Machine Learning Methods and Docking For Predicting Human Pregnane X Receptor Activation

Abstract

The pregnane X receptor (PXR) regulates the expression of genes involved in xenobiotic metabolism and transport. In vitro methods to screen for PXR agonists are used widely. In the current study, computational models for human PXR activators and PXR non-activators were developed using recursive partitioning (RP), random forest (RF) and support vector machine (SVM) algorithms with VolSurf descriptors. Following 10 fold randomization, the models correctly predicted 82.6 % to 98.9 % of activators and 62.0 % to 88.6 % of non-activators. The models were validated using separate test sets. The overall (n = 15) test set prediction accuracy for PXR activators with RP, RF and SVM PXR models is 80 to 93.3 %, representing an improvement over models previously reported. All models were tested with a second test set (n =145) and prediction accuracy ranged from 63−67 % overall. These test set molecules were found to cover the same area in a principal component analysis plot as the training set, suggesting the predictions were within the applicability domain. The FlexX docking method combined with logistic regression performed poorly in classifying this PXR test set compared with RP, RF and SVM, but may be useful for qualitative interpretion of interactions within the LBD. From this analysis, VolSurf descriptors and machine learning methods had good classification accuracy and made reliable predictions within the model applicability domain. These methods could be used for high throughput virtual screening to assess for PXR activation, prior to in vitro testing to predict potential drug-drug interactions.

Introduction

The human pregnane X receptor, PXR (NR1I2; also known as SXR or PAR) is a transcriptional regulator of a large number of genes involved in xenobiotic metabolism and excretion. The genes regulated by PXR include cytochrome P450 (CYP) 3A4 (1-3), CYP2B6 (4), aldehyde dehydrogenases, glutathione-S-transferase, sulfotransferases, organic anion transporter peptide 2, and multi drug resistance protein 1 and 2 (5, 6) as well as others. Human PXR activators include a wide range of prescription and herbal drugs such as paclitaxel, troglitazone, rifampicin, ritonavir, clotrimazole, and St. John's Wort which can be involved in clinically relevant drug-drug interactions (7). In addition to xenobiotics, PXR is also activated by pregnanes, androstanes, bile acids, hormones, dietary vitamins and a wide array of endogenous molecules reviewed recently (8).

The PXR ligand binding domain (LBD) consists of 12 α-helices that fold to form a hydrophobic pocket and a short region of β-strands. The pocket is lined with twenty eight amino acid residues, twenty hydrophobic, four polar and four charged (9-13). The potential for molecules to bind in numerous locations in the LBD complicates the reliable prediction of PXR activators (A) or non-activators (N) using structure based drug design methods alone. Computational models ranging from ligand based pharmacophores (14-17), quantitative structure activity relationships (QSAR) (18-20), and machine learning methods (20), to homology modeling with molecular dynamics (21) (for identifying protein-co-repressor interactions), represent predominantly reports to predict PXR ligand binding (8) to differing degrees. These previously described computational methods focused on diverse structural types for agonists and in one case used structural analogs (8) which may have assessed specific binding locations within the LBD, such as that for steroidal compounds. A likely consensus has emerged across the different QSAR modeling methods that PXR agonists are required to fit to multiple hydrophobic features and at least one hydrogen bond acceptor (and in some cases an additional hydrogen bond donor feature) (8). A further qualitative observation from these previous studies is the dependence of the resulting agonist QSAR or pharmacophore models on the molecules used in the training set, and potential for overlap of multiple models derived from different molecules (8). It should also be noted that rarely do the published QSAR models utilize a large external test set to validate the predictive nature or assess the applicability domain (22-24) of the training and test sets, i.e. how structurally similar do the molecules in the training and test set have to be for accurate predictions. This is especially important to build confidence in the use of these methods with such structurally promiscuous proteins as PXR. One of the limitations of using published data for PXR is that only a small fraction of the data available reports quantitative EC₅₀ data, (e.g. much of the work is published as greater or less than a cutoff value e.g. 100 μM). Therefore there are currently no widely available large, diverse continuous datasets to enable quantitative QSAR modeling for human PXR. Two PXR machine learning studies have been published recently with relatively large training sets (≥ 99 molecules) using recursive partitioning (19), support vector machine (SVM), K- nearest neighbors (k-NN) and probabilistic neural network (PNN) (20). In the latter case, binary classification data for 98 human PXR activators and 79 non-activators were used (20) to predict between 80.8 to 85.0 % of human PXR activators and 67.7 to 73.6 % of human PXR non-activators (in the training set); the test set prediction accuracies in this same study ranged from 53.3−66.7 % for 15 known human PXR activators across the three machine learning methods, with SVM performing the best (20).

