Generation of QSAR Sets Using a Self-Organizing Map

As mentioned in Chapter 2, self-organizing maps (SOM) are a class of unsupervised neural networks whose characteristic feature is their ability to map nonlinear relations in multi-dimensional datasets into easily visualizable two-dimensional grids of neurons. SOM’s are also referred to as self-organized topological feature maps since the basic function of a SOM is to display the topology of a dataset, that is, the relationships between members of the set. SOM’s were first developed by Kohonen in the 1980’s, and since then they have been used as pattern recognition and classification tools in various fields including robotics, astronomy, and chemistry. Neural networks, in general, have been used extensively in chemistry and chemometrics and examples of applications in chemistry include spectroscopy, prediction of NMR properties and prediction of reaction products SOM’s have also been applied to studies in the field of QSAR/QSPR. The fundamental premise of QSAR studies is that structurally related (similar) compounds will have similar properties. Determining similarity is a complex task, and many methods exist such as principal components analysis and hierarchical cluster analysis. The fact that a SOM is able to extract topological information from a dataset makes it a valuable tool for detecting similarities in a dataset. Thus, it is to be expected that neighboring neurons in a two-dimensional SOM grid will be similar to each other. If each neuron in such a SOM grid can be assigned a molecule, groups of similar molecules can be identified. Many studies have used a SOM to perform the actual QSAR analysis by detecting relationships between structures and activities of interest. Other applications use SOM’s at different stages of the QSAR study, for example, the use of a SOM to choose the best subset of molecular descriptors to perform a QSAR analysis. However, another important step in QSAR study is the generation of training, cross-validation,