Descriptor generation from Morgan fingerprint using persistent homology.

Abstract

In cheminformatics, molecular fingerprints (FPs) are used in various tasks such as regression and classification. However, predictive models often underutilize Morgan FP for regression and related tasks in machine learning. This study introduced descriptors derived from reshaped Morgan FPs using persistent homology for the predictive accuracy improvement. In the solvation free energy (FreeSolv) and water solubility (ESOL) datasets, persistent homology was found to enhance predictive accuracy compared to the use of only Morgan FPs. Notably, using the first-order persistence diagram (PD1) for descriptor generation resulted in more significant improvements than using the zeroth-order persistence diagram (PD0). Combining 4096 bits Morgan FPs with PD1-generated descriptors increased the average coefficient of determination in the Gaussian process regression from 0.597 to 0.667 for FreeSolv and from 0.629 to 0.654 for ESOL. Adjusting the grid size parameter during PD-based descriptor generation is crucial, as finer grids, especially with PD0, generate more descriptors but reduce predictive accuracy. Coarsening the grid or applying principal component analysis (PCA) mitigates overfitting and enhances accuracy. When descriptors were generated from Morgan FPs with randomly shuffled bit positions, coarsening the grid and/or applying PCA achieved similar accuracy improvements as when the persistent homology of the original Morgan FPs was used.