Improving Predictions of Molecular Properties with Graph Featurization and Heterogeneous Ensemble Models

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Accurate molecular property prediction is crucial for drug discovery and materials science.
• Existing methods often rely on either general descriptors or learned features, with limitations in scope.
• Integrating diverse feature types and models can potentially improve predictive performance.

Purpose of the Study:

• To develop and evaluate a "best-of-both" approach for molecular property modeling.
• To introduce a MetaModel framework for aggregating predictions from multiple machine learning models.
• To investigate the synergy between graph neural network (GNN)-derived features and conventional molecular descriptors.

Main Methods:

• A featurization scheme combining GNN-learned descriptors with general-purpose molecular descriptors.
• Implementation of a MetaModel framework to aggregate predictions from an ensemble of machine learning models.
• Benchmarking against the state-of-the-art ChemProp model on various regression and classification tasks.

Main Results:

• The proposed framework demonstrated superior performance compared to ChemProp across all tested regression datasets.
• The model achieved higher accuracy on 6 out of 9 classification datasets.
• Incorporating GNN features notably improved ensemble model performance on datasets where it initially underperformed.

Conclusions:

• Combining general-purpose descriptors with task-specific learned features is vital for optimal molecular property prediction.
• Utilizing a diverse ensemble of machine learning models enhances predictive robustness and accuracy.
• The "best-of-both" approach offers a powerful strategy for advancing computational chemistry and cheminformatics.