A Fragmentation-Based Graph Embedding Framework for QM/ML.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• Approximate quantum chemistry methods like DFT have limitations in accuracy.
• Machine learning (ML) offers potential for high-accuracy molecular property prediction.
• Integrating chemical principles into ML models is crucial for reliable predictions.

Purpose of the Study:

• To introduce FragGraph, a novel fragmentation-based molecular representation for QM/ML.
• To enhance the accuracy of approximate quantum mechanical methods using delta-machine learning (Δ-ML).
• To develop a graph network framework that incorporates fragmentation and error cancellation principles.

Main Methods:

• Developed a fragmentation-based attributed graph representation using fragment-wise molecular fingerprints.
• Implemented a graph network fingerprint encoder for generating global fingerprints via message passing.
• Combined fragmentation, error cancellation (generalized Connectivity-Based Hierarchy), and deep learning.

Main Results:

• Achieved a mean absolute error below 1 kJ/mol on the GDB9 dataset for energy prediction.
• Demonstrated that FragGraph rivals state-of-the-art deep learning methods in accuracy.
• Showcased reduced computational scaling compared to existing methods.

Conclusions:

• FragGraph provides an accurate and computationally efficient molecular representation for QM/ML.
• The framework effectively integrates molecular structure and fragment information for improved predictions.
• This approach holds promise for advancing high-accuracy molecular modeling in computational chemistry.