Advances in Relevant Descriptor Selection

Background

During the last few decades the number of available molecular descriptors became substantially larger. In most situations this number is practically infinite. Without previous detailed knowledge about the phenomenon at hand, it is difficult to conduct the unsupervised off-line descriptor selection that is theoretically justified. Modern QSAR studies, therefore, frequently include thousands of descriptors generated for a few dozens of molecules. From the statistical point of view, the estimation of QSAR model parameters is highly unreliable or impossible in the described settings. Even the application of the machine learning approaches to the model development leads to serious problems. So-called “curse of dimensionality” (Guyon, 2003), caused by a large number of descriptors and their possible interactions, makes model development intractable. Even if obtained models pass the external validation criteria, their interpretation is blurred by numerous descriptors of unknown relevance. This property makes the rational molecular design or data base mining a very complex task.