A structure-based alternative to understanding small molecule-protein interactions is to dock molecules into proteins. These methods have resulted in the discovery of novel inhibitors for many targets (25-28) and have been applied to proteins such as drug metabolizing enzymes which are also relatively promiscuous in their ligand binding (27-29). Docking and scoring ligands in target protein binding sites is a challenging process (30-32) and the performance of different implementations of scoring functions have been found to be target dependent (33). The use of consensus scoring has also been recommended to improve the “hit” enrichment (34). Docking of a small number of molecules into the PXR structure 1NRL has been used previously to design molecules that are weaker agonists (35) and an automated docking method (GOLD (36) which maintains the protein as rigid), was used to flexibly dock azole PXR antagonists onto the outer surface at the AF-2 site (8). We are not aware of any studies using docking as a screening tool to predict the potential for PXR activation on a larger scale with diverse drug-like libraries of molecules and this may be due to the challenge of the protein promiscuity.

In the current study we have compared RP, RF and SVM machine learning methods for building human PXR models derived with VolSurf 3D descriptors (which have been widely applied for ADME/Tox and drug discovery target modeling (37-39)). In addition for comparison, we have used FlexX docking (40, 41) of molecules into the crystal structures of human PXR combined with logistic regression. The predictive ability of the classification models and docking models was evaluated using a novel large external test set containing 145 human PXR activators and non-activators (using data for molecules generated in this study and published in the literature) (42-44). Our aim was to identify the most appropriate computational approach to predict whether a molecule was likely to be a human PXR agonist (activator) and in so doing, prioritize molecules for in vitro testing.

Materials and Methods

Reagents and Plasmids

The construction of a HepG2 (human liver) cell line stably expressing human SLC10A1, a transporter that can take up conjugated bile salts has been reported before in detail (45). Human PXR was expressed as a full-length protein and CYP3A4-PXRE-Luc, which contains promoter elements from CYP3A4 recognized by the PXR DNA-binding domain, was used as the reporter construct. The plasmids for human PXR, human SLC10A1, CYP3A4-PXRE-Luc, and empty vectors pSG5 were generously provided by S.A. Kliewer, J.T. Moore, and L.B. Moore (GlaxoSmithKline, Research Triangle Park, NC).

Reporter gene assay with HepG2 cells

Human PXR activation in the HepG2 human liver cell line was determined by a luciferase-based reporter assay as previously described (8, 46).

Data sets and molecular descriptors

A previously published dataset was used for model building (training) in which compounds with EC₅₀<100 μM (n = 98) were classified as human PXR activators, whereas compounds with EC₅₀>100 μM (n = 79) were classified as human PXR non-activators (20). One small test set also published by the same group consisted of 15 human PXR activators (20). The molecular structures encoded as SMILES strings (47) were downloaded from the supplementary information tables in the original publication (20). We have additionally developed our own novel test set (n = 145) consisting of PXR activators (n = 82) and PXR non-activators (n = 63) from our exhaustive literature searching (over 5 years) and data generated in this study (see above) that are not in the current training sets (8, 42, 48). The SMILES string for each molecule was downloaded from either PubChem (http://pubchem.ncbi.nlm.nih.gov/), ChemSpider (http://www.chemspider.com/) or sketched using the BUILDER module of SYBYL. The SMILES strings for the previously published training and test sets were converted to SYBYL MOL files using an in-house script. The MOL files were then subsequently minimized in SYBYL (Tripos, St Louis, MO). Energy minimizations were performed using the Tripos force field (49) and Gasteiger-Hückel charges with a distance-dependent dielectric constant and conjugate gradient method with a convergence criterion of 0.001 kcal/mol.

Eighty-six VolSurf descriptors (38) were calculated from the 3D molecular fields using VolSurf 4.0 implemented in SYBYL. Three different probes including water (OH2), dry and amphipathic (BOTH) were used for descriptor calculation. VolSurf descriptors include descriptors for size, shape, hydrophilic and hydrophobic regions, and interaction energy as well as others.

Machine learning model building and validation

RP calculations were performed using the rpart module of the R package (50). Briefly, there are (n) number of compounds and each compound contains a descriptor (x variables) and a class label (activator or non-activator). RP is a technique that builds a set of classification rules based on descriptor information. At the base of the tree, all the n compounds are mixed into a single group. The base is then split into smaller and smaller samples (every sub sample is called node) by choosing a descriptor and the corresponding threshold value to divide the sample. If the compound's descriptor value is above the threshold then it is assigned to one branch and if it is lower then it is assigned to the other branch. The ultimate goal is to separate PXR activators from non-activators. The grown tree can then be used to classify test compounds to activator or non-activator classes.

The R program was also used for RF calculations (51). In the RF method, which is an extension of RP, multiple trees are grown and predictions are made by averaging the multiple trees. The total number of trees was set to 1000. The other optimizable parameter in the random forest is m_try i.e. the number of descriptors (p) randomly sampled as candidates for splitting at each node. The number of descriptors was increased systematically with an increment of 5. The out of bag error (OBB) estimate can be considered as equivalent to a cross validation study. In OBB, one third of the compounds are randomly selected as a test set and a model is developed from the remaining compounds, this is then repeated multiple times. The optimum m_try was chosen such that %OBB is a minimum. Thus, a lower %OBB indicates higher accuracy of the model.

