Prediction of permeability of drug-like compounds across polydimethylsiloxane membranes by machine learning methods

Abstract

The prediction of maximum steady state flux values of a chemical compound from its structural features plays an important role in design of transdermal drug delivery systems. In this study, we developed the quantitative structure property relationship (QSPR) models to estimate the maximum steady state flux of 245 drugs-like compounds through the polydimethylsiloxane membranes. A correlation-based feature selection was used for descriptor selection. The selected descriptors, surface tension, polarity, and count of hydrogen accept sites, which are interpretable and can be used to explain the permeability of chemicals. These descriptors are used for developing the QSPR prediction models by multiple linear regression, artificial neural network, support vector machine (SVM) and Instance-Based Learning algorithms using K nearest neighbor machine learning approaches. The models were assessed by internal and external validation. All four approaches yield the QSPR models with good statistics. The models developed by SVM have better prediction performance. These models can be useful for predicting the permeability new untested compounds.