PySIDT: Subgraph Isomorphic Decision Trees for Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Materials Science

Background:

• Accurate molecular property prediction is crucial in chemistry.
• Deep Neural Networks (DNNs) are popular but require large datasets, are difficult to interpret, and struggle with incorporating chemical knowledge.
• Existing methods often lack interpretability and struggle with smaller datasets.

Purpose of the Study:

• Introduce Subgraph Isomorphic Decision Trees (SIDTs) as a novel approach for molecular property prediction.
• Develop and present PySIDT software for training and inference using SIDTs.
• Demonstrate the advantages of SIDTs over DNNs and gradient boosted trees in terms of data efficiency, interpretability, and performance.

Main Methods:

• Developed Subgraph Isomorphic Decision Trees (SIDTs), a graph-based decision tree method using molecular substructures.
• Implemented PySIDT software for training and running inference on SIDTs.
• Applied SIDTs to diverse molecular prediction tasks including rate coefficients, thermochemistry, and stability.

Main Results:

• SIDTs demonstrated strong performance across various molecular property prediction tasks.
• PySIDT showed superior performance compared to popular DNN (Chemprop) and gradient boosted trees (XGBoost) methods, particularly with limited training data.
• In enthalpy of formation prediction, PySIDT outperformed Chemprop and XGBoost across all tested training/validation set sizes.

Conclusions:

• SIDTs offer a scalable, interpretable, and data-efficient alternative to DNNs for molecular property prediction.
• PySIDT facilitates the integration of chemical knowledge and uncertainty estimation.
• The SIDT approach shows significant promise for advancing computational chemistry and materials science applications.