Molecular graph augmentation with rings and functional groups.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Molecular graphs offer a concise representation but may limit generalization in graph-based machine learning.
• Essential chemical knowledge, such as functional groups and ring structures, is often not explicit in standard molecular graphs.

Purpose of the Study:

• To introduce a method for incorporating chemical background knowledge into molecular graph representations.
• To improve the predictive performance of graph-based quantitative structure-activity relationship (QSAR) models.

Main Methods:

• Augmenting molecular graphs by adding vertices and edges for identified functional groups and ring structures.
• Evaluating the augmented graphs using graph kernel-based QSAR models, including pairwise maximal common subgraphs kernel.

Main Results:

• The proposed augmentation method significantly improves predictive performance across various ligand-based tasks and datasets.
• State-of-the-art performance was achieved on the NCI-60 cancer dataset for 28 out of 60 cell lines.
• Near-optimal predictions were obtained on the Bursi mutagenicity dataset.

Conclusions:

• Incorporating chemical background knowledge into molecular graphs is an effective strategy for enhancing machine learning model generalization.
• This approach offers a promising avenue for improving QSAR modeling and drug discovery efforts.