Figure 8

Scheme of the different algorithms in the SPACE method. The figure depicts the different proposed stages that are performed after the background removal and normalization across the arrays. The initial input to the algorithm is the probe expression matrix for a particular gene (in this case the WNK1 gene). The selected experiment is a set of arrays hybridized against different tissues provided by Affymetrix. Probe signal matrix may have some outliers that are identified and corrected in stage A (outlier removal) of the method. In stage B (NMF factorization), the corrected matrix is factorized into two matrices W (related to concentration) and H (related with the structure). NMF factorization is not unique. We select the factorization whose $\tilde{W}$ matrix is closer to the AG factorization suggested by Wang et al. This adaptation is performed in stage C. It can be seen in the zoomed figures that the $\tilde{W}$ matrix have columns that are closer to each other. Once we have the $\tilde{W}$ matrix it can be factorized again into two matrices: a matrix of affinities (step D.1) and a continuous $\tilde{W}$ = $A\tilde{G}$ property matrix $\tilde{G}$. Finally, the discretization process (step E) provides a binary structure of the G matrix that is compatible with the characteristics of the exon arrays, i.e., the probes that belong to the same exon probeset have the same hybridization pattern.