Logistic Classification Models for pH–Permeability Profile: Predicting Permeability Classes for the Biopharmaceutical Classification System

• Published in: Journal of chemical information and modeling 2019-05, Vol.59 (5), p.2442-2455
• Main Authors: Oja, Mare, Sild, Sulev, Maran, Uko
• Format: Article
• Language: English
• Subjects: Absorption; Accuracy; Classification; Decision analysis; Decision theory; Decision trees; Gastrointestinal system; Hydrophobicity; Ionization; Model accuracy; Octanol; Permeability; Predictions; Regression analysis; Variations
• ISSN: 1549-9596; 1549-960X
• DOI: 10.1021/acs.jcim.8b00833

Permeability is used to describe and evaluate the absorption of drug substances in the human gastrointestinal tract (GIT). Permeability is largely dependent on fluctuating pH that causes the ionization of drug substances and also influences regional absorption in the GIT. Therefore, classification models that characterize permeability at wide ranges of pH were derived in the current study. For this, drug substances were described with six data series that were measured with a parallel artificial membrane permeability assay (PAMPA), including a permeability profile at four pH values (3, 5, 7.4, and 9), and the highest and intrinsic membrane permeability. Logistic regression classification models were developed and compared by using two distinct sets of descriptors: (1) a hydrophobicity descriptor, the logarithm of the octanol–water partition (logPow) or distribution (logD) coefficient and (2) theoretical molecular descriptors. In both cases, models have good classification and descriptive capabilities for the training set (accuracy: 0.76–0.91). Triple validation with three sets of drug substances shows good prediction capability for all models: validation set (accuracy: 0.73–0.91), external validation set (accuracy: 0.72–0.9), and the permeability classes of FDA reference drugs for the biopharmaceutical classification system (BCS) (accuracy: 0.72–0.88). The identification of BCS permeability classes was further improved with decision trees that consolidated predictions from models with each descriptor type. These decision trees have higher confidence and accuracy (0.91 for theoretical molecular descriptors and 0.81 for hydrophobicity descriptors) than the individual models in assigning drug substances into BCS permeability classes. A detailed analysis of classification models and related decision trees suggests that they are suitable for predicting classes of permeability for passively transported drug substances, including specifically within the BCS framework. All developed models are available at the QsarDB repository (http://dx.doi.org/10.15152/QDB.206).