Tumor classification by Bayesian classifier with Markov model using
                     microarray gene expression data

                               Jan Dupac

Tissue classification using gene expression data is an important
problem in genetics. Closely related problem is the selection of
informative genes.  Both problems are addressed in this research.  The
Bayesian classifier with Markov model for tumor classification using
microarray gene expression (GEP) data is introduced.

We used two colon cancer GEP data sets. The main data set containing
205 samples comes from Osaka University. The second data set
containing only 62 samples is from Princeton University and served for
comparison with the other researchers.  The classification of tissue
to normal versus cancer using GEP has almost no practical effect. The
more challenging problem is to distinguish patients with and without
metastasis according to primary cancer samples.

The classifier using Markov model without feature extraction was
compared with the classifier using the model of Normal distribution
with full covariance matrix and Principal Component analysis for
feature extraction.  It was found that feature extraction is not
necessary for classification if appropriate statistical model is used
but that the gene selection has important effect to classification
results. The selection methods using both gene ranking and
classification feedback was developed and tested.  However, the
feedback from classification error on training set had minimal effect
on classifier performance.

The diagnosis of metastasis to liver against no metastasis based on
GEP is possible, the classification error is about 5\%.  Non-liver
metastasis is not so well separated. However, the classification no
metastasis versus both liver and non-liver metastasis is also possible
with classification error about 15\%.