The Kernlab package in R was used for generating SVM models using the radial basis kernel function. The grid optimization approach was followed to obtain optimum values of C and gamma by varying these parameters from 2⁻⁵ to 2¹⁵ and 2⁻¹⁵ to 2³, respectively. The optimal C and gamma values for the resultant PXR model were 8 and 0.00781, respectively based on the lowest cross-validation error (52).

Docking

The molecules in the training and validation set were docked into 4 different ligand co-crystallized structures of PXR with hyperforin (PDB ID: 1M13, resolution 2.00 Å)(10), rifampicin (PDB ID: 1SKX, resolution 2.80 Å)(53) , T0901317 (PDB ID: 2O9I, resolution 2.80 Å)(54) and SR12813 (PDB ID: 1NRL, resolution 2.00 Å)(9) using FlexX (BioSolveIT, GmbH, Sankt Augustin) (40, 41). The FlexX program considers ligand flexibility by an incremental ligand placement technique while the receptor is considered as rigid. In total, 322 molecules (177 training and 145 testing) were docked into each crystal structure giving rise to 322*4 = 1288 molecules. For each ligand (n = 1288), 30 different docked poses were generated (1288*30 = 38,640 overall poses) and the best pose was selected based on the FlexX score. In all cases the active site was defined as the amino acid residues within 6Å of the co-crystallized ligand. The FlexX scores were then used as the independent (x) variable and actual PXR activators (A) / PXR non-activators (N) as the dependent (y) variable in a logistic regression analysis (Eq. 1). The PXR activators and PXR non-activators were assigned scores of 1 and 0 respectively. This logistic regression analysis was performed for molecules docked in all 4 structures. (Eq. 1).

$$\mathrm{Prob}(\mathrm{Y}=\mathrm{A})=\frac{1}{1+{\mathrm{e}}^{-({\alpha }^{\ast }\mathrm{FlexX}+\beta )}}\mathrm{or}\log \mathrm{odds}(\mathrm{Y}=\mathrm{A})={\alpha }^{\ast }\mathrm{FlexX}+\beta$$ (1)

Where, Prob and A represents probability and activator, respectively and α and β represents adjustable parameters. Two different confusion matrices were built for each crystal structure i.e. one using only successfully docked compounds and the other from both successfully docked and failed compounds. In the latter case the molecules which failed to dock inside the LBD of PXR were considered as non-activators.

Data analysis

The performance of RP, RF, SVM and docking were evaluated using true positive (TP), true negative (TN), false positive (FP), false negative (FN), sensitivity (SE) SE=TP/(TP+FN), specificity (SP) SP=TN/(TN+FP), overall prediction accuracy (Q) Q=(TP+TN)/(TP+TN+FP+FN) and Matthew's correlation coefficient (C) $\mathrm{C}=\left(\left({\mathrm{TP}}^{\ast }\mathrm{TN}\right)\text{-}\left({\mathrm{FN}}^{\ast }\mathrm{FP}\right)\right)∕\sqrt{(\mathrm{TP}+\mathrm{FN})(\mathrm{TP}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})}$ (Table 1) (20). Sensitivity and specificity were used to represent the prediction accuracy of PXR activators and non-activators, respectively. In the present context, TP is the number of true PXR activators, TN is the number of true PXR non-activators, FP is the number of falsely classified PXR activators and FN is the number of falsely classified PXR non-activators.

Table 1.

The predictive performance of RP, RF, SVM and docking with logistic regression for the training set (n = 177, 10-fold cross-validation study) and test set (n = 145). The values in the parenthesis (docking and logistic regression) include both failed and successfully docked compounds. RP = recursive partitioning; RF = random forest; SVM = support vector machine

Method	% SEa	% SPb	% Qc	Cd
Training
RP	95.92	77.21	87.57	0.754
RF	82.65	62.02	73.45	0.459
SVM	98.98	88.61	94.35	0.888
Docking + logistic regression	60.0 (45.92)	40.28 (45.57)	50.34 (45.76)	0.002 (−0.085)
Test set
RP	64.63	61.9	63.45	0.264
RF	64.63	66.67	65.52	0.310
SVM	68.29	65.08	66.9	0.332
Docking + logistic regression	73.42 (70.73)	17.54 (25.4)	50.0 (51.03)	−0.106 (−0.043)

^aSE=TP/(TP+FN)

^bSP=TN/(TN+FP)

^cQ=(TP+TN)/(TP+TN+FP+FN)

^d$\mathrm{C}=\left(\left({\mathrm{TP}}^{\ast }\mathrm{TN}\right)\text{-}\left({\mathrm{FN}}^{\ast }\mathrm{FP}\right)\right)∕\sqrt{(\mathrm{TP}+\mathrm{FN})(\mathrm{TP}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})}$

True positive (TP), true negative (TN), false positive (FP), false negative (FN), sensitivity (SE), specificity (SP), overall prediction accuracy (Q) and Matthew's correlation coefficient (C).

