Signature Selection for Grouped Features with a Case Study on Exon Microarrays - Feature Selection for Data and Pattern Recognition - page 335

Fig. 14.4 Probabilistic outcomes of logistic regression classifiers for a single random test set. The
outcomes of LASSO, GL and SGL are compared to the true labels

14.3.3.3 Selected Features

Another important factor of feature selection in this case study is how the three
methods (LASSO, GL, and SGL) provide insights on our research question about
alternative splicing.
Table 14.1 summarizes the features selected by the three methods in a single
random subsampling trial. Here the PID represents a 7-digit unique number assigned
by Affymetrix, to a set of probes (also known as probesets) on microarrays to detect
exons. As we consider core probesets with unique hybridization, each probeset has
a correspondence to a unique exon, a coding subsequence of a gene on DNA. A
PID also maps to a feature in our discussion. The “Coef” refers to the value in the
coefficient vector corresponding to each chosen feature: the larger its magnitude

is, the more contribution is made by a feature to prediction.
Although some genes (DDR2 and HIST1H1A) were detected in all of the three
feature sets, the results were quite different. First, the exons chosen by LASSO were
all different except for the gene EPB41L3. As the genetic relations of exons are
essentially ignored and any exon with high correlation to labels can be selected by
LASSO, it is hard to say if an exon was chosen by a possible change in a single
exon or by a new combination of exons (alternative splicing). In fact, the coefficient
magnitude values were small except for a single exon from the gene IPW. Since this
gene is non-functional (non-protein coding), it is likely that correlations in measure-
ment errors have been captured rather than the true signals in this case. On the other
hand, several exons were chosen from the same genes (DDR2 and HIST1H1A) with
relatively large coefficient values in case of GL and SGL (a single exon was chosen
from HIST1H1A in case of SGL). They could be indicatives of possible alternative

ʲ