Bulletin of Experimental Biology and Medicine, Vol. 154, No. 4, February, 2013
521
METHODS Computer Evaluation of Drug Interactions with P-Glycoprotein A. A. Lagunin, T. A. Gloriozova, A. V. Dmitriev, N. E. Volgina*, and V. V. Poroikov
Translated from Byulleten’ Eksperimental’noi Biologii i Meditsiny, Vol. 154, No. 10, pp. 520-524, October, 2012 Original article submitted May 13, 2011 The (Q)SAR models for evaluating the structure–property relationships, fit for prediction of drug interactions with P-glycoprotein as inhibitors or substrates, were constructed using PASS and GUSAR software. The models were constructed and validated on the basis of information on the structure and characteristics of 256 and 94 compounds used as P-glycoprotein substrates and inhibitors, respectively. The initial samples were divided 80:20 into training and test samples. The best prediction accuracy for the test samples was 78% for P-glycoprotein substrate prediction (PASS) and 89% for inhibitor prediction (GUSAR). Key Words: (Q)SAR; P-glycoprotein; PASS; GUSAR The blood–brain barrier (BBB) functions as a filter through which nutrients from circulating blood get into the brain and the nervous tissue vital activity products are eliminated in the opposite direction. BBB protects the nervous tissue from microorganisms, toxins, cellular and humoral immunity factors, circulating in the blood, and prevent the penetration of xenobiotics, including drugs, into the brain. Xenobiotic elimination from the brain into circulating blood is realized by active transport with participation of P-glycoprotein (Pgp) [9]. The absence of Pgp in BBB (in genetically modified mice) or blockade of its functions by injection of specific inhibitors (chemical knockout) results in a 10-50-fold increase in the levels of some drugs in the brain (the majority of these drugs are highly lipophilic) [9]. The drugs modulating the CNS have V. N. Orekhovich Institute of Biomedical Chemistry, the Russian Academy of Medical Sciences; *N. I. Pirogov Russian Medical University, the Ministry of Health and Social Development of the Russian Federation, Moscow, Russia. Address for correspondence: alexey.
[email protected]. A. A. Lagunin
to easily penetrate into this system by passive diffusion and at the same time they should not be Pgp substrates or inhibitors [6]. By contrast, drugs of peripheral action should poorly penetrate into the CNS and/or be effectively eliminated back into the blood with participation of Pgp. Hence, evaluation of drug interactions with Pgp as substrates or inhibitors is an important step in creation of new drugs. That is why the attention of scientists has been recently focused on the construction of computer models, (Q)SAR models of the structure–property relationships for prediction of drug interactions with Pgp [2-5,12,13]. An advantage of computer methods is the possibility of their application to virtual, heretofore not synthesized drug structures, which makes it possible to select the most promising of them at the early stage of studies. The accuracy of the proposed methods [2-5,12,13] for the test samples varies from 0.51 to 0.88 for prediction of Pgp substrates and up to 0.81 for Pgp inhibitors. In order to make highly accurate prognosis of the substrate specificity, 3D pharmacophores [12] are often used, which limits the application of this approach to other (than
0007-4888/13/15440521 © 2013 Springer Science+Business Media New York
Bulletin of Experimental Biology and Medicine, Vol. 154, No. 4, February, 2013 METHODS
522
the training sample) classes of organic compounds and involves significant computation. We studied the possibility of constructing (Q)SAR models of the structure–property relationships fit for prediction of drug interactions with Pgp as inhibitors and substrates using PASS [1,7,10] and GUSAR [8,11] software we created. The efficiency of software has been demonstrated not once [1,8], which gives grounds to construct stable predictive (Q)SAR models on this base.
