Quantum chemistry-augmented neural networks for reactivity prediction: Performance, generalizability, and...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Predictive Modeling

Background:

• A perceived gap exists between structure-based and descriptor-based molecular representations in predictive chemistry.
• Traditional methods often struggle with generalizability and explainability in complex chemical predictions.

Purpose of the Study:

• To evaluate the performance, generalizability, and explainability of a quantum mechanics-augmented graph neural network (ml-QM-GNN) architecture.
• To predict chemical regioselectivity (classification) and activation energies (regression).
• To bridge data-driven predictions with established theoretical frameworks for chemical reactivity.

Main Methods:

• Developed a hybrid QM-augmented model architecture combining structure-based representations with QM-derived reactivity descriptors.
• Utilized density functional theory (DFT) calculations to derive atom- and bond-level reactivity descriptors.
• Integrated predicted descriptors with original structure-based inputs for final reactivity predictions.

Main Results:

• The ml-QM-GNN architecture demonstrated significant improvements in accuracy and generalization to unseen compounds compared to structure-based GNNs.
• Outperformed state-of-the-art structure-based architectures and descriptor-based regressions, even with small training datasets.
• Established a link between data-driven predictions and qualitative reactivity insights.

Conclusions:

• The ml-QM-GNN offers a synergistic approach, combining computational chemistry theory with data science for robust chemical predictions.
• The model's grounding in QM descriptors enhances explainability and confirms qualitative chemical analyses.
• This hybrid approach facilitates a deeper understanding of chemical reactivity and machine learning model decision-making.