In Silico Prediction of PAMPA Effective Permeability Using a Two-QSAR Approach

Abstract

Oral administration is the preferred and predominant route of choice for medication. As such, drug absorption is one of critical drug metabolism and pharmacokinetics (DM/PK) parameters that should be taken into consideration in the process of drug discovery and development. The cell-free in vitro parallel artificial membrane permeability assay (PAMPA) has been adopted as the primary screening to assess the passive diffusion of compounds in the practical applications. A classical quantitative structure-activity relationship (QSAR) model and a machine learning (ML)-based QSAR model were derived using the partial least square (PLS) scheme and hierarchical support vector regression (HSVR) scheme to elucidate the underlying passive diffusion mechanism and to predict the PAMPA effective permeability, respectively, in this study. It was observed that HSVR executed better than PLS as manifested by the predictions of the samples in the training set, test set, and outlier set as well as various statistical assessments. When applied to the mock test, which was designated to mimic real challenges, HSVR also showed better predictive performance. PLS, conversely, cannot cover some mechanistically interpretable relationships between descriptors and permeability. Accordingly, the synergy of predictive HSVR and interpretable PLS models can be greatly useful in facilitating drug discovery and development by predicting passive diffusion.

Keywords: effective permeability coefficient (Pe); hierarchical support vector regression; in silico; parallel artificial membrane permeability assay (PAMPA); partial least square; two-QSAR.

Figure 1

Administration routes and the corresponding ratios for those unique drugs approved by the FDA in 2018.

Figure 2

Molecule distribution for the molecules selected for this study in the training set (solid square), test set (open circle), and outlier set (grey triangle) in the chemical space spanned by three principal components.

Figure 3

Observed log of the effective permeability coefficient (P_e) vs. the log P_e predicted by SVR A (solid circle), SVR B (solid diamond), hierarchical support vector regression (HSVR; green square), and partial least square (PLS; red triangle) for the molecules in the training set. The green and red solid lines, dashed lines, and dotted lines correspond to the HSVR and PLS regressions of the data, 95% confidence intervals for the HSVR and PLS regressions, and 95% confidence intervals for the prediction, respectively.

Figure 4

Observed log P_e vs. the log P_e predicted by SVR A (solid circle), SVR B (solid diamond), HSVR (green square), and PLS (red triangle) for the molecules in the test set. The green and red solid lines, dashed lines, and dotted lines correspond to the HSVR and PLS regressions of the data, 95% confidence intervals for the HSVR and PLS regressions, and 95% confidence intervals for the prediction, respectively.

Figure 5

Observed log P_e vs. the log P_e predicted by SVR A (solid circle), SVR B (solid diamond), HSVR (green square), and PLS (red triangle) for the molecules in the outlier set. The green and red solid lines, dashed lines, and dotted lines correspond to the HSVR and PLS regressions of the data, 95% confidence intervals for the HSVR and PLS regressions, and 95% confidence intervals for the prediction, respectively.

Figure 6

Residual vs. the log P_e predicted by HSVR (green) and PLS (red) in the training set (square), test set (circle), and outlier set (triangle).

Figure 7

The observed log of the intrinsic permeability coefficient (P_o) values vs. the observed log P_e values.

Figure 8

The observed log P_o values vs. the log P_e values predicted by HSVR (green square) and PLS (red triangle), and their regression lines.

Figure 9

Relationship among log P_e, fractional polar surface area (FPSA), and log P in 3D presentation.

Figure 10

Box plot of log P_e values for different ion classes, where the boxes represent the distribution of log P_e from the 25th to the 75th percentile, the black and red lines depict the median and mean values, the whiskers denote the minimum and maximum values, and the asterisk indicates significant difference between neutral and the others (p < 0.05).