Discovering Significant Structures in Clustered Bio-molecular Data Through the Bernstein Inequality

Searching for structures in complex bio-molecular data is a central issue in several branches of bioinformatics. In particular, the reliability of clusters discovered by a given clustering algorithm have been recently assessed through methods based on the concept of stability with respect to random perturbations of the data. In this context, a major problem is to assess the confidence of the measures of reliability. We discuss a partially ”distribution independent” method based on the classical Bernstein inequality to assess the statistical significance of the discovered clusterings. Experimental results with gene expression data show the effectiveness of the proposed approach.