A probability density function describes how densely observations occur across the possible values of a continuous variable. Unlike a histogram, which counts observations in arbitrarily chosen intervals, a density estimate preserves a smoother view of the underlying distribution. section 7.3.3
Kernel density estimation (KDE) estimates this distribution by placing a smooth, symmetric, non-negative kernel on every observation. Nearby points contribute more strongly than distant points, and normalization makes the estimated density integrate to one. The bandwidth controls how broadly each observation influences the estimate. section 7.3.3
A practical fragment-screening workflow uses density estimates to model pharmacophore geometry:
- Extract donor, acceptor, and aromatic pharmacophore features and their three-dimensional coordinates from fragment molecules. section 7.3.3
- Calculate pairwise distances between pharmacophore types across distinct fragments, producing observations for each pharmacophore pair. section 7.3.3
- Fit Gaussian KDE models with scikit-learn, selecting bandwidth through cross-validated grid search. section 7.3.3
- Score a new molecule by comparing its pharmacophore-pair distances with the learned distributions. section 7.3.3
Pharmacophore-pair density
A bimodal sample can reveal two high-density regions that a jagged histogram may obscure. In fragment screening, KDE models capture the observed distance distributions between pharmacophore pairs and help evaluate whether a new molecule fits those patterns. section 7.3.3
Use the KDE equation to relate the estimated density to the kernel, bandwidth, observations, and normalization factor. section 7.3.3







