Predictive human intestinal absorption QSAR models using Bayesian regularized neural networks

An oral dosage form is generally the most popular with patients. Many drug candidates fail in late development because of unfavourable absorption and pharmacokinetic profiles, or toxicity, among other factors (ADMET properties). This contributes to the fall in the efficiency of the pharmaceutical industry and to the rise in health costs. The ability to predict ADMET properties of drug leads can contribute to overcoming this problem. We have modelled intestinal absorption using several types of molecular descriptors and a non-linear Bayesian regularized neural network. Our models show very good predictive properties and are able to account for essentially all of the variance in the data that is not due to experimental error.