MATERIALS AND METHODS Samples of substances. Pgp substrates and inhibitors (256 and 94 compounds, respectively). were used for construction and validation of (Q)SAR models. Information on these substances is presented in ChEMBLdb [https://www.ebi.ac.uk/chemldb/]. The predictive capacity of (Q)SAR models is usually evaluated in an independent test sample. Because of rather small scope of available information on the known Pgp substrates and inhibitors, the initial samples (before creation of models) were divided into training and test samples (in 80:20 ratio). In order to make the test sample representative by a variety of compound structure and by even distribution of these substances, we classified them as follows: compounds of the initial sample were ranked by decrease of activity; each fifth compound was taken into the test sample, while the remaining compounds constituted the training sample. The accuracy of quantitative models was evaluated by the following criteria: R2model is a multiple coefficient of determination calculated when creating the QSAR model by the formula: R2=1-Σ(yobs-ycalc)2/Σ(yobs-ymean)2, where yobs is the observed dependent variable, ycalc estimated dependent variable, and ymean the mean dependent variable. The predictive capacity of the model was calculated by the sliding control method with exclusion by one: Q2=1-Σ(yobs-ypred)2/Σ(yobs-ymean)2, where ypred was the predicted value of excluded compound. R2test is evaluation of prediction accuracy for the test sample: R2test=1-Σ(yobs-ypred)2/Σ(yobs-ymean)2, where yobs was the observed variable in the test sample, ypred predicted variable in the test sample, and ymean the mean variable in the test sample. RMSEtest was the root mean square error for test sample:
RMSE=
Σ(y
obs
-ypred)2, n
where n was the number of objects. The qualitative (classification) structure–activity models were constructed using the PASS (Prediction of Activity Spectra for Substances) system [1,10]. The biological activity was described qualitatively (active/ inert), the prediction algorithm making use of the “naïve” Bayes approach. The organic compound structure was described using fragmented atomic centered MNA descriptors (Multilevel Neighborhoods of Atoms) [7]. The resultant prognosis was presented as an orderly list of names of the respective activities and probabilities Pa (be active) and Pi (be inert); the compound was assumed to be active if Pa>Pi. Quantitative structure–activity models were constructed using GUSAR computer system (General Unrestricted Structure Activity Relationships), based on the use of QNA (Quantitative Neighborhoods of Atoms) descriptors and the results of prediction of biological activity spectrum by the PASS software as descriptions of molecules and on the use of autocoordinated regression method [8,11]. By varying the parameters of the descriptors used, the program constructed numerous QSAR models. The best models were selected with consideration for estimated accuracy values, and the consensus model, averaging the predicted results of the selected models, was constructed on this base. The applicability sphere, that is, evaluation of the similarity of the test compound and the compounds used for model creation, was estimated for the test compounds. If the test compound did not fall in the applicability sphere, the results of prognosis were unreliable.
RESULTS Construction of models for prediction of Pgp substrates. Two samples of compounds, tested for Pgp substrate specificity, were formed. The training sample consisted of 210 compounds, the test sample of 53 ones. The value describing EfluxRatio BA/AB was the dependent variable and was calculated from permeability coefficients (Papp): Papp(b→a) EfluxRatio BA/AB= , Papp(a→b) where Papp=dC/dt Vr/AC0, b→a is substance migration from the basolateral to the apical side; a→b is substance migration from the apical to the basolateral side; C0 was the initial concentration, dC/dt migration through the monolayer, Vr compartment volume, and A area of membrane surface. EfluxRatio BA/AB>1.5 indicated that the substance was a Pgp substrate.
A. A. Lagunin, T. A. Gloriozova, et al.
The structure–characteristic relationship for the training sample was analyzed using GUSAR software. Based on the 10 best QSAR models, the consensus model for predicting EfluxRatio BA/AB for Pgp substrates was constructed with permissible (by precision) parameters: R2=0.696, Q2=0.635. The EfluxRatio BA/ AB for the test sample compounds was predicted using the consensus model. The results of prognosis for test sample were used to calculate the RMSEtest and R2test (coefficient of determination):
523 on the base of the training sample and included 7 best QSAR models. The consensus model had acceptable accuracy parameters: R2=0.647, Q2=0.554. The consensus model was used to predict the Ki for the test sample. The RMSEtest and R2test were calculated from the results of prognosis for the test sample: RMSEtest(Log10(Ki(nM)=0.334, R2test=0.60.
