Three-Dimensional Convolutional Neural Networks Utilizing Molecular Topological Features for Accurate Atomization...

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Deep learning, particularly 3D-convolutional neural networks, shows promise for correlating molecular structure with properties.
• Traditional 3D data structures for molecules are sparse, leading to learning instabilities and underfitting.
• Predicting molecular properties accurately is crucial for drug discovery and materials design.

Purpose of the Study:

• To develop a more stable and accurate deep learning model for predicting molecular properties.
• To address the sparsity issue in 3D molecular data representation.
• To improve the prediction of atomization energies using novel molecular descriptors.

Main Methods:

• Utilized quantum-chemically derived molecular topological features, including localized orbital locator and electron localization function, as dense 3D input descriptors.
• Employed 3D-convolutional neural networks for property prediction.
• Integrated the delta-machine learning approach to enhance prediction accuracy.

Main Results:

• Achieved high accuracy in predicting atomization energies for the QM9-G4MP2 dataset (∼134k molecules).
• Demonstrated mean absolute errors of approximately 0.1 kcal mol⁻¹ (∼0.42 kJ mol⁻¹) compared to G4(MP2) theory.
• Exceeded benchmark accuracy levels, reaching beyond traditional limitations.

Conclusions:

• The proposed model effectively uses dense 3D topological features for accurate molecular property prediction.
• The integration of delta-machine learning significantly enhances predictive performance.
• The approach offers a pathway for systematic improvement in molecular property prediction with increased computational resources.