1. Development of computational tools for
Integrated structure- and ligand based in silico approach to predict inhibition of cytochrome P450 2D6
Development of computational tools for
drug design
Seminar

2. INTRODUCTION
CYP2D6 (Cytochrome P450, family 2, subfamily D, polypeptide 6)
Member of cytochrome P450 superfamily of proteins.
Expressed in liver, intestine, kidney and brain cells.
Interacts with many drugs used for regulation of central nervous system
(psychtopics) or the cardiovascular system
(anti-arrhythmic drugs).
Metabolizes about 30% of drugs.

3. INTRODUCTION
CYP2D6 (Cytochrome P450, family 2, subfamily D, polypeptide 6)
Many drugs can inhibit the activity of CYP.
Inhibition of CYP by drug leading then to drug-drug interactions.
Thus, predicting potential CYP
inhibition is important in
early-stage drug discovery.
יש תרופות שעוברות מטבוליזם רק ע"י CYPD6

4. INTRODUCTION Modeling studies of CYP2D6-drug interactions
- A number of modeling studies have been undertaken to understand the molecular basis of CYP2D6-drug interactions and CYP2D6-related metabolism or inhibition.

5. AIMS
To develop an in silico approach for the prediction of CYP2D6 inhibition combining the knowledge of the protein structure and its dynamic behavior.
Structure-based in silico approach
Ligand-based in silico approach
Comprehensive analysis of CYP2D6 inhibition.
To develop an integrated structure- and ligand- based in silico approach able to predict inhibition of CYP2D6.
Structural information for CYP2D6 based on the available crystal structures and molecular dynamic simulations which take into account conformational changes of the binding site occurring due to the presence of diverse ligands.
The approach also includes information from experimental inhibition studies and chemical information from ligands.

6. Conformational ensemble
METHODS
MD simulations
X-ray structures of CYP2D6
Free
Ligand (7 structures, RMSD<1A)
Ligand binding conformation (without ligand)
Conformational ensemble
generation
PubChem
Figure S6. Workflow of the performed protocol for prediction of CYP2D6 inhibitors.
Ligand based Quantitative Structure-Property Relationships (QSAR) modeling.
Structural information for CYP2D6 based on available crystal structures (wang et al., 2012)
Large number of experimentaly validated CYP2D6 binders available in the PubChem BioAssay database
molecular dynamics simulations (MD)- to take into account conformational changes of the binding site.
Constructed combined models based on topological information of knowen CYP2D6 inhibitors and predicted binding energies computed by docking on both X-ray and MD protein conformations.
Modeling using three learning algorithms-
- support vector machine
- RandomForest
- NaiveBayesian
In the field of molecular modeling,docking is a method which predicts the preferred orientation of one molecule to a second whenbound to each other to form a stable complex.[1] Knowledge of the preferred orientation in turn may be used to predict the strength of association or binding affinitybetween two molecules using, for example, scoring functions. Molecular docking is one of the most frequently used methods instructure-based drug design, due to its ability to predict the binding-conformation of small moleculeligands to the appropriate targetbinding site. Characterisation of the binding behaviour plays an important role in rational design of drugs as well as to elucidate fundamental biochemical processes.
Protein-Ligand
Docking
Energy Minimization
Hierachical
classification

7. METHODS Compound dataset preparation Compound dataset preparation
Small molecules identified to be competitive binders (potentially inhibitors) or non-binders (non-inhibitors) for CYP2D6
have been collected in PubChem BioAssay (Wang, et al., 2012). We extracted 8614 molecules from the bioassay ID 891
(Fig. S3) In order to select only drug-like molecules, the entire collection of compounds was filtered using the FAF-Drugs3
server (Lagorce, et al., 2011) and an in-house developed “soft” drug-like filter for physicochemical properties (see Fig. S3)
without removing toxic/reactive/PAINS (Pan Assay Interference Compounds). After filtering, 1996 molecules were discarded.
Among the 6463 remaining drug-like compounds, the bioassay tagged 740 compounds as active, 4426 as inactive and
others as uncertain. We kept the most active inhibitors with AC50 ≤ 10 μM, which represent 427 compounds (AC50, “activity
concentration 50” refers to the concentration that is required to elicit half-maximal effect), and 4426 non-inhibitors showing
<20% inhibition at 57 μM concentration. In order to select chemically diverse molecules for validation of our approach
and to avoid possible over-representation of a chemical series, we performed several initial tests to cluster the compounds
using the fingerprint FCFP_4 as implemented in Pipeline Pilot v.7.5 (SciTegic, Inc/Accelrys) with a Tanimoto similarity criterion
of 0.8. The final dataset was composed of the centroid of each cluster derived from the classification, in total 3345
molecules, with 343 active and 3002 inactive compounds. Only 70 molecules among the 3002 non-inhibitors have structural
similarity with 60 molecules among the 343 inhibitors (Tanimoto > 0.8 when using fingerprint FCFP_4). The 3D structures
of the molecules were generated using the freely available web-server Frog2 (Miteva, et al., 2010). The procedure was
launched keeping a maximum of 1 stereoisomer per compound without generating multiple ring conformations. The molecules
were protonated at pH 7 using the major macrospecies option of the ChemAxon ( calculator plugins.

