Kernel density estimation (KDE) estimates a continuous probability density by summing smooth, symmetric, non-negative kernel functions centered on observed data points. Unlike a histogram, it avoids arbitrary bins and provides a continuous view of how crowded regions are. In drug discovery, KDE can model distributions of distances between 3D pharmacophore features, helping characterize fragment ensembles and score new molecules against learned structural patterns. section 7.3.3
The estimator averages kernel contributions from all observations; bandwidth controls how broadly each observation influences nearby locations. section 7.3.3
A practical pharmacophore-density workflow is:
- Featurize fragment molecules with RDKit 3D pharmacophore features such as donors, acceptors, and aromatics. section 7.3.3
- Calculate pairwise distances between selected pharmacophore types across distinct fragments. section 7.3.3
- Fit Gaussian KDE models with scikit-learn, selecting bandwidth through cross-validated grid search. section 7.3.3
- Compute corresponding feature-pair distances for a new molecule and use the fitted distributions for scoring. section 7.3.3
Fragment ensemble screening
A fragment ensemble’s donor–acceptor distance distribution can be represented by a fitted KDE rather than a jagged histogram. A candidate molecule is evaluated by calculating its matching pharmacophore distances and consulting the learned pair-distribution models. section 7.3.3


