DASH properties: Estimating atomic and molecular properties from a dynamic attention-based substructure hierarchy.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Accurate prediction of atomic properties is crucial for molecular modeling.
• Existing methods often require computationally expensive quantum mechanical (QM) calculations.
• Graph neural networks offer a data-driven approach to chemical property prediction.

Purpose of the Study:

• To demonstrate the utility of the dynamic attention-based substructure hierarchy (DASH) tree for predicting diverse atomic properties.
• To show that a single DASH tree can capture features for multiple properties, avoiding property-specific model retraining.
• To provide an efficient, explainable alternative to QM calculations for generating molecular features.

Main Methods:

• Utilized a previously developed DASH tree construction method.
• Leveraged graph neural network attention values to define "rules" for property assignment.
• Applied the DASH tree to predict various atomic properties without QM calculations.
• Made the DASH tree, atomic properties dataset, and wave functions publicly available.

Main Results:

• The DASH tree effectively captures local atomic environment features.
• Accurate prediction of multiple atomic properties was achieved using the same DASH tree.
• The method offers high efficiency and QM-like accuracy.
• The approach provides explainable predictions through visualization of substructures.

Conclusions:

• The DASH tree is a versatile and efficient tool for predicting atomic properties.
• This method significantly reduces the computational cost associated with generating molecular features for machine learning.
• The explainable nature of the DASH tree enhances trust and understanding in the predictions.