In the present study a systematic approach was used to model the anti-T. cruzi activity of a series of N-oxide containing heterocycles belonging to four chemical families with a wide structural diversity. The proposed mode of action implies the reduction of the N-oxide moiety; however, the biochemical mechanism underlying the anti-T. cruzi activity is still unkown. For structural representation two types of descriptors were analyzed: quantum chemical (AM1) global descriptors and properties coded by radial distribution function (RDF). Both types of descriptors point to the relevance of electronic properties. The local-RDF (LRDF) identified an electrophilic center at 4.1-4.9 A from the oxygen atom of the N-oxide moiety, although other properties are required to explain the biological activity. While the mode of action of N-oxide containing heterocycles is still unknown, the results obtained here strengthen the importance of the electrophilic character of the molecule and the possible participation of the heterocycle in a reduction process. The ability of these descriptors to distinguish among activity classes was assessed using Kohonen neural networks, and the best clustering descriptors were later used for model building. Different learning algorithms were used for model development, and stratified 10-fold cross-validation was used to evaluate the performance of each classifier. The best results were obtained using k-nearest neighbors (k-NN) and decision tree (J48) methods combined with global descriptors. Since tree-based methods are easily translated into classification rules, the J48 model is a useful tool in the de novo construction of new N-oxide containing heterocycle lead structures.