Molecular Connectivity Index-Based Data Augmentation for Molecular Property Prediction.

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Machine learning is increasingly used in cheminformatics, but insufficient training data is a major challenge.
• Existing data augmentation techniques often overlook the impact of construction rules and domain information on data quality.
• Molecular graph topology and topological indices like the molecular connectivity index are crucial for understanding physicochemical properties and biological activities.

Purpose of the Study:

• To develop a novel data augmentation technique for cheminformatics that addresses the limitations of current methods.
• To generate augmented molecular data that retains essential topology-based properties by preserving the molecular connectivity index.
• To improve the reliability and predictive accuracy of machine learning models in cheminformatics through enhanced data augmentation.

Main Methods:

• A new data augmentation technique was proposed that modifies the topology of molecular graphs.
• The method ensures that the augmented data maintains the same molecular connectivity index as the original data.
• The approach focuses on retaining crucial topology-based molecular properties during the augmentation process.

Main Results:

• The proposed data augmentation technique effectively generates reliable augmented data.
• Augmented data generated using molecular topology features led to significant improvements in prediction accuracy for molecular properties.
• Testing on five benchmark datasets confirmed the efficacy of the novel approach.

Conclusions:

• Modifying molecular graph topology while preserving the molecular connectivity index is a viable strategy for data augmentation in cheminformatics.
• This method enhances the retention of topology-based molecular properties, leading to more dependable augmented data.
• The findings offer a new perspective for data augmentation, improving machine learning model performance in cheminformatics studies.