Interpretable Molecular Property Predictions Using Marginalized Graph Kernels.

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Marginalized graph kernels offer strong performance in molecular machine learning.
• Current methods lack interpretability, hindering trust, bias detection, and molecular optimization.
• Interpretable AI is crucial for advancing molecular science.

Purpose of the Study:

• To introduce and implement two novel interpretability measures for Gaussian process regression using marginalized graph kernels (GPR-MGK).
• To quantify the influence of training data and graph nodes on predictions.
• To enhance the trustworthiness and applicability of GPR-MGK in molecular property prediction.

Main Methods:

• Development of two interpretability measures for GPR-MGK.
• Quantification of training data contribution to predictions.
• Quantification of node contribution to predictions.
• Comparison with graph neural networks on logic and toxicology datasets.
• Molecular attribution analysis on the FreeSolv dataset.

Main Results:

• GPR-MGK interpretability measures successfully applied to molecular property prediction.
• GPR-MGK demonstrated generally superior atomic attribution compared to graph neural networks across datasets.
• Detailed analysis revealed how training molecules impact predictions and explained limitations of Morgan fingerprints.

Conclusions:

• The developed interpretability measures significantly enhance the understanding of GPR-MGK models.
• GPR-MGK offers a more interpretable alternative to graph neural networks for molecular property prediction.
• This work represents a key advancement in interpretable marginalized graph kernel methods for molecular machine learning.