Results

The human PXR data for model development (training set n = 177) and a small initial test set (n = 15 activators) was taken from a recent publication (20), while the experimental data for the large external test set (n = 145 activators and non-activators) was collated in the current study and several recent publications (8, 42, 48). The predictive performance of RP, RF and SVM with VolSurf descriptors for predicting PXR activators and PXR non-activators were initially based on a 10-fold cross validation metric. A 10-fold cross validation was chosen in order to make direct comparisons with previously published models in order to assess the validity of our selected descriptors and algorithms (20). The overall training set prediction accuracy of PXR models were 87.5, 73.4, and 94.3%, respectively (Table 1 and Supplemental Table 1). The C value of the RP, RF and SVM PXR models were: 0.75, 0.46 and 0.89, respectively (Table 1). For comparison the previous best SVM reported (20) had a training set accuracy of 79.6% compared with 94.3% in this study. The C value ranges between +1 and −1, where a value of +1 indicates perfect prediction, −1 represents an inverse prediction and 0 indicates that the prediction is equivalent to a completely random prediction. The previous best SVM reported (20) had a training set accuracy of C value of 0.598 compared with 0.888 in this study. Based on the data distribution, the probability of randomly selecting a PXR activator or a PXR non-activator is 55.4 % and 44.6 %, respectively. RF would therefore appear to perform poorly compared with the SVM and RP. The 10 fold cross validation results would also indicate that in all cases the classification predictions are better than random. All the models were also validated using external test sets. The first small test set (using 15 known activators from a previous study (20)) resulted in a prediction accuracy for RP, RF and SVM PXR models from 80 to 93.33 % (Supplemental Table 1). The previous best SVM reported (20) correctly predicted 66.7 % of the 15 molecules compared with 93.33 % in this study. Therefore the overall prediction accuracies for the training and test set machine learning models reported here represents an improvement over the models reported previously using other molecular descriptors (20). The large 145 molecule test set (using PXR activators and PXR non-activators) resulted in overall prediction accuracies for RP, RF, SVM and the docking PXR logistic regression model of 63.4 %, 65.5 %, 66.9 % and 51 % respectively. (Table 1 and Supplemental Table 2). Interestingly RF performs better with the external test set than with internal testing for the training set, which would suggest that this model is not over-trained. Consensus scoring has been used previously with QSAR and docking applications (34, 55-57) and in some cases can improve predictions over individual methods (58, 59). In the current study a combination of the models indicated that a consensus approach did not improve the overall prediction accuracy (data not shown).

The RP method produces more interpretable models unlike those obtained from RF and SVM. The RP tree from the PXR activator and PXR non-activator model is shown in Figure 1. In general, the descriptors important for PXR activators and PXR non-activators are W1, CW1, CW5, IW7, ID1, ID7 and molecular weight (MW). These represent the first 4 descriptors from the OH2 probe and the next two are from the DRY probe. W1 accounts for polarizability and dispersion forces, CW1 and CW5 represent the extent of hydrophilic regions per surface unit, IW7 is an integer moment and measure the imbalance between the center of mass of a molecule and hydrophilic regions around it, ID1 and ID7 are also integer moments which measure the imbalance between the center of mass of a molecule and hydrophobic regions around it. The RP tree can also be used to develop simple rules for separating PXR activators from PXR non-activators.

Figure 1.

The recursive partitioning tree model separating PXR activators and PXR non-activators developed in this study. The tree method is described in the Materials and Methods. Molecules are assigned to branches when descriptor values are greater than or less than a threshold value described in the preceding node. The subscripts represent the OH2 and dry probe used in VolSurf calculations.

When comparing predictions for individual molecules, we made the following observations: a) members of the initial test set of activators such as diflubenzuron, is predicted as a non-activator by all the three machine learning methods in this study. Interestingly, the same molecule is also predicted as a PXR non-activator by the PNN and k-NN models in a previous study that used the same training set (20). b) fenbuconazole, another PXR activator from the previously used test set, is predicted as a PXR non-activator by all the models used previously (20), whereas the current RP and RF models correctly predicted it as a PXR activator in this study. The large test set used in this study contains 82 PXR activators and 63 non-activators and are predominantly a sub-set of available drugs (42), imidazole derivatives similar to clotrimazole, steroids (8), molecules with different heterocyclic ring systems (48) as well as many other diverse molecules. In this large test set, the PXR non-activators 5α-petromyzonol and 5β-cholan-3α,7α,12α,24-tetrol are predicted as PXR activators by all three QSAR methods in this study. The PXR activators 7-ketolithocholic acid and 12-ketolithocholic acid are predicted as PXR non-activators by all three QSAR methods. Interestingly, the unsubstituted compound lithocholic acid is correctly predicted as a PXR activator by all three methods. The generally low affinity of PXR activators raises the possibility that for some of the molecules in the published data, toxicity of the compounds may complicate determination of PXR activation. It may also be possible to improve the predictions for PXR activation using additional different descriptors and modeling approaches.

