Characterization of lung tumor subtypes through

gene expression cluster validity assessment

Supplementary Information

Giorgio Valentini and Francesca Ruffino

DSI - Dipartimento di Scienze dell'Informazione
Universita' degli Studi di Milano
Via Comelico 39, Milano, Italia
e-mail: {valenti,ruffino}@dsi.unimi.it

This web page contains supplementary information related to the paper "Characterization of lung tumor subtypes through gene expression cluster validity assessment". This work performs an experimental analysis of the validity of clusters discovered in lung tumor patients  using stability measures based on random projections.
From this web page  the full experimental data and the R source code used for the experiments are available.
In particular are downloadable the following supplemntary information and R source code:

Clusterings obtained with different clustering algorithms

The clusterings obtained with different clustering algorithms are reported, considering from 2 to 20 clusters. In the hierarchical clustering file the examples belonging to each cluster are labeled with a conventional number used in our experiments (that corresponds to the ordering of the examples in the original Bhattacharjee et al. data set) together with the  label used in the Bhattacharjee et al. original experiments.
In the other two clustering files (PAM and c-mean), the same number labels are used to identify the examples of each cluster, but the original labels of Bhattacharjee et al. are substituted with labels corresponding to classes histologically defined.
More precisely:
  • NL corresponds to  normal lung sample (n=17 specimens)
  • AD corresponds to  lung adenocarcinoma (n=139)
  • SCLC corresponds to small-cell lung carcinoma (n=6)
  • SQ corresponds to squamous cell lung carcinoma (n=21)
  • COID corresponds to pulmonary carcinoids (n=20)
The text files with the clusterings obtained with different clustering algorithms are the following:
  1. Clusterings obtained with the Ward's hierarchical algorithm
  2. Clusterings obtained with the Prediction Around Medoids (PAM) algorithm
  3. Clusterings obtained with the c-mean algorithm

Cluster validity results with different clustering algorithms

In the following text files the overall stability index S and the individual stability index s  of the of the clusterings and clusters described in the previous section are reported. We considered from 2 to 20 clusters for epsilon values in the set {0.1, 0.2, 0.3, 0.4, 0.5}. Each epsilon value corresponds to different dimensions of the projected subspace, according to the Johnson Lindenstrauss lemma. We used in all cases Plus-Minus-One (PMO) random projections. For more details about random projections see the clusterv web-site.

  1. Hierarchical clustering validity results
  2. Prediction Around Medoids (PAM) clustering validity results
  3. C-mean clustering validity results

R source code used to perform the cluster validity analysis

  1. R code to perform cluster validity analysis (hierarchical clustering)
  2. R code to perform cluster validity analysis (PAM clustering)
  3. R code to perform cluster validity analysis (c-mean clustering)
Note that you need the Biobase R package  (downloadable from the Bioconductor web site) and the clusterv package (downloadable for the clusterv web site) as well as the Bhatta.Lung.filtered.norm binary file to run the above R scripts.
Bhatta.lung.filtered.norm  stores the homonymous object of the class exprSet that represents the gene expression data.