Measures model generalization capabilities across scaffold-based molecular datasets.

Methodology:

Meta-GAT uses bilevel optimization to learn meta-knowledge from scaffold family molecular datasets in the source domain, enabling predictions in the target domain through internal iteration of a few examples.