Figure 2

Gene expression correlation space obtained by HiT-MDS-2. Surrogate gene space of developing barley endosperm tissue. A total number of 4824 temporally regulated, log-normalized gene profiles with 14 time points are embedded using powers of p = 8 of Pearson correlation for profile comparison. The bipolar sandglass shape of points is labeled by 64 centroids obtained from correlation-based neural gas (NG) clustering which yields a good density-related covering of the data space. The clusters fall into five biologically meaningful regulatory different patterns: each of the five small panels displays the average and standard deviation of aligned genes belonging to the marked centroids – the number of genes within each category 1–5 are given in parentheses. The HiT-MDS-2 embedding shows a consistent spatial arrangement of clusters within the manually selected categories (apart from clusters 8, 13, and 51); pre-storage phase (cluster group 1), intermediate phase (cluster groups 2 and 5, anti-correlated to each other), and storage phase (cluster groups 3 and 4). Since the exponent p = 8 magnifies small differences, the resulting large spatial variability for down- and up-regulation represented by cluster groups 3 and 5 do still correspond to only small standard deviations. The small asymmetric bias of the gene cloud to the right indicates more subtle patterns found in up-regulation rather than in down-regulation, which underlines the benefit of the visualization for screening and validation.