Chemistry Central Journal Poster presentation Validation of predicitve modelling techniques in drug design – influence of test set composition

© 2009 Matz et al; licensee BioMed Central Ltd. In chemoinformatics and in the analysis of Quantitative Structure-Activity Relationships (QSAR) experimental data of a molecular property of interest are routinely mathematically related to a set of carefully chosen structure descriptors which represent the molecules under study. Many different mathematical techniques can be used for this purpose. For the sake of simplicity, we focus here on the simplest technique to predict continuous properties, which is linear regression. In a typical model building process the data analyst needs to make several decisions with respect to model complexity and model parameters. Mostly, these decisions are data-driven which makes some form of internal validation necessary to obtain information how the model complexity or mode