8. METHODS Machine Learning - Protein ligand binding energies
Figure S6. Workflow of the performed protocol for prediction of CYP2D6 inhibitors.
Ligand based Quantitative Structure-Property Relationships (QSAR) modeling.
Structural information for CYP2D6 based on available crystal structures (wang et al., 2012)
Large number of experimentaly validated CYP2D6 binders available in the PubChem BioAssay database
Machine Learning
- Protein ligand binding energies
- Molecular properties (ECFP)

9. RESULTS Analysis of CYP2D6 X-ray structures CYP2D6 +Ligand CYP2D6 Free
Figure S2. X-ray structures of CYP2 family represented in cartoon and the heme moiety in sticks colored in orange. A:
Rebuilt apo x-ray structure PDB ID 2F9Q colored in magenta; B: Holo X-ray structure PDB ID 3QM4 colored in cyan and prinomastat
co-crystallized colored in orange; C: Superimposition of both structures, apo in magenta and holo in cyan. The α helix
F’ seen in the holo structure is circled in black; D: Superimposition of CYP2A6 holo structure PDB ID 1Z10 colored in green,
CYP2C8 holo structure PDB ID 2NNJ colored in cyan and CYP2C9 apo structure PDB ID 1OG2 colored in magenta.

10. RESULTS
Active site key residues of CYP2D6 X-ray structures from different points of view
Figure S3. Active site key residues of CYP2D6 X-ray structures. Prinomastat and the heme moiety are represented
in sticks colored in orange, the X-ray structures in cartoon, with the key residues in stick colored in magenta and cyan
for 2F9Q and 3QM4 respectively, from different points of view, A, B et C.
CYP2D6 Free
CYP2D6 +Ligand
Ligand

11. RESULTS Docking of ligands into the apo X-ray CYP2D6 (prior to MD)
Docking of ligands into the apo X-ray CYP2D6 (shown in green) (PDB ID 2F9Q) prior to MD. (A) Propafenone shown in sticks in magenta atom type. (B) Mexiletine shown in sticks in cyan atom type. (C) Codeine shown in sticks in mauve atom type
Propafenone
Mexiletine
Codeine
-8.9 Kcal/mol
-8.2 Kcal/mol
-7 Kcal/mol

12. RESULTS MD-derived structure classification Selection of on per group
Conformational sampling
Structure 6
3000
CYP2D6 free X-ray 3
CYP2D6+ Ligand X-ray 7
CYP2D6 -Ligand X-ray 8
CYP2D6 + propatenone 5
CYP2D6 + mexiletine
CYP2D6 + codeine 34
18,000
MD-derived structure classification. This protocol has been applied in the same way for the X-ray structure 3QM4 apo and the X-ray structure 3QM4 in complex with codeine

13. RESULTS Enrichment curves of CYP2D6 MD1- CYP2D6 free
Figure S5. Enrichment curves of CYP2D6. A: Enrichment of the two X-ray structures, MD1 (derived from MD
started from 2F9Q apo) and MD2 (derived from MD started from 3QM4 without bound ligand); B: Enrichment of the
two X-ray structures, MD3 and MD4 (derived from MD started from 2F9Q and docked propafenone); C: Enrichment of
the two X-ray structures, MD5 and MD6 (derived from MD started from 2F9Q and docked mexiletine
MD1- CYP2D6 free
MD2- CYP2D6 -ligand
MD3- CYP2D6 +propafenone
MD4- CYP2D6 +propafenone
MD5- CYP2D6 +mexiletine
MD6- CYP2D6 +mexiletine

14. RESULTS
Key residues of binding sites of the MD-derived structures in comparison with the free X-ray structure.
CYP2D6 Free
MD1- CYP2D6 free
MD2- CYP2D6 -ligand
MD3- CYP2D6 +propafenone
MD4- CYP2D6 +propafenone
Key residues of binding sites of the MD-derived structures in comparison with the holo X-ray structure colored in cyan. (A) MD1 is colored in mauve; (B) MD2 is colored in blue; (C) MD3 is colored in green and MD4 in salmon; (D) MD5 is colored in yellow and MD6 in orange
MD5- CYP2D6 +mexiletine
MD6- CYP2D6 +mexiletine

15. RESULTS
Enrichment curves obtained using docking into the structures: the X-ray ones: 2FQ9, 3QM4; the MD structures: MD2, MD4 and MD6
Accuracy of the combined conformations: 70%
Enrichment curves obtained using docking into the structures: the X-ray ones: 2FQ9, 3QM4; the MD structures: MD2, MD4 and MD6. The combined enrichment is obtained by taking the best binding energies predicted by docking into MD2, MD4 and MD6
Identified three MD-derived structures that are capable all together to better discriminate inhibitors and non-inhibitors compared with individual CYP2D6 conformations, thus ensuring complementary ligand profiles.

16. Training set validation
METHOD
Machine learning classification modelling
3002 inactive
343 active
343 inactive
Inhibition models based on classical molecular descriptors and predicted binding energies were able to predicted binding energies CYP2D6 inhibition with an accuracy of 78% on the training set and 75% on the external validation set.
20%
External validation
80%
Training set validation
- Support vector machine
- Random forest
- Naïve Bayesian
- Protein ligand binding energies
- Molecular properties (ECFP)

17. RESULTS
Inhibition models based on classical molecular descriptors and predicted binding energies were able to predicted binding energies CYP2D6 inhibition with an accuracy of 78% on the training set and 75% on the external validation set.

18. CONCLUSIONS
The method that takes into account conformational changes of CYP2D6 and ligand properties allow to better discriminate inhibitors and non-inhibitors compared to a single CYP2D6 conformations, thus ensuring complementary ligand profiles.
We found a set of modeled CYP2D6 conformations, which all together are able to correctly retrieve, after docking, 70% of the knowen CYp binders.
The computational protocol integrating docking into this pool and machine learning-based modeling can be successfully applied to predict CYP2D6 inhibitors with 75% of Success.

19. FUTURE WORK
CYP2D6 is highly polymorphic with more than hundred of allelic variants.
Structure based approach will be essential to account for patient mutations in order to predict inhibition for personalized medication.

20. Acknowledgement