We aim at computing the different parts of a dataset of molecules, considering both their chemical structures and their activity towards a biological receptor – in this case Bcr-Abl. To this aim, we start by representing a molecule with three molecular fingerprinting either based on state of the art ECFP4 descriptors, on the FCFP4s or a new description from our laboratories that uses frequent pharmacophores associated to a dataset. Then we perform two learned feature transformations by passing the data through two feed forward neural network (FFNN), a supervised FFNN and an unsupervised one. Those two feature transformations allow to weight the data. Finally, we relied on clustering techniques to identify different families of ligands in the studied dataset. The partitions obtained by using the former computation initiate a comparative study between the different fingerprinting methods: the pharmacophores and ECFP4 and FCFP4 representations. The results have demonstrated that our pharmacophores obtain the best results and allow to distinguish between three principal families of ligands in addition to efficient results on pharmacophoric decoys.