All tests done were based on comparing the differences in normalized, mean log-transformed intensities between classes of samples.
Unlike methods that attempt to select individual genes by their significant differential expression between classes of samples (e.g., SAM [ 12]), the gene set is selected as a set of markers that discriminate when used collectively but not necessarily when each gene is used individually.
Orthogonal partial least square discriminant analysis (OPLS-DA), a supervised multivariate regression technique, was performed to identify components that best differentiate between predefined classes of samples while dissecting orthogonal components which do not differentiate between these classes.
Unlike PCA, OPLS-DA can identify sets of metabolites that best distinguish between predefined classes of samples.
This analysis allows the identification of defined sets of genes with differential expression between two classes of samples [ 31, 32].
Paired miRNA and mRNA expression profiling provides an excellent platform for capturing those miRNA expression changes between two classes of samples that lead, positively or negatively, to the changes in mRNA expressions between the two classes of samples.
The difference is that, while GSEA gives only an overall enrichment score of a pathway between two classes of samples, ASSESS provides an enrichment score of the pathway for each sample.
After the validation, a final list of significant genes differentially expressed between three classes of samples (Her2+, ER + and TNBC) was determined by using one way ANOVA on the full set of samples, except for the one Her2+/ER + sample (TB148).
So, applying FA to microarray data also means that to select differentially expressed genes between different classes of samples, the classes do not have to be defined a-priori.
To reduce the effect of such genes on the data normalization step, Robinson and Oshlack reported a simple but effective global scaling method, the trimmed mean of M values (TMM) method, where a scaling factor for the normalization is calculated as a weighted trimmed mean of the log ratios between two classes of samples (i.e., Samples A vs. B) [ 16].
If there are systematic differences in RNA quality between two classes of samples being compared, we predict that the effect of RNA quality on relative estimates of gene expression levels would be responsible for much of the signal in the data.
