Robust differential expression analysis by learning discriminant boundary in multi-dimensional space of statistical attributes

Table 2 Compares the results of different DC configurations in a typical simulation test (6 vs. 6; G ⁵⁰⁰₅₀₀ )

Baseline	Target FDR	Add s^T	Add s^R	Add s^NB	Use s^T _, s^R, and s^NB
Use s^T alone	0.01	-	+4.62% (1.84e-08)	+2.84% (8.74e-06)	+5.52% (2.18e-12)
Use s^T alone	0.05	-	+2.32% (2.04e-14)	+3.71% (1.06e-23)	+5.31% (3.18e-35)
Use s^NB alone	0.01	+24.48% (1.98e-41)	+16.35% (8.87e-30)	-	+27.72% (8.76e-44)
Use s^NB alone	0.05	+14.71% (1.23e-51)	+9.62% (2.32e-35)	-	+16.48% (8.96e-56)

The 1^st column lists the single-attribute DC configurations. DC^R was not displayed because it failed to detect any true DEFs under both FDR targets. The 2^nd column lists the target FDR levels (0.01 or 0.05) at which performances are compared. Cells in the 3^rd~6^th columns show the improvements in percentage of multi-attribute DC configurations (indicated by the column headers) over single-attribute DC configurations (indicated by the 1^st cells in the corresponding rows). The numbers in parentheses are the paired t-test p-values showing the significance of the improvement. For example, the cell at the 3^rd column and 4^th row shows that DC^T+NB outperformed DC^NB by 24.48% with a paired t-test p-value of 1.98e-41 at FDR < 0.01. Although DC^R as a single attribute failed detect any DEFs, adding s^R to the other two attributes (s^T and s^NB) significantly improved the performance, as indicated by the 4^th column

ISSN: 1471-2105