AABBA Graph Kernel: Atom-Atom, Bond-Bond, and Bond-Atom Autocorrelations for Machine Learning.

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Molecular graphs are powerful representations for chemical structures.
• Graph kernels transform molecular graphs into vectors for machine learning.
• Existing graph kernels primarily focus on atomic nodes.

Purpose of the Study:

• Develop a novel graph kernel incorporating atom-atom, bond-bond, and bond-atom (AABBA) autocorrelations.
• Evaluate the AABBA kernel's performance on regression tasks involving transition metal complexes.
• Improve molecular representations by considering both atomic and bond properties.

Main Methods:

• Developed the AABBA graph kernel.
• Applied the kernel to generate vector representations of molecular graphs.
• Tested the representations on regression machine learning tasks, including predicting energy barriers and bond distances for Vaska's complex.
• Utilized various machine learning models such as neural networks, gradient boosting machines, and Gaussian processes.

Main Results:

• The AABBA graph kernel demonstrated superior performance compared to baseline methods (atom-atom autocorrelations only).
• Dimensionality reduction revealed that bond-bond and bond-atom autocorrelations contribute significantly to feature relevance.
• The AABBA kernel effectively predicted energy barriers and bond distances for transition metal complexes.

Conclusions:

• The AABBA graph kernel offers a more comprehensive molecular representation by integrating atomic and bond properties.
• This novel approach can accelerate the exploration of large chemical spaces.
• The AABBA kernel provides a foundation for developing new molecular representations that leverage both atom and bond information.