Previously we have used molecular similarity using the Tanimoto coefficient (60), or principal component analysis (PCA) to assess test and training set similarity (61). In our hands either of these approaches are useful to understand potential outlier predictions. In this study PCA was performed using VolSurf descriptors for the PXR data. The PCA score plot provides an estimate of the descriptor space of training and test set molecules (Figure 2) (61). Predictions can be considered as an extrapolation if the test set occupies a different descriptor space to the training set. It is clear from Figure 2 that the test set molecules occupy approximately the same descriptor space as that of the training set molecules and one would therefore expect the predictions to be relatively accurate. The first three principal components of the training and test sets explain 55.6 %, to 68.6 % of the variance, respectively. PCA analysis therefore did not explain the observed versus predicted PXR activation differences.

Figure 2.

The principal component analysis scores plot of PXR training (20) and test and validation sets using the VolSurf descriptors. The PXR activators are shown as filled circles training (black), small test set ((20), blue) and large test set (red), whereas, the non-activators are shown as empty circles training (black) and large test set (red). This orthogonal linear transformation, transforms the data to a new coordinate system. The x and y axes represent PCA scores from first and second principal components, respectively. The first three principal components of the training and test sets explain 55.6 %, to 68.6 % of the variance, respectively.

The logistic regression analysis method is a useful tool when the outcome of an experiment is binary (62), as in the current study either PXR activator or PXR non-activator. Four different models were generated by docking molecules to the four available co-crystal structures for PXR. The FlexX scores were used as the independent variable and the PXR activators or PXR non-activators (class variables) were used as dependent variables in the logistic regression analysis. Two different confusion matrices were built for each crystal structure i.e. one using only successfully docked compounds and the other from both successfully docked and failed compounds. In the second case, molecules which failed to dock inside the LBD of PXR were considered as non-activators. The logistic regression coefficients obtained from the FlexX scores were then used as the independent variable to predict activators/non-activators in the test set. The best model was obtained using the structure with SR12813 (PDB ID: 1NRL) which resulted in Eq 2 and was used to predict the large test set (Tables 1 and 2). The results for the other 3 complexes are not shown.

$$\mathrm{Prob}(\mathrm{Y}=\mathrm{A})=\frac{1}{1+{\mathrm{e}}^{-({0.0398}^{\ast }\mathrm{FlexX}-0.374)}}\mathrm{or}\log \mathrm{odds}(\mathrm{Y}=\mathrm{A})=-0.374-{0.0398}^{\ast }\mathrm{FlexX}$$ (2)

Table 2.

Prediction results for the large 145 molecule test set. Classifications: A = PXR activator (agonist); N = PXR non-activator. RP = recursive partitioning; RF = random forest; SVM = support vector machine

