ML based QSAR Models for Prediction of Pharmacological Permeability of Caco-2 Cell

In the initial stages of de novo drug discovery, numerous drug components need to be considered, in order to determine those candidates which bind to a particular disease protease. The greater the binding effect the better the drug efficacy. However, mapping every potentially relevant drug and its effect on the protein is a time consuming task. To discard the drugs at the initial stage we can know how much permeable a drug is through a particular layer or cell membrane. A potential approach to determine this by measuring the permeability of a compound through a specific layer. In this paper, an approach for QSAR regression for predicting pharmacological permeability of the Caco-2 cell is proposed. The compounds are represented by chemical descriptors calculated from their construction properties and structural properties sets of descriptors were derived from the chemical compounds structures. Linear regression, nonlinear regression and nonlinear artificial neural network models were experimented with to correlate their reported permeability value. Two different sets of chemical descriptors were derived and each set was used for training different machine learning and neural network models. The results were evaluated using standard metrics like mean square error and R-squared error, during which it was observed that boosting based ML models achieved the lowest values when compared to other regression models.