High-Resolution Segmentation of ENCODE Data by Self-Organising Maps (SOM)The training of the self-organising map (panel A) and analysis of the results (panels B and C) are shown. Initially we arbitrarily placed genomic segments from the chromHMM segmentation on to the toroidal map surface, although the SOM does not use the chromHMM state assignments (panel A). We then trained the map using the signal of the 12 different ChIP-seq and DNase-seq assays in the six cell types analysed. Each unit of the SOM is represented here by an hexagonal cell in a planar two-dimensional view of the toroidal map. Curved arrows indicate that traversing the edges of two dimensional view leads back to the opposite edge. The resulting map can be overlaid with any class of ENCODE or other data to view the distribution of that data within this high-resolution segmentation. In panel A the distributions of genome bases across the untrained and trained map (left and right, respectively) are shown using heatmap colours for log10 values. Panel B shows the distribution of TSSs from CAGE experiments of GENCODE annotation on the planar representations of either the initial random organisation (left) or the final trained SOM (right) using heat maps coloured according to the accompanying scales. The bottom half of panel B expands the different distributions in the SOM for all expressed TSSs (left) or TSSs specifically expressed in two example cell lines, H1 hESC (centre) and HepG2 (right). Panel C shows the association of Gene Ontology (GO) terms on the same representation of the same trained SOM. We assigned genes that are within 20 kb of a genomic segment in a SOM unit to that unit, and then associated this set of genes with GO terms using a hypergeometric distribution after correcting for multiple testing. Map units that are significantly associated to GO terms are now coloured green, with increasing strength of colour reflecting increasing numbers of genes significantly associated with the GO terms for either immune response (left) or sequence-specific TF activity (centre). In each case, specific SOM units show association with these terms. The right-hand panel shows the distribution on the same SOM of all significantly associated GO terms, now colouring by GO term count per SOM unit. For sequence-specific TF activity, two example genomic regions are extracted at the bottom of panel C from neighbouring SOM units. These are regions around the DBX1 (from SOM unit 26,31, left panel) and IRX6 (SOM unit 27,30, right panel) genes, respectively, along with their H3K27me3 ChIP-seq signal for each of the Tier 1 and 2 cell types. For DBX1, representative of a set of primarily neuronal TFs associated with unit 26,31, there is a repressive H3K27me3 signal in both H1 hESC and HUVEC cells; for IRX6, representative of a set of body patterning TFs associated with SOM unit 27,30, the repressive mark is restricted largely to the embryonic stem cell.
