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

sample cleansing process and these encompass 4,248 ProbeSets. Borrowing from machine
learning practices, a gain ratio was calculated from k-means clustering analysis of individual
ProbeSets to perform down selection to the most significant features [3, 4]. This gain ratio was
also calculated at the individual probe level and then averaged across the probes belonging to an
accepted set, to give a ProbeSet response. We demonstrate that both approaches improve the
model’s performance in a supervised classification analysis, implemented using either the kNN
classifier based on Euclidean space [3, 5, 6], or Fisher’s linear discriminant analysis (LDA) [7, 8].
In this case the emphasis is not on comparing analysis methods, but rather focuses on discovery
of intriguing biological phenomenon revealed by using the BaFL pipeline to select the most
unambiguous signal. Eighteen significant ProbeSets were selected, possessing a gain criterion
greater than 0.8, and this set of genes suggest that an important mechanism underlying
tumorogenesis is abnormal cytokinesis. The biological significance of these genes are validated
by a literature survey in the discussion section.
Materials and Methods
A subset of samples were selected from the BaFL cleansed Bhattacharjee dataset to construct a
multiclass dataset containing NSCLC tumor biopsies and adjacent/normal biopsies. This dataset
contained the 125 adenocarcinomas and 13 normals, which comprised the two state disease model
previously considered, and an additional 17 squamous samples. The BaFL cleansing pipeline
was applied across all samples, with the result that 24,022 probes were found to be common to
the three states; these lie in 4,248 ProbeSets (with at least 4 acceptable probes each).
115