Compounds	Classification	Docking + logistic regression	RP	RF	SVM	Reference / EC50 value determined in this study (μM)
1,9-Dideoxydforskolin	A	N	A	A	A	(42)
Amlodipine	A	A	A	A	N	(42)
Bergamottin	A	N	A	A	A	(42)
Celecoxib	A	A	A	N	N	(42)
Diltiazem	A	A	A	N	A	(42)
Felodipine	A	A	A	A	A	(42)
Fluvastatin	A	A	N	A	A	(42)
Forskolin	A	A	A	A	A	(42)
Glimepiride	A	N (failed)	N	A	A	(42)
Haloperidol	A	A	A	N	N	(42)
Lansoprazole	A	A	A	N	N	(42)
Mevastatin	A	N	A	A	A	(42)
Montelukast	A	A	N	A	A	(42)
Omeprazole	A	A	A	A	A	(42)
Phenylbutazone	A	A	A	A	A	(42)
Pioglitazone	A	A	A	A	A	(42)
Rabeprazole	A	A	A	A	N	(42)
Reserpine	A	N	A	N	N	(42)
Rifabutin	A	N	A	A	A	(42)
Rifapentine	A	N (failed)	A	N	N	(42)
CDD3508	A	A	A	A	A	(8)
CDD3532	A	A	A	A	A	(8)
CDD3538	A	A	A	A	A	(8)
CDD3543	A	A	A	A	A	(8)
CDD3501	A	A	A	A	A	(8)
CDD3530	A	A	A	A	A	(8)
CDD3536	A	A	A	A	A	(8)
CDD3540	A	A	A	A	A	(8)
5β-Androstan-3α-ol	A	A	N	A	A	(8)
Petromyzonol	N	N	A	A	A	(8)
17β-Dihydroandrosterone	A	N	A	A	A	(8)
Dihydrotestosterone	A	N	A	A	A	(8)
Etiocholanolone	A	A	A	A	A	(8)
Taurochenodeoxycholic acid	N	A	A	N	N	(8)
Lithocholic acid	A	N	A	A	A	(8)
7-Ketolithocholic acid	A	A	N	N	N	(8)
12-Ketolithocholic acid	A	N	N	N	N	(8)
Estrone	A	A	A	A	A	(8)
Estriol	A	N	N	A	A	(8)
Allocholic acid	N	N	A	N	N	(8)
Cortolone	A	A	A	A	A	(8)
Estetrol	A	N	N	A	A	(8)
Epitestosterone sulfate	A	N	N	A	A	(8)
5β-Pregnan-3α,20α-diol	A	N	A	A	A	(8)
5α-Androstan-3α-ol	A	N	A	A	A	(8)
Lithocholicacid acetate	A	N (failed)	A	A	A	(8)
5β-Cholan-3α,7α,12α,24-tetrol	N	N	A	A	A	(8)
ω-Muricholicacid	A	N	A	A	A	(8)
16,(5α)-Androsten-3β-ol	A	N	A	A	A	(8)
Tauro-β-muricholic acid	N	N	A	N	N	(8)
C2BA-10	A	A	A	N	N	(66)
C2BA-11	N	A	A	A	A	(66)
C2BA-12	N	A	A	A	A	(66)
C2BA-13	A	A	N	A	N	(66)
C2BA-248	A	A	N	A	N	(66)
C2BA-251	A	A	N	A	A	(66)
C2BA-3	N	A	A	A	N	(66)
C2BA-5	A	A	A	A	A	(66)
C2BA-6	A	A	A	N	A	(66)
C2BA-7	A	A	A	A	A	(66)
C2BA-8	A	A	A	A	A	(66)
C2BA-9	N	A	A	N	N	(66)
Budesonide	A	A	A	A	A	(67)
Carbamazepine	A	A	N	N	A	(68)
Chlorpromazine	N	A	N	N	N	(69)
Cyclophosphamide	A	N	A	A	A	(70)
Efavirenz	A	A	N	N	N	(71)
Etoposide	A	N	N	A	A	(71)
Fluconazole	N	A	N	N	N	(72)
Gemfibrozil	N	A	N	N	N	(73)
Ketoconazole	N	N	N	A	A	(71)
Nicardipine	A	A	A	N	N	(74)
Rosiglitazone	A	A	A	N	A	(75)
Topotecan	A	N	N	N	A	(71)
Valproic acid	A	N	N	A	A	(76)
Bupropion	A	A	N	N	N	(42)
Diclofenac	A	A	N	N	A	(42)
Flutamide	A	A	N	N	N	(42)
Isotretinoin	A	A	N	A	A	(42)
Loratadine	A	A	A	N	N	(42)
Meclizine	A	A	A	N	N	(42)
Meloxicam	A	A	A	A	A	(42)
Midazolam	A	A	A	N	N	(42)
Nabumetone	A	A	A	A	A	(42)
Naloxone	A	A	N	N	N	(42)
Ondansetron	A	A	N	N	N	(42)
Quinapril	A	A	A	N	N	(42)
Raloxifene	A	A	A	N	A	(42)
Sildenafil	A	A	N	N	N	(42)
Simvastatin	A	N	A	A	A	(42)
Terbinafine	A	A	N	N	N	(42)
Triazolam	A	A	N	N	N	(42)
Zolpidem	A	A	A	A	A	(42)
Acetaminophen	N	A	A	A	A	(42) and Inactive this study
Acyclovir	N	A	N	N	N	(42)
Albuterol	N	A	N	N	N	(42)
Amitriptyline	N	A	N	N	N	(42)
Atenolol	N	A	N	N	N	(42)
Azithromycin	N	N (failed)	N	N	A	(42)
Captopril	N	A	N	N	N	(42)
Cimetidine	N	A	A	A	A	(42)
Ciprofloxacin	N	A	A	N	N	(42)
Clarithromycin	N	N (failed)	A	A	N	(42)
Dapsone	N	A	A	A	N	(42)
Desipramine	N	A	N	N	N	(42)
Doxazosin	N	N	N	N	N	(42)
Doxepin	N	A	N	N	N	(42)
Doxorubicin	N	A	N	N	N	(42)
Enalapril	N	N	A	A	A	(42)
Erythromycin	N	N (failed)	A	A	A	(42)
Ethosuximide	N	A	N	N	A	(42)
Etodolac	N	A	N	N	A	(42)
Famotidine	N	A	N	N	N	(42)
Fexofenadine	N	N (failed)	A	N	N	(42)
Furosemide	N	A	N	N	N	(42)
Ibuprofen	N	A	N	A	A	(42)
Lamotrigine	N	A	A	N	N	(42)
Levofloxacin	N	A	N	N	N	(42)
Linezolid	N	A	N	A	A	(42)
Lisinopril	N	N (failed)	N	N	N	(42)
Metoprolol	N	A	N	N	N	(42)
Metronidazole	N	A	N	N	N	(42)
Morphine	N	A	N	N	N	(42)
Nadolol	N	A	N	N	A	(42)
Naproxen	N	A	N	A	A	(42)
Nevirapine	N	A	A	A	A	(42)
Paroxetine	N	A	N	N	N	(42)
Propranolol	N	A	N	N	N	(42)
Ranitidine	N	N	N	N	N	(42)
Risperidone	N	A	A	N	N	(42)
Sumatriptan	N	A	N	N	N	(42)
Tacrine	N	A	A	A	A	(42)
Terazosin	N	A	A	A	A	(42)
Theophylline	N	A	A	A	N	(42)
Timolol	N	A	N	N	N	(42)
Tocainide	N	A	N	N	N	(42)
Tolmetin	N	A	N	N	N	(42)
Valsartan	N	A	N	N	A	(42)
Venlafaxine	N	N	N	N	N	(42)
Zomitriptan	N	A	N	N	N	(42)
Cholecalciferol	N	N (failed)	A	A	A	Inactive this study
Flurbiprofen	A	A	N	N	N	80.1 μM this study
Mycophenolic acid	N	N	N	A	A	Inactive this study
Oxcarbazepine	A	A	N	A	A	18.1 μM this study
Oxycodone	N	A	N	N	N	Inactive this study

