Ligand- and Structure-Based Pregnane X Receptor Models

Abstract

The human pregnane X receptor (PXR) is a ligand dependent transcription factor that can be activated by structurally diverse agonists including steroid hormones, bile acids, herbal drugs, and prescription medications. PXR regulates the transcription of several genes involved in xenobiotic detoxification and apoptosis. Activation of PXR has the potential to initiate adverse effects by altering drug pharmacokinetics or perturbing physiological processes. Hence, more reliable prediction of PXR activators would be valuable for pharmaceutical drug discovery to avoid potential toxic effects. Ligand- and protein structure-based computational models for PXR activation have been developed in several studies. There has been limited success with structure-based modeling approaches to predict human PXR activators, which can be attributed to the large and promiscuous site of this protein. Slightly better success has been achieved with ligand-based modeling methods including quantitative structure–activity relationship (QSAR) analysis, pharmacophore modeling and machine learning that use appropriate descriptors to account for the diversity of the ligand classes that bind to PXR. These combined computational approaches using molecular shape information may assist scientists to more confidently identify PXR activators. This chapter reviews the various ligand and structure based methods undertaken to date and their results.

Appendix

Structure-based Methods. Data collection: A comprehensive hPXR dataset consisting of a variety of chemical classes, namely, androstanes, pregnanes, estratrienes, bile acids, and regular xenobiotics is available from previously published studies. (53, 54). The Environmental Protection Agency has developed a dataset comprising of everyday chemicals, pesticides and other related chemicals aptly called Toxcast dataset (Toxcast^TM) which is available for use by researchers ((47, 49). For all these datasets, either the direct binding data to hPXR or relative fold induction changes have been published. However, the EC₅₀ cutoff value used for classifying the molecules into hPXR activators and nonactivators is not always available or does not follow a strict pattern (25, 33), SMILES strings for compounds listed in datasets or 2D structures in mol or sdf formats can be obtained from the respective publications (25). Using these as input, single low energy three-dimensional structure of molecules can be generated using Molecular Operating Environment (MOE, Chemical Computing Group, Montreal, Canada) or CORINA (Molecular Networks GmbH, Nägelsbachstr. 25, 91052 Erlangen, Germany. http://www.mol-net.de) with partial charges assigned according to the Gasteiger-Marsili scheme (55).

Molecular Descriptors. hPXR binds a variety of ligands that differ in shape, size, and chemical composition. However, analysis of well known hPXR agonists show that the interactions are dominated by hydrophobic and hydrogen bonding features of the ligands. Hence to capture these properties, molecular descriptors that represent shape, size, flexibility and hydrogen bonding, and hydrophobic properties must be chosen. These include FCFP_6 fingerprints, volume, weight, KierA1-A3, Kier1-3, number of rotatable bonds, number of rings and KierFlex, electrostatic features like logP, TPSA, logs of lip_don, lip_acc, number of N atoms, and number of O atoms. The values for these specific molecular descriptors can be derived from MOE. In addition to analyze the specific role of shape and electrostatics, specific shape based descriptors such as shape signatures can be derived (56).

Molecular Docking. Five crystal structures of hPXR cocrystalized with a variety of ligands are available in the protein databank (PDB) under the codes 1 M13, resolution 2.00 Å (36), 1SKX, resolution 2.80 Å (38), 2O9I, resolution 2.80 Å (37), 1NRL, resolution 2.00 Å (36), and 2QNV, resolution 2.80 Å (37). In addition, another structure in complex with 17-β estradiol that is yet to be deposited in PDB was also used for docking studies (57). In all cases, the protein structure is first prepared by removing the crystal structure ligand and adding hydrogen atoms to the amino acids and the resulting structures are energy minimized to remove any steric contacts. All amino acids within 6 Å of the cocrystallized ligand are generally chosen as being part of the binding site. The docking program GOLD (ver 4) (58) or FlexX could be used for docking. GOLD uses a genetic algorithm to explore the various conformations of ligands and flexible receptor side chains in the binding pocket. In our studies, we have chosen to perform 20 independent docking runs for each ligand to sample the ligand and protein conformational space. The resulting docked complexes can be scored using GoldScore and ChemScore.

