Semi-supervised class discovery using quantitative phenotypes – CVD as a case study
© Steinfeld et al; licensee BioMed Central Ltd. 2007
Published: 20 November 2007
Genomic studies typically focus on comparing disease to healthy population. In our work, various parameters, including peripheral blood mononuclear (PBM) cells expression profiling, were stratified solely from healthy subjects. To analyze the data we developed a semi-supervised class discovery method, constraining the search space to patterns that respect an order induced by the rich quantitative annotations. We show that our method is robust enough to detect known clinical parameters with accordance to expected values. We also use our method to elucidate cardiovascular disease (CVD) putative risk factors.
One of the basic tasks in gene expression data analysis is finding differentially expressed genes between 2 classes (such as tumor vs. normal). Among the various methods for measuring differential expression (e.g. Student t-test), we focus on TnoM  which is a non-parametric statistical score that affords an exact p-value.
When many partitions of the sample set are possible, one would like to assess the statistical significance of any partition considered, and to compare between partitions. In overabundance  analysis the exact p-value of the TNoM score is used to estimate the expected number of differentially expressed genes. By comparing to the actually observed number we can calculate the overabundance of differentially expressed genes. This quantity can be used as a figure of merit: higher overabundance indicating a more profound change in the cell state.
We applied our method to PBMC gene expression profiling data, collected from 49 healthy subjects. Clinical, laboratory measurement and CVD prognostic indicators were also collected, adding more then 160 phenotypic parameters for each subject. One of the interesting phenotypic parameters is Carotid Intima-Media Thickness (IMT) , a CVD prognostic indicator. Using semi-supervised class discovery with the IMT values we received IMT threshold levels that are in agreement with the known prognosis values (Figure 1). The differentially expressed genes in this partition were enriched with GO terms related to vesicle-mediated transport (p < 10-8) and glycolysis (p < 10-6), giving mechanistic insights to the difference between the two cell states.
This study is supported by a EU FP6 grant in the framework of the Multi Knowledge Project. We thank Eran Eden, Doron Lipson and Benny Chor for useful discussions.
- Ben-Dor A, Bruhn L, Friedman N, Nachman I, Schummer M, Yakhini Z: Tissue classification with gene expression profiles. J Comput Biol 2000, 7(3–4):559–83. 10.1089/106652700750050943View ArticlePubMedGoogle Scholar
- Ben-Dor A, Friedman N, Yakhini Z: Overabundance Analysis and Class Discovery in Gene Expression Data. Agilent Technical Report 2002. AGL-2002–4 AGL-2002–4Google Scholar
- von Heydebreck A, Huber W, Poustka A, Vingron M: Identifying splits with clear separation: A new class discovery method for gene expression data. Bioinformatics 2001, 17: S107-S114. 10.1093/bioinformatics/17.1.107View ArticlePubMedGoogle Scholar
- Lorenz MW, Markus HS, Bots ML, Rosvall M, Sitzer M: Prediction of clinical cardiovascular events with carotid intima-media thickness: a systematic review and meta-analysis. Circulation 115(4):459–67. 2007, Jan 30; Epub 2007 Jan 22 2007, Jan 30; Epub 2007 Jan 22 10.1161/CIRCULATIONAHA.106.628875View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd.