Prediction of membrane permeability of molecules using Machine Learning

Understanding the permeability of small drug-like molecules across a lipid membrane is fundamental in pharmaceutical applications. Physics-based models have found that the main physio-chemical properties determining the permeability are bulk partitioning free energy ∆G and pKa, but they are expensive to compute experimentally or computationally. Hence, some research exists in the area of using solely molecular structure information to predict permeability. In this work, by using different combinations of the main physio-chemical properties and molecular structure information, we predict the membrane permeability of small molecules using three machine learning (ML) algorithms, the Lasso model, the Multi-layer Perceptron (MLP) model, and the combined model of the Lasso and MLP. Molecular descriptors and fingerprints, obtained from the SMILES string of a molecule, are used with/without the key features (∆G and pKa) as input to the ML models. The Lasso model shows R2 values of 0.80 and 0.72 with and without the key features, respectively. With the key features, the prediction of the MLP model is very accurate (R2=0.99), while it decreases without the key features (R2=0.89). The combined Lasso-MLP model shows a slightly better prediction without the key features (R2=0.90). Our results validate the possibility of predicting the permeability well simply based on data extracted from the SMILES string, without the main physio-chemical properties. You can find the project technical report here.