A combinatorial approach for the discovery of cytochrome P450 2D6 inhibitors from nature

The human cytochrome P450 2D6 (CYP2D6) enzyme is part of phase-I metabolism and metabolizes at least 20% of all clinically relevant drugs. Therefore, it is an important target for drug-drug interaction (DDI) studies. High-throughput screening (HTS) assays are commonly used tools to examine DDI, but show certain drawbacks with regard to their applicability to natural products. We propose an in silico – in vitro workflow for the reliable identification of natural products with CYP2D6 inhibitory potential. In order to identify candidates from natural product-based databases that share similar structural features with established inhibitors, a pharmacophore model was applied. The virtual hits were tested for the inhibition of recombinant human CYP2D6 in a bioluminescence-based assay. By controlling for unspecific interferences of the test compounds with the detection reaction, the number of false positives were reduced. The success rate of the reported workflow was 76%, as most of the candidates identified in the in silico approach were able to inhibit CYP2D6 activity. In summary, the workflow presented here is a suitable and cost-efficient strategy for the discovery of new CYP2D6 inhibitors with natural product libraries.


In silico dataset for the screening of novel CYP2D6 inhibitors
Building of the 3D-database Every single 2D-database was treated and built up separately using the BEST mode and creating a maximum of 255 conformers for each single substance in the 2D-database, according to the findings from the validation of the in silico workflow. From all of the 2,147 single substances that make up the 17 2D-databases we could recover 86 % i.e. 1,847 compounds that account for the 3D-databases, whereas 11 databases were built without declines and 6 databases were built with an overall loss of 300 compounds (Supplementary Table S1).

Screening of the 3D-database
We screened the 17 3D-databases with overall 1,847 compounds and applied a rigid search using the pharmacophore model for CYP2D6 inhibitors 1 that proved to be an appropriate tool in the preceding validation workflow. From the screened 1,847 compounds in 17 3D-databases, we identified 4.1 % i.e. 75 compounds that fitted the model for potential CYP2D6 inhibitors and that were located in 6 different 3D-databases. In 10 databases we were not able to identify compounds that fitted the pharmacophore model for CYP2D6 inhibitors (Supplementary Table  S1).

CYP2D6 inhibition of virtual hits already reported in the literature
In the course of the screening of the various databases we found 75 hits that fitted the pharmacophore model. From these hits, 23 were selected for the in vitro screening. We selected the hits according to non-available CYP2D6 inhibition data, plant origin and availability. The 52 not-tested hits included some duplicate entries and were composed of natural, synthetic or semisynthetic compounds. Focusing on the compounds from plant origin, highly potent and well-known CYP2D6 inhibitors were found by the pharmacophore model, like ajmalicine and quinidine. On the one hand, this was expected because some of these compounds have been used for model generation 1 already. On the other hand, additional active compounds were found, which underlines the power of the in silico pre-selection step (Supplementary Table S2).

Preparation of the compounds from the university collaborators-UIBK
The compounds 1 and 2 were isolated at the Institute of Pharmacy / Pharmacognosy, University of Innsbruck. The identification of the structure was performed by mass and NMR spectroscopy as well as by comparison (TLC, HPLC) with authentic samples. The compound 3 was isolated and identified as described previously 14 .

Preparation of the compounds from the university collaborators-Rohan A. Davis
The isolation and identification of compounds 4 and 5 has been described elsewhere 15,16 . The compounds 6, 7, 21, 22 and 23 were purchased from PhytoLab (www.phytolab.com).

Incubation conditions for the in vitro assay and the luminescence quenching control
Supplementary

Establishment of the re-docking process
In order to predict the binding pose of an inhibitor in the active site cavity, the program settings were validated by re-docking of ligands to a CYP2D6 structure that was solved as a cocrystallized complex with inhibitors in the active site cavity by Wang et al (PDB entry 4WNT 19 ). The re-docking helps to identify the software settings with which the crystallographic and thus known binding pose can be reproduced by a computational docking method within a range of a root mean square deviation (RMSD) at or below 2 Å. The re-docking experiments were performed using the genetic algorithm implemented in GOLD 17,18 . Independent redocking models for each of the crystal structures 4WNT 19 , 4WNU 19 and 4XRZ 20 were generated that exhibit a RMSD of 0.19, 0.95 and 0.60 Å for the core to the co-crystallized ligand compared to the re-docked ligand, respectively. Furthermore, each re-docking model was challenged with a self-generated ligand to avoid bias by using the bioactive conformation of the ligand for re-docking. For all three crystal structures i.e. 4WNT, 4WNU and 4XRZ, the RMSD of the core, co-crystallized ligand to the self-generated re-docked ligand was 0.37, 0.51 and 0.74, respectively. The docking and re-docking workflow obtained from 4WNT, with ajmalicine co-crystallized in the active site cavity ( Supplementary Fig. S2) turned out to be best suited for the further studies.
Supplementary Figure S2. Visualization of ajmalicine in the active site cavity of CYP2D6 with a docking model. The docking model for CYP2D6 was based on the crystal structure cocrystallized with the inhibitor ajmalicine (PDB code 4WNT), having a RMSD value below 0.4 Å of the core molecule to the re-docked molecule. The amino acids highlighted in magenta listed on the left side were found to be important interaction and binding partners in previous studies. The prosthetic heme-b group is highlighted in red.

Interactions of the three most potent inhibitors in the active site cavity of 4WNT
Supplementary

Supplementary Figure S3. Comparison of the docking poses of the three most potent inhibitors in different PDB structures:
The docking model selected in the redocking process was used for the prediction of the potential binding poses of the newly found inhibitors. Figure  S3 shows the compounds 21 (I), 19 (II) and 9 (III), which were the three most potent inhibitors in the in vitro screening. Each compound was docked in the active site cavity of the PDB structures 4WNT 19 , 4WNU 19 and 4XRZ 20 , respectively. For a better comparison, the docked inhibitor is presented on the left sides, whereas the core and the redocked core molecule of the corresponding crystal structure can be found on the right sides (I, II and III_A-F). The docked inhibitors are colored in grey/light blue, the heme-moiety in yellow and the amino acids in magenta. The numbers indicate the follwing amino acids: 1-Ser304, 2-Asp301, 3-Phe120, 4-Phe247, 5-Glu216, 6-Leu213, 7-Gln244 and 8-Phe483.

CypRules performance comparison
The direct comparison of the pharmacophore model performance with CypRules is actually rather challenging. The data set used for comparison needs to be unknown prior to both of the classifiers, because literature datasets , e.g. PubChem, will certainly contain inhibitors that have been used for training the models and thus are already known. The only data set for which this novelty was granted was our own test compound set from this study.
Both the active and the inactive compounds ( Table 2 in the main manuscript) were tested on the freely accessible CypRules platform, which is a rule-based P450 inhibition prediction server (http://cyprules.cmdm.tw) 21 . For the calculations, those compounds were classified as active, which exhibited CYP2D6 IC50 values up to 10 µM, whereas inactives showed less than 50 % total luminescence at a concentration of 100 µM. The rule-based CYP2D6 inhibition prediction server correctly identified eight out of the thirteen actives as inhibitors. A direct comparison of the performance with the approach presented in this manuscript is difficult because only one of the compounds that was predicted as inactive by the model (compound 24) was part of our test-set. Thus an assessment of the prospective model performance regarding compounds that are predicted to be inactive is not possible. However, the screening of the PubChem BioAssay database showed that our model successfully classified 96 % of the inactive compounds correctly.