Using Alternative Multi-variate Data Analysis Methods for Predicting Physical-chemical Properties of Organic Compounds

In drug design, the knowledge of physical properties of organic compounds is vital. The special field of ADMET (Administration, Distribution, Metabolism, Excretion, Toxicity) involves the determination of properties governing the effect of a drug on the human body, such as aqueous solubility, intestinal absorption and permeability. Assessing these properties experimentally, however, is often laborious and time-consuming. This master’s thesis investigates the applicability of three neural network methods as computer aided means for predicting physical properties from molecular structure. The examined methods include fuzzy ARTMAP, Adaptive Network-based Fuzzy Inference System (ANFIS) and Support Vector Machines (SVM). Furthermore, the possibility of constructing a so-called consensus model, in which the results of the above mentioned models are weighted together in order to receive improved predictions, is investigated. The results from this study show that the SVM technique produced the most accurate and stable predictions on a majority of the data sets. Both fuzzy ARTMAP and ANFIS experienced difficulties in generalising to unseen data, as a consequence of over-training. This was in particular valid for the smaller data sets (< 30 compounds). Nevertheless, some satisfactory results were obtained. Fuzzy ARTMAP made several excellent test set predictions, even though the cross-validation results were genuinely poor for the small data sets. ANFIS turned out to be less appropriate for biological applications, since the number of adjustable network parameters grows exponentially with the number of molecular descriptors, owing to the network’s complex architecture. For practical use, the number of descriptors was restricted to at most ten. Finally, utilising a simple consensus model was found to be an excellent method for improving the predictions. ANVÄNDNING AV ALTERNATIVA DATAANALYSMETODER FÖR PREDIKTION AV FYSIKALISKA OCH KEMISKA EGENSKAPER HOS ORGANISKA FÖRENINGAR