Augmenting MACCS Keys with Persistent Homology Fingerprints for Protein-Ligand Binding Classification

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Machine learning is vital in computational drug design for analyzing molecular data and predicting interactions.
• Current methods often rely on traditional fingerprints, which may not fully capture molecular complexity.

Purpose of the Study:

• To develop a novel molecular representation by integrating topological descriptors (persistence images) with MACCS Keys.
• To enhance the predictive performance of machine learning models in computational drug design.

Main Methods:

• Generated persistence images using topological data analysis to represent molecular shape.
• Concatenated persistence images with MACCS Keys to create an augmented molecular descriptor.
• Evaluated the augmented representation using a consistent artificial neural network on 19 ChEMBL protein-ligand bioactivity datasets.

Main Results:

• The augmented molecular representation consistently outperformed individual components (persistence images and MACCS Keys).
• A higher average validation Matthews correlation coefficient was achieved across most datasets.
• The integration of shape-based and traditional descriptors improved classification performance.

Conclusions:

• Integrating topological shape features with traditional cheminformatic descriptors offers a more robust molecular representation.
• This approach shows significant potential for enhancing predictive accuracy in computer-aided drug design.
• The findings suggest a promising direction for future drug discovery workflows.