Where, Prob and A represents probability and activator, respectively. The value of the FlexX score is negative and the associated coefficient is also negative (−0.0398) suggesting that the log odds (and, therefore, the probability) of a compound being a PXR activator decrease with the increase in FlexX score (positive value). In other words, for a one unit increase in FlexX score, the odds in favor of an activator are estimated to be decreased by a multiplicative factor of 0.961 (exp-0.0398 = 0.961). Figure 3 shows several representative molecules from the test set that were correctly predicted as PXR activators by all methods in this study. Apart from the SE which was higher for the test set than the machine learning methods, the performance of RF, SVM and RP appear to be better than the docking logistic regression model, based on the prediction accuracies, SP, Q and C of the large set (Table 1). The overall poor prediction accuracy of the docking results may be due to: a) poor performance of the scoring function in characterizing ligand-receptor interactions; b) treating the protein as rigid rather than flexible; and c) the overall promiscuous nature of the PXR LBD. It should also be noted that 9 out of 145 (6.2%) test set molecules failed to dock and were classed as non-activators, whereas predictions were made with the other machine learning models for all test set molecules.

Figure 3.

The FlexX docked (1NRL) conformations of selected correctly predicted PXR activators - omeprazole (A), phenylbutazone (B), pioglitazone (C), and felodipine (D). Omeprazole, pioglitazone and felodipine are predicted to form hydrogen bonding interactions with the side chain NH of Gln-285. The hydrogen bonding interaction is shown as red dotted lines. The amino acid residues are shown in the stick mode whereas the PXR activators are shown in the ball and stick mode (purple). The 3D images were created using Chimera.(77)

Discussion

VolSurf descriptors have been widely used for drug discovery and ADME/Tox (37-39, 63). In the course of this study we have developed novel predictive global models using VolSurf descriptors with human PXR activation classification data to identify compounds that are PXR activators. Such PXR activators may in turn lead to undesirable effects on drug and endogenous compound metabolism and transport. The machine learning models we have evaluated in the course of this work can also be used for virtual screening to identify molecules previously unrecognized as PXR activators that can then be selected for in vitro testing. We have used a similar approach previously with transporters (64, 65). In this study we have also used a large test set of 145 molecules that have not been included in model building to validate all of the computational approaches. Our classification results were a significant improvement on those published previously (20). It is important to note that the new test set was also structurally diverse yet covered the same descriptor space as the training set as analyzed using PCA analysis, which reassures us that we were unlikely to be extrapolating far beyond the training set (61). Although the results of the large test set are not as good as those achieved for internal validation of the training set, a common weakness of many computational approaches, this could be partially due to the promiscuous nature of PXR. We would add that in contrast to the previous study by Ung et al (20) who used a test set of 15 PXR activators, we not only were able to predict more of these correctly, but we also further evaluated the model with a much larger test set and obtained predictions which we felt were acceptable considering the difficulty in predicting activation with this protein. The probability of randomly selecting a PXR activator or a PXR non-activator is 56.55 % and 43.44 %, respectively. Our SVM results are therefore better than random for the large test set as we achieved approximately 68% and 65% correct for PXR activator or PXR non-activator, respectively (Supplemental Table 2). These machine learning methods may be improved by using additional descriptor types which we are currently evaluating for inclusion in future studies.

We have also utilized a structure-based docking approach using FlexX combined with logistic regression for comparison purposes which performs poorly compared with SVM, RP and RF. This is perhaps not surprising considering the flexibility of PXR and the differences in the ligand bound structures reported (9, 10, 53, 54). In this case the final FlexX logistic regression model was based on 1NRL which had a resolution of 2.00Å (9). To our knowledge there have been no direct comparisons of docking, QSAR or machine learning methods for PXR with a large external test set containing diverse xenobiotics. Although a recent study has used both rigid and flexible docking to destabilize the key agonist-protein interactions for PXR in a small number of molecules, no binding scores or classification was reported (35). Assessment of other docking approaches in the future with a large test set like that used in the current analysis, is justified in an attempt to improve the results and at least match the standard set by the machine learning models described herein. With several PXR crystal structures in the PDB (9, 10, 53, 54), this study has highlighted the difficulty in using docking to predict a large set of molecules as potential agonists and suggests that classification models developed with machine learning methods (35) may have an important role alongside pharmacophores (8, 14) and smaller “local” quantitative QSAR models in prioritizing compounds for in vitro testing. The combination of the training and test sets into a single larger model (322 molecules) would be worth assessing in the future with other descriptors and machine learning or docking methods and may lead to the creation of custom PXR scoring functions.

