Comparative Analysis of Machine Learning Techniques for the Prediction of the DMPK

Several machine learning techniques were evaluated for the prediction of parameters relevant in pharmacology and drug discovery including rat and human microsomal intrinsic clearance as well as plasma protein binding represented as the fraction of unbound compound. The algorithms assessed in this study include artificial neural networks (ANN), support vector machines (SVM) with the extension for regression, kappa nearest neighbor (KNN), and Kohonen Networks. The data sets, obtained through literature data mining, were described through a series of scalar, two- and three-dimensional descriptors including 2-D and 3-D autocorrelation, and radial distribution function. The feature sets were optimized for each data set individually for each machine learning technique using sequential forward feature selection. The data sets range from 400 to 600 compounds with experimentally determined values. Intrinsic clearance (CLint) is a measure of metabolism by cytochrome P-450 enzymes primarily in the vesicles of the smooth endoplasmic reticulum. These important enzymes contribute to the metabolism of an estimated 75 % of the most frequently prescribed drugs in the U.S. The fraction of unbound compound (fu) greatly influences pharmacokinetics, efficacy, and toxicology. In this study, machine learning models were constructed by systematically optimizing feature sets and algorithmic parameters to calculate these parameters of interest with cross validated correlation/RMSD values reaching 9.53 over the normalized data set. These fully in silico models are useful in guiding early stages of drug discovery, such as analogue prioritization prior to synthesis and biological testing while reducing costs associated with the in vitro determination of these parameters. These models are made freely available for academic use.