Fuzzy subgroup mining for gene associations

When studying the therapeutic efficacy of potential new drugs, it would be much more efficient to use predictors in order to assess their toxicity before going into clinical trials. One promising line of research has focused on the discovery of sets of candidate gene profiles to be used as toxicity indicators in future drug development. In particular genomic microarrays may be used to analyze the causality relationship between the administration of the drugs and the so-called gene expression, a parameter typically used by biologists to measure its influence at gene level. This kind of experiments involves a high throughput analysis of noisy and particularly unreliable data, which makes the application of many data mining techniques very difficult. In this paper we explore a fuzzy formulation of the a priori algorithm, a technique whose crisp version is commonly used to mine for subgroups in large datasets; the purpose is to extend the original method, already suitable to deal with large amount of data, in a way that naturally allows the user to deal with the intrinsic imprecision in the data. The algorithm is tested on real data coming from experimental genomic data.