Scoring functions. One of the bottlenecks in docking studies is the scoring functions. Although most docking programs are capable of sampling the ligand in the binding pocket and generating solutions, the scoring functions are not sensitive enough to identify the truly best docked solutions. This is because, most scoring functions are empirical energy based schemes as shown below:

$$ \Delta {G_{\rm{bind}}} = \Delta {G_{\rm{solvent}}} + \Delta {G_{\rm{conf}}} + \Delta {G_{\rm{int}}} + \Delta {G_{\rm{rot}}} + \Delta {G_{\rm{t}}}/r + \Delta {G_{\rm{vib}}} $$

Where ΔG_bind is the binding free energy, ΔG_solvent is the penalty for desolvation, ΔG_int is the internal energy, ΔG_rot represents the energy contribution for bond rotation, and ΔG_vib represents the energy contribution for vibration component. However, the terms for solvation is approximated to a best fit model and the terms for entropy are most likely dropped from the energy equation due to the complexity involved in computing the entropy factors. Thus, in practice no single scoring function can work for every target and has to be customized to suit the needs of the target. In case of hPXR, given that the binding site is very promiscuous and binds a variety of ligands, developing a scoring scheme is challenging. We have used a number of methods to derive consensus scoring schemes including contact score, shape based scores, and molecular descriptor based schemes.

• 1. Contact scoring scheme. The docked receptor–ligand complexes were scored using a contact based scoring function. Accordingly, an in-house program was used to scan the docked complexes for contacts between the ligand and protein atoms (56). These contacts were scored based on a weighting scheme that was derived from the nature of interaction between the ligands cocrystallized with hhPXR. For example, hyperforin has hydrogen bond interactions with residues Gln285, His407, and Ser247 of the hhPXR protein in the crystal structure (PDB ID:1 M13) (Fig. 3). Thus the contact scoring function weighted all those docked protein–ligand complexes that featured the hydrogen bonding between the ligands and these three residues, higher than the rest of the interactions. Similarly, other nonbonded interactions were weighted based on the interactions of the ligands in the hPXR crystal structures. All interaction scores were then summed and normalized against all crystal structures. A consensus scoring scheme was developed for final classification based on the following rule: Only those compounds that had at least half the value of the highest GoldScore and a nonzero contact score were assigned as activators and the rest of the molecules were classified as nonactivators.

• 2. Shape based scoring scheme. In this scheme, the ligands were compared with the hPXR ligands from the five crystal structures for their shape based similarities using two different approaches. The first was based on the 2D similarity encoded in MDL public fingerprint keys calculation using Discovery Studio 2.0 (Accelrys, San Diego, CA). The Tanimoto coefficient was used as the metric to compare the molecular fingerprints. The coefficients varied between 0 and 1, where 0 meant maximally dissimilar and 1 coded for maximally similar. The Tanimoto coefficient between fingerprints X and Y has been defined to be: [number of features in intersect (A,B)]/[number of features in union(A,B)], where A and B are two compounds.

  In an approach, the 3D shapes of the molecules from the combined dataset were compared with the shapes of each of the four crystal structure ligands. This was achieved by comparing their corresponding 1D Shape Signatures and a dissimilarity score was computed for each ligand pair. The dissimilarity score was then converted to a similarity score, which was in turn used as weighting factor for the GoldScore. In all these scoring schemes the consensus score was calculated as shown below in Eq. (1).

• 3. Molecular descriptor based scoring. In this scheme, the molecular descriptors computed using MOE were used to calculate Euclidean distances from the crystal structure ligands. These Euclidean distances were used as weighting factors to GoldScore. Similarly, the values of the molecular descriptors were also used to calculate Tanimoto similarity indices (59) with reference to the cocrystal structure ligands. The values for the Tanimoto indices for each ligand in the combined set were calculated against each of the crystal structure ligands and then used as weighting factors to the GoldScores.

  The weighted docking score of an active compound j with i conformations was described as

  $$ {S_{{i,j}}} = {w_i}{s_{{ij}}} $$

  (1)

  where s _ij was the original GoldScore for the compound i in its jth conformation and w _i is the weighting factor for compound i from either of the schemes described above.