Only one of the 53 test sample compounds did not fall into the applicability area of our consensus model. The accuracy of quantitative prediction of EfluxRatio BA/AB values was not high (Fig. 1). A model describing interactions with Pgp at the qualitative level (Pgp substrate, Pgp nonsubstrate) was created using PASS software on the base of the described training sample. The model accuracy (IAP – invariant analysis of accuracy) [1], calculated by the sliding control method with exclusion by one, was 77%. The accuracy of PASS qualitative prognosis for the test sample compounds was evaluated in comparison with the results of GUSAR prognosis (Table 1). Qualitative prognosis of Pgp substrates made by PASS software had certain advantages in comparison with the GUSAR prognosis, offering better balanced specificity and sensitivity characteristics and higher accuracy (Table 1). Construction of models for prediction of Pgp inhibition. Two samples of compounds were tested for Pgp inhibition. The training sample consisted of 76 compounds, the test samples of 18. The inhibition constant Ki, presented in Log10 (nM), was the dependent variable. The consensus model for Pgp inhibitors Ki prediction was constructed using GUSAR software
All test sample compounds fell in the field of applicability of the resultant consensus model. The accuracy of Ki values was acceptable (Fig. 2). As with Pgp substrates we obtained models of better quality with the use of PASS software, we used it for analysis of the training sample. A model describing interactions with Pgp at a qualitative level was constructed (Pgp inhibitor – Pgp noninhibitor). The model accuracy (IAP), calculated by the sliding control method with exclusion by one, was 57.7%. Prediction for test sample compounds was made on the base of this model (Table 2). The sensitivity and accuracy of this prognosis were inferior to those of the GUSAR prognosis. Hence, we carried out a simulation study of the relationships between organic compounds and Pgp transporter in order to evaluate the BBB permeability for drugs. By the accuracy of prognosis for substrates (0.78) the PASS software was comparable with the best known methods [2,5,13] used for a heterogenic test sample, but was worse than the 3D pharmacophore method [12], the accuracy of which in the test was 0.88. The prognosis for inhibitors was more accurately made by the GUSAR software (0.89) – higher than in a previous [4] study (0.81). The results indicate a relationship between the substance structure and activity towards Pgp. The accuracy of prognosis suggests these models for prediction of drug interactions with Pgp in drug creation.
Fig. 1. Predicted (ordinate) and experimental (abscissa) EfluxRatio BA/AB for test sample compounds.
Fig. 2. Predicted (ordinate) and experimental (abscissa) Ki values for test sample compounds.
RMSEtest(Log10(EfluxRatio BA/AB=0.49; R2test=0.38.
Bulletin of Experimental Biology and Medicine, Vol. 154, No. 4, February, 2013 METHODS
524
TABLE 1. Qualitative Prognosis of Pgp Substrates for Test Sample at EfluxRatio BA/AB>1.5 Software
TP
TN
FP
FN
SN
SP
AC
PASS
20
21
5
7
0.74
0.81
0.78
GUSAR
23
14
12
4
0.85
0.54
0.70
Note. TP: True Positive (number of Pgp substrates predicted as substrates); TN: True Negative (number of Pgp nonsubstrates predicted as nonsubstrates); FP: False Positive (number of Pgp nonsubstrates predicted as substrates); FN: False Negative (number of Pgp substrates predicted as nonsubstrates); SN: sensitivity of prognosis – TP/(TP+FN); SP: specificity of prediction – TN/(TN+FP); AC: accuracy of prediction – (TP+TN)/(TP+TN+FP+FN).
TABLE 2. Evaluation of Prognosis Quality for Test Sample Compounds (Compounds, Reacting with Pgp with Ki (nM)
