Inhibitors of Drug Efflux Pumps – A Pharmacoinformatic Approach

1) Artificial Neural Networks for Identification of new Lead Compounds (Dominik Kaiser)
In silico screening of large compound libraries is a versatile approach for identification of new lead compounds in the drug discovery and development process. Once a target has been identified and a set of active and inactive compounds are known, this information can be used to design filters for virtual screening of compound libraries.

This project focuses on the development of inhibitors of drug efflux pumps, such as P-glycoprotein (P-gp). P-gp is a membrane bound ATPase which transports a broad pannel of strucutrally and funtionally diverse cytotoxic drugs out of tumor cells. This leads to a decreased accumulation of the drugs in the cell and gives rise to multiple drug resistance. Analogous transport systems were identified in bacteria and fungi. Inhibition of these pumps thus leads to restoration of drug-sensitivity to multiple drug resistant tumours. Currently, 4 compounds are in clinical phase III studies for this indication.

Within our studies on propafenone-type inhibitors of P-gp, we used also self organising maps (SOMs) to identify new inhibitors of P-gp. A training set of 131 propafenones was used to train a SOM to distinguish between active and inactive P-gp inhibitors. Descriptors used were a set of autocorrelation vectors provided by the group of J. Gasteiger, Erlangen. In the next step, the map was enlarged, the propafenones were merged with the SPECS compound library (156.000 compounds) and the SOM was trained again. SPECS-compounds co-localizing with highly active propafenones were regarded as new lead compounds, ordered and pharmacologically tested in the group of P- Chiba, Vienna. 6 out of 7 compounds showed activity values in the low micromolar range. A further validation of this approach was achieved via identification of a set of inactive compounds. Within this set, only 1 out of 8 compounds showed moderate activity, the others were inactive. Next the approach will be used for virtual screening of even larger compound libraries, such as ChemDiv (450.000 compounds) or iResearch (12.000.000 compounds)


2) Similarity Based Structure Activity Relationships - SIBAR (Rita Schwaha)
The SIBAR-approach developed in our group is based on the concept, that similar compounds should show similar biological behavior. However, similarity between compounds is almost exclusively calculated on basis of chemical structures rather than pharmacophoric features. First, a reference set of compounds is defined on basis of chemical and/or biological diversity (i.e. active/inactive, drug like/non drug like). Subsequently, similarity values for compounds of the training set to the reference compounds are calculated (euclidian distance, tanimoto indices, ...). Descriptors used for calculation of similarity values are chosen according to the given problem (physicochemical parameters, autocorrelation vectors, fingerprints, molecular holograms,...). The similarity matrix obtained (SIBAR descriptors) is subject to PLS analysis or serves as input vector for artificial neural networks. Preliminary results show, that SIBAR seems to be best suited for ADME profiling.

3) Protein Homology Modeling of Drug Efflux Pumps (Michael Demel)
Overexpression of membrane-bound drug efflux pumps was identified as one of the predominant mechanisms responsible for development of multiple drug resistance in tumor therapy and antibacterial treatment. These ATP-dependent, highly efficient efflux proteins transport a wide variety of structurally and functionally diverse drugs. Inhibition of these efflux pumps gives rise to restoration of drug sensitivity to multiresistant cells and bacteria. The structure and molecular mechanisms of action of these pumps needs still to be elucidated. Recently, X-ray structures of three bacterial efflux pumps have been published. These structures serve as template for our protein homology modeling approach targeting human transporters. However, models obtained need to be refined using molecular dynamics simulations. Due to the fact, that these proteins usually show 12 transmembrane regions and that the interaction with both substrates and inhibitors takes place within the membrane environment, dynamics simulations have to include both membrane and aqueous phase. This requires front end computing facilities. After refinement, the models will be used for in silico screening of compound libraries to identify of new inhibitors.