This web page was produced as an assignment for a course on
Statistical Analysis on Microarray Data at
Discussion
This project has focused on differentiating different types of tumors and normal tissue based on microarrays of microRNA. The major difficulties in this project arose from the poor quality of the microarrays. Although the published paper had a stringent filter, I was more interested in working with lots of data. Out of 768 miRNA on the array, only 87 miRNA passed the filtering criteria set by the researchers. I used a LOESS normalization procedure on the data.
2 of the 34 arrays had been given to the researchers misdiagnosed. The researchers noticed that these two arrays consistent classified differently, and upon a new biopsy the tumors were found to be different than the original diagnosis. This spurred my interest in differentiating between ARMS and ERMS (ARMS was misdiagnosed as ERMS) and between PRMS and GIST (PRMS was misdiagnosed as GIST). I used LIMMA to find miRNA that distinguished these tumors.
Using SAM I looked for miRNA that differentiated between normal and tumor tissues and again that differentiated between any of the tissues. Using the top 20 significant miRNA from the normal versus tumor tissues I constructed dendrograms and decided to use the 1-abs(cor) metric and a complete linkage method. With that metric and method I looked at the 100 most significant miRNA and compared the hierarchical clustering to clustering of PAM (Partitioning Around Mediods). The hierarchical 1-abs(cor) complete linkage method resulted in the same clusters as PAM.
Using classification of Prediction Analysis of Microarrays (PAMr) we see that the miRNA do a pretty good job of classifying the tissue types correctly. It is good that several of the miRNA that are very important in classifying the arrays are the same miRNA that showed up as being the most significant ones in SAM to differentiate between the arrays.
This website was designed by Austen Head.
Email: austen
[dot] head [at]