Experiments on automatic drug activity characterization using support vector classification

The characterization of pharmacological properties from their chemical structure has become a challenging and promising technique in computer aided drug design. The idea consists of finding appropriate representations of candidate compounds in terms of their chemical formulae and try to apply a particular machine learning method able to appropriately characterize certain desired properties or kinds of pharmacological activity. In this particular work antibacterial activity has been considered. Several classic pattern classification methods have already been applied to this problem with promising results. In this work, the support vector machine model is considered and compared to multilayer perceptrons in this particular context. The natural and unpredictable imbalance and the fact that only relatively small samples can be used for learning make this a challenging and interesting problem.