02.08.2013 Views

Sample A: Cover Page of Thesis, Project, or Dissertation Proposal

Sample A: Cover Page of Thesis, Project, or Dissertation Proposal

Sample A: Cover Page of Thesis, Project, or Dissertation Proposal

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

decrease perf<strong>or</strong>mance in the weighted AUC metric. This list <strong>of</strong> candidate genes demonstrate a<br />

decrease true positive classification rate f<strong>or</strong> the squamous samples, with co<strong>or</strong>dinating increases in<br />

false positive rates f<strong>or</strong> the adenocarcinoma and n<strong>or</strong>mal samples (data not shown).<br />

Although ProbeSets selected by their average probe gain criteria demonstrate enhanced<br />

perf<strong>or</strong>mance f<strong>or</strong> the kNN alg<strong>or</strong>ithm and similar perf<strong>or</strong>mance f<strong>or</strong> the LDA alg<strong>or</strong>ithm, it likely not<br />

a suitable approach. We observed little overlap (6) between the 29 ProbeSets selected by their<br />

average probe gain (>=0.6) and the 18 ProbeSets selected by their ProbeSe5t gain criteria<br />

(>=0.8). This translates to only 1/5 th <strong>of</strong> th ProbeSets having an enhanced inf<strong>or</strong>mation gain f<strong>or</strong> the<br />

aggregated ProbeSet, and in fact 2/3 rd <strong>of</strong> them demonstrate diminished inf<strong>or</strong>mation gain f<strong>or</strong> the<br />

aggregated ProbeSet. Theref<strong>or</strong>e, we believe that the enhanced classification perf<strong>or</strong>mance f<strong>or</strong> the<br />

average probe inf<strong>or</strong>mation gain selected lists is likely the result <strong>of</strong> the ‘good’ classifiers being<br />

m<strong>or</strong>e significant in the feature subspace.<br />

However the average probe gain suggests probes to investigate f<strong>or</strong> transcript measurement<br />

anomalies. F<strong>or</strong> example, the same list <strong>of</strong> 9 ProbeSets included 6 ProbeSets containing distinct<br />

probes demonstrating po<strong>or</strong> clustering perf<strong>or</strong>mance, and these have been show to skew the gain<br />

distributions and presumably the aggregate’s clustering perf<strong>or</strong>mance. F<strong>or</strong> example, the KRT17<br />

ProbeSet has 2 probes (index 6 and 10) with low probe perf<strong>or</strong>mances, 0.033 and 0.06 gain ratios<br />

respectively. Selection by the average probe gain per ProbeSet appears to have a high false<br />

positive rate; furtherm<strong>or</strong>e these ProbeSets appear to have limited connection to the ProbeSets<br />

selected by the ProbeSet gain criterion. As a counterexample, ProbeSet 33108_i_at, SOX2,<br />

includes five probes all <strong>of</strong> whose gain values range from 0.73 to 0.99, yet the ProbeSet aggregate<br />

gain ratio value is a dismal 0.18. This ProbeSet is up-regulated in the squamous samples, by the<br />

124

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!