Figure 1

Graphical overview of Q-mode Factor Analysis (FA) [9]. Each sample is described by a vector x _i , containing the expression levels for all genes in the chip. The complete expression for all samples is contained in the matrix X = {x _i }. The expression levels of each sample are assumed to be generated by a linear combination of a small number of underlying transcriptional programs, the latent (non-observable) variables, contained in the set of vectors {F _i }, forming matrix F. The relative contribution of each program is given by the thickness of the arrows connecting factors and samples, stored in variables l_ij, altogether forming the loading matrix L. Each l_ij element can be understood as the correlation coefficient between the expression levels of the sample and the corresponding latent variable. Residuals are kept in vectors {ε _i }, giving rise to matrix E. Note that small loadings connecting a given sample (i.e., X ₄ with the factor model implies large residuals.