Although the docking approach was not ideal for classification it could still be useful in those cases were the predictions were correct to enable a qualitative evaluation of the likely ligand-protein interactions. This could assist in defining what molecular changes can be introduced chemically to reduce or avoid PXR activation. For example, others have suggested attaching hydrogen bonding groups on one of the hydrophobic features, adding larger more rigid groups as well as removing central H-bond acceptors (35). The large test set created in this study (Table 2) can be used to find structurally similar compounds that are either observed human PXR activators or PXR non-activators (regardless of their predicted activity), and provides further instruction as to what types of structural modification can decrease activity. While the test set contains numerous β-adrenergic blockers they are all PXR non-activators, and molecules in the calcium channel blocker and proton pump inhibitor series are all PXR activators. We were able to find some examples that had differing activity and visualize those that were correctly predicted as PXR activators based on the docking orientation. For example oxycodone and morphine are PXR non-activators while naloxone is a PXR activator (Figure 4). While there are several subtle differences between the three compounds, naloxone contains an additional ethylene moiety which may interact with a hydrophobic region in the LBD in addition to being a reactive functional group. Naproxen (PXR non-activator) and nabumetone (PXR activator) are non steroidal anti-inflammatory drugs, the former possesses a carboxylic acid group while the latter has a longer alkyl chain with a methyl substituted for the hydroxyl which may serve to preferentially position the C=O hydrogen bond acceptor (Figure 4). Nevirapine is a PXR non-activator while oxcarbazepine and carbamazepine are PXR activators. The latter two molecules feature carboxamide moieties which will serve as potential hydrogen bond acceptors while in the corresponding position nevirapine has a hydrophobic group (Figure 4). A further series of tricyclic compounds includes the PXR non-activators desipramine, doxepin and amitriptyline which contain a basic primary amine group at the end of an aliphatic chain, while the PXR activator loratadine has a neutral hydrogen bond acceptor in the corresponding position as well as ring substitutions. Interestingly the training set also contains additional tricyclic molecules (protriptyline, nortriptyline, trimipramine and clomipramine) that are PXR non-activators, yet structurally they closely resemble desipramine, doxepin and amitriptyline. These qualitative examples demonstrate some of the benefits of a large database of such binary classification information as it is valuable in biasing design of future molecules away from PXR activators that may be accomplished with the assistance of computational and in vitro methods.

Figure 4.

Examples of PXR non-activator and PXR activator compounds in the test set that are structurally similar. Red dashed circles highlight likely important structural features that may contribute to activity. Corresponding PXR activator compounds (that had correctly predicted classifications) were docked in the PXR structure (1NRL) with FlexX (as described in the Materials and Methods). Carbamazepine (circled group), loratadine (circled group) and nabumetone (C=O group next to circled group) are predicted to form hydrogen bonding interactions with the side chain NH of Gln-285, additionally the methyl group in nabumetone may interact with the Phe288. Naloxone is predicted to form a H-bond with the side chain C=O and His407 and the ethylene group may also have a hydrophobic interaction with the Phe288. Oxcarbazepine docks differently to Carbamazepine and may also form a H-bond between the side chain C=O and His407. The hydrogen bonding interactions are shown as a red dotted line. The amino acid residues are shown in the stick mode whereas the PXR activators are shown in the ball and stick mode (purple). The 3D images were created using Chimera.(77)

In conclusion, we have developed new machine learning models for PXR that outperform those previously developed as well as docking models when tested with a new large external test set. We suggest it is likely quite difficult to definitively describe functional groups that will globally decrease affinity due to the size and flexibility of the human PXR LBD, rather, as in this and previous cases (35) it will be important to address structural series individually using all of the available computational and in vitro tools at our disposal. These approaches should also enable increased efficiency of screening and testing, aiding pharmaceutical research to avoid developing potent PXR agonists.

Abbreviations

ADME/Tox

Absorption, distribution, metabolism excretion/ Toxicology

LDB

Ligand binding domain

PCA

Principal Component analysis

PXR

Pregnane X Receptor

PDB

Protein data bank

QSAR

Quantitative Structure Activity Relationship

RF

Random Forest

RP

Recursive Partitioning

SVM

Support Vector Machine.

Supplementary Material

1

Supplemental Table 1. Confusion matrix for the training and small test set of the PXR dataset (20). The values in the parenthesis are for the small test set (20). The numbers in italics (docking and logistic regression) includes failed molecules.

Supplemental Table 2. PXR confusion matrix analysis of the large 145 molecule test set. The numbers in italics (docking and logistic regression) includes failed molecules.