Functional Groups Are All You Need for Chemically Interpretable Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Deep learning (DL) models accelerate discovery but often lack chemical interpretability.
• This hinders the adoption of DL models by chemists in drug and materials science.

Purpose of the Study:

• To develop interpretable and high-performing DL models for molecular property prediction.
• Introduce a novel molecule representation framework based on chemical functional groups.

Main Methods:

• Developed the functional group representation (FGR) framework.
• Integrated curated functional groups (FG) and mined functional groups (MFG) using sequential pattern mining.
• Utilized pretraining on unlabeled molecules to encode representations in a latent space.

Main Results:

• Achieved state-of-the-art performance on 33 diverse benchmark datasets.
• Demonstrated chemical interpretability, linking predictions to specific functional groups.
• Showcased improved structure-property relationship insights.

Conclusions:

• Incorporating chemical knowledge into DL models enhances both performance and interpretability.
• The FGR framework provides chemically meaningful representations for property prediction.
• This approach facilitates the adoption of DL by chemists in discovery.