Artificial neural networks, the support vector machine (SVM), and other machine learning methods for the classification of molecules are often considered as a "black box", since the molecular features that are most relevant for a given classifier are usually not presented in a human-interpretable form. We report on an SVM-based algorithm for the selection of relevant molecular features from a trained classifier that might be important for an understanding of ligand-receptor interactions. The original SVM approach was extended to allow for feature selection. The method was applied to characterize focused libraries of enzyme inhibitors. A comparison with classical Kolmogorov-Smirnov (KS)-based feature selection was performed. In most of the applications the SVM method showed sustained classification accuracy, thereby relying on a smaller number of molecular features than KS-based classifiers. In one case both methods produced comparable results. Limiting the calculation of descriptors to only the most relevant ones for a certain biological activity can also be used to speed up high-throughput virtual screening.