Probes were variance-filtered, collapsed into per-gene transcripts by selecting the highest variance probe for each gene and filtered for univariate survival association using a 1,000,000-round permutation test based on the FAST statistic (Gorst-Rasmussen & Scheike, 2012).
For selecting the most informative probes we applied the following procedure: 1) we log2 transformed the normalized absolute intensity fluorescence value of each probe; 2) selected, for the FFPE and FF samples separately, the 20% highest variance probes, and 3) took the union of the two probe groups (FF and FFPE).
After univariate exploration of the data, variables with significant differences (p<0.05) between low- and high-variance probes were investigated further in multivariate analysis using a logistic regression model.
For univariate analysis, we compared the distribution of predictor variables between low- and high-variance probes using Student's t-test for continuous variables and a chi-square test for categorical variables.
Known variables with microarrays include high variance in intensity between probe spots, high variance in a single probe spot's intensity between chips, as well as background (non-specific binding) intensity differences between chips.
Because genes that discriminate between molecular subtypes are by definition differentially expressed, and thus show a relatively high variance in expression, we selected the probe set with highest variance per gene (n=17 583 probe sets (6.2% of all 'core' probe sets); log2 normalised data).
Examining the distribution of variance revealed that the majority of probes have low variance (<0.1, n = 378,057), with only ∼2% of probes showing high variance (>0.1 and <4.97, n = 7,417).
Examining the distribution of variance revealed that the majority of probes have low variance (<0.1, n = 378,057), with only a fraction of probes displaying high variance (>0.1 and <4.97, n = 7417).
However, a selection on the most informative probe sets (based on highest variance for each probe set/gene and highest variance between genes) is required.
To limit the study to genes under high variance and to limit the number of probes used in calculating the metagene fit, probes were considered for metagene analysis based on the interquartile range (IQR) of the probe being in the upper 80th percentile.
