B2 NCBI Library, Building 38A

11am, January 8, 2009

 

Sumitra Nair

British Columbia Cancer Research Center, Vancouver, Canada

 

Determination of telomere length using an automated learning model

 

The flow-FISH is a technique for measuring the telomere length of blood cell populations and it makes use of multi-parameter
 measurements in flow cytometry (FCM) and fluorescent in-situ hybridization (FISH). The flow-FISH analysis involves manual 
gating of cell population. As the flow-FISH data is multidimensional, the manual analysis is a time consuming process. 
The aim of this work was to automate the flow-FISH analysis technique. The proposed automated learning model makes use of 
the Kohonen self organizing map (KSOM) for identifying the blood cell populations of interest, namely, control cells
 (a group with known telomere length), lymphocytes and granulocytes. The main bottle neck in the analysis was the noise 
involved in the FCM files. The model is equipped with filtering techniques for noise removal, which had been developed on
 the basis of the properties of the FCM attributes of the data. Our study consisted of the data of 140 subjects. Corresponding 
to each subject there existed two pairs of files, of which one file in a pair is with and the other is without peptide nucleic
 acid probe. The telomere length was determined using each pair and their mean was calculated. However for some cases 
the automated model used only one pair of files and rejected the other pair. The algorithm determined the lymphocytes
 telomere length of 129 subjects (92%) and flagged the remaining ones as problematic. In the case of granulocytes, 
the manual gating telomere length results of 132 subjects were known and of which the automated model calculated 
the telomere length of 120 subjects (92%) and flagged the remaining 12 ones as potentially incorrect. Of the calculated 
telomere lengths, the relative error percentage between the automated and the manual calculation was less than 5% for 98% 
of subjects in the case of lymphocytes and 94% of subjects in the case of granulocytes. Thus the proposed learning model
 showed a clinical utility for the analysis of flow-FISH data.