Descriptor generation from Morgan fingerprint using persistent homology.

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Molecular fingerprints (FPs) are vital in cheminformatics for tasks like regression and classification.
• Existing predictive models often do not fully leverage Morgan FPs for regression tasks.
• There is a need for improved methods to enhance the predictive accuracy of molecular descriptors.

Purpose of the Study:

• To introduce novel descriptors derived from Morgan FPs using persistent homology.
• To improve the predictive accuracy of machine learning models in cheminformatics.
• To investigate the impact of persistent homology order and parameter tuning on descriptor performance.

Main Methods:

• Generated descriptors from reshaped Morgan FPs using persistent homology.
• Applied persistent homology to zeroth-order (PD0) and first-order (PD1) persistence diagrams.
• Utilized Gaussian process regression for predictive modeling on FreeSolv and ESOL datasets.
• Investigated the effect of grid size and principal component analysis (PCA) for descriptor optimization.

Main Results:

• Persistent homology significantly enhanced predictive accuracy compared to using Morgan FPs alone.
• PD1-generated descriptors yielded greater improvements than PD0-generated descriptors.
• Combining 4096 bits Morgan FPs with PD1 descriptors increased R-squared values for FreeSolv and ESOL datasets.
• Optimizing grid size and applying PCA mitigated overfitting and improved accuracy, even with shuffled FPs.

Conclusions:

• Persistent homology offers a valuable approach to enhance molecular descriptors derived from Morgan FPs.
• The choice of persistence diagram order (PD1 vs. PD0) and parameter tuning (grid size, PCA) is critical for optimal performance.
• This method holds promise for improving predictive modeling in cheminformatics, particularly for physical-chemical properties.