On the (lack of) robustness of gene expression data clustering

We assess the robustness of partitional clustering algorithms applied to gene expression data. A number of clusterings are made with identical parameter settings and input data using SOM and k-means algorithms, which both rely on random initialisation and may produce different clusterings with different seeds. We define a reproducibility index and use it to assess the algorithms. The index is based on the number of pairs of genes consistently clustered together in different clusterings. The effect of noise applied to the original data is also studied. Our results show a lack of robustness for both classes of algorithms, with slightly higher reproducibility for SOM than for k-means