Recently, the analysis of Structure-Activity Relationships has been confronted with the high dimensionality of chemical representations of molecular datasets, making analysis for drug discovery more complicated. To address this, different machine learning (ML) approaches have been exploited and proved their effectiveness by extracting relevant information. Generally, before applying ML methods, the raw data must be preprocessed to obtain better results. In our work, we start with a dataset described by pharmacophores obtained by the Norns tool 1 . Norns considers a dataset of molecules for which both structure and activity are given, and extracts a set of pharmacophores whose occurrences in the data set fulfill specified properties. For example, Norns allows the automatic extraction of 112047 pharmacophores from ligands tested on BCR-ABL, without any prior supervised selection. However, the size of this set is too large to perform efficient analysis and offer explicable results based on ML algorithms. This is why, in a first step, we select a subset of pharmacophores via grouping them into equivalence classes: pharmacophores occurring in the same set of molecules. This first unsupervised selection step allows us to retain 22127 pharmacophores. As the latter step removes redundant pharmacophores without losing statistical information, it allows to cope with the high dimensionality exhibited by chemogenomics datasets. We can then perform more sophisticated operations using the new representation: we do so by passing the data through two neural network-based transformations. The first neural network (NN) performs an unsupervised transformation 2 by reducing a loss function based on similarity computations. Its goal is to reduce the distance between similar data and increase it between dissimilar ones. The second NN exploits (a part of) the activity information with the aim of obtaining a more structured data space. While we use labels contained in the data in our work, the necessary information could also be the result of an expert interacting with the process, a first step towards interactive mining of pharmacophores. We obtain good clustering performances which lead to an easier and more efficient analysis. In addition, feature weights derived from the NN-based transformations could help to explain the results of the clustering step.