Methods to Improve the Reliability, Validity and Interpretability of Qsar Models

Quantitative structure activity relationship (QSAR) models are are a statisti-cal solution to the problem of directly calculating physical and biological properties ofmolecules from their physical structure. The direct prediction of properties is in generalnot feasible either owing to lack of computing resources or lack of knowledge about therelationship between structure and property. The goal of a QSAR model is to extract in-formation from a set of numerical descriptors characterizing molecular structure and usethis information to develop inductively a relationship between structure and property.Two important questions arise during the modeling process. First, are the data used tobuild the model representative of the whole dataset and can the model be extended topredict properties for new molecules? Second, given that a model encodes informationabout the structures of molecules and relates this to their properties, can we extractand interpret the encoded information? The focus of the work reported in this thesisis on the validation and interpretation of QSAR models and presents both applicationsof interpretation techniques as well as the development of validation and interpretationmethodologies.The first study describes a technique to develop representative QSAR sets using aself-organizing map (SOM). The SOM was used to classify a dataset consisting of dihy-drofolate reductase inhibitors with the help of an external set of global descriptors. Theresultant classification was used to generate training, cross-validation and prediction sets(collectively known as QSAR sets) for QSAR modeling using the ADAPT methodology.The results were compared to those of QSAR models generated using sets created byactivity binning and a sphere exclusion method. The results indicated that the SOM wasable to generate QSAR sets that were representative of the composition of the overalldataset in terms of similarity. The resulting QSAR models were half the size of thosepublished and had comparable RMS errors. Furthermore, the RMS errors of the QSARsets were consistent, indicating good predictive capabilities as well as generalizability.The determination of the validity of a QSAR model when applied to new com-pounds is an important concern in the field of QSAR modeling. Various scoring tech-niques can be applied to specific types of models to obtain measures of confidence in thepredicted property for new compounds. The second study describes the development ofa methodology which allows one to state whether a new compound will be well predicted