A Shown is numbers of the differentially correlated genes discovered using a meta-analysis procedure (Methods) between AN and TU tissues. For comparison the same analysis was run both in the real data and in a permutation of the AN and TU assignments. The top panel shows the number of differential gene pairs (Y-axis) for the real data (AN vs TU, in blue) in comparison to the permutation (red) as a function of the p value (shown as negative log10[P], on the X-axis). The number of differentially connected genes (middle panel) and false discovery rates (bottom panel) are also shown. B To establish that the differentially correlated gene pairs resulted from the difference between AN and TU tissue and not for example differences between individuals the same analysis was run using AN vs AN compared to AN vs TU using the same number of samples (Methods). The top panel shows the number of differential gene pairs (Y-axis) for both AN vs TU (blue) and AN vs AN (red) as a function of the p value (shown as negative log10[P], on the X-axis). The number of differentially connected genes (middle panel) and false discovery rates (bottom panel) are also shown.

To illustrate both loss and gain of correlation two cell cycle genes were chosen that have many differential connections: the cyclin dependent kinase inhibitor CDKN2C (A, 149 loss and 11 gains of connectivity) and the chromatin licensing and DNA replication factor CDT1 (B, 0 loss and 38 gains of connectivity). The expression intensity values for CDNK2C (X-axis) and DCN (Y-axis) are shown (A) in AN (blue, cc 0.548, p = 2e-20) and TU tissues (red, cc -0.358, p = 7.14e-10). The intensity values for CDT1 (X-axis) and MCM3 (Y-axis) are shown (B) in AN (blue, cc-0.153, p = 0.0166) and TU tissues (red, cc 0.673, p = 3.05e-38). Shown in C is the degree distribution for the 8,736 differentially connected genes as described in the text. The numbers of differential connections for each gene (log10, X-axis) was compared to the count (log10, Y-axis). As shown the distribution was scale-free, indicating that the identified genes tend to preferentially attach to a small number of hub genes in either tumor or adjacent normal tissues, but not in both.

Distribution of correlations (X-axes) between genes (Y-axes) and the closest sCNV marker (in cis) in TU (A) and AN (B) tissues. The distribution of the real data (blue solid lines) was compared to permutation of the gene to marker connection (green dashed lines). No significant associations were seen in AN above what would be expected by chance. In TU there was a pronounced bias to positive correlations. C. Distribution of correlations between all genes (Y-axis) and all sCNV markers, shown linearly by chromosome location through the genome (X-axis). Chromosome boundaries are indicated by the vertical green lines and are numbered. Using a correlation cut-off of >0.3 (p<4.7e-4, FDR <0.02), the count of all genes (not including cis genes, green), positively correlated trans genes (red), or negatively correlated trans genes (blue) is indicated for each marker (trans here was defined as genes and sCNV markers falling on separate chromosomes). Several hotspots were apparent in which many genes were associated with sCNV at a particular locus, especially for regions on chromosomes 1, 2, 6, 7, 12, 14 and 20 (see text for additional discussion). D. Genes associated with sCNV hotspots in HCC and cell lines significantly overlap. Hotspot sCNV markers were selected by identifying regions associated with >500 genes (Pearson correlation coefficient>0.3) and then selecting the single top marker per chromosome. The genes associated with each hotspot marker were compared (blue circles, size equivalent to number of genes) and significant overlaps (Fishers Exact Test p<1e-6) are shown as edges connecting pairs of nodes. A similar analysis was performed on a collection of cancer cell lines (CCL, green circles). In this case a smaller number of total genes were measured (23,404 vs 37,585), so an equivalent fraction of the total genes (>370) significantly associated with a sCNV marker was required. The thickness of the edges connecting the nodes represents the enrichment of the overlapping genes in comparison to that expected by chance (observed overlap divided by expected; enrichment <5 fold – thin line, >5 – thick lines). As shown the genes associated with hotspots were shared within datasets. For example, the genes in HCC linked to the hotspots of chromosomes 2, 6, 12, 14 and 20 significantly overlap each other. Hotspot overlaps between CLL and HCC involving the same genomic regions are highlighted in red (chromosomes 1, 14 and 20).

Co-expression modules were derived from both the AN and TU tissues (see Results and Methods) and are represented here as circles. Green (top) and red (bottom) circles are modules derived from AN and TU tissues respectively. The size of the circles indicates the number of genes in each including less than 100 (small), 100–500 (medium) and greater than 500 (large). Genes which were variant within the tissues but did not fall within a co-expression module are collectively represented by diamonds (one for each tissue). The largest modules from each tissue (AN-turquoise and TU-turquoise) are indicated by the boxes. Co-expression modules in AN (top) or TU (bottom) that were enriched for AN-survival genes (green arrows) or TU-survival genes (red arrows) are indicated. Additionally the concentration of cell cycle genes in the TU-salmon module is indicated.

The top 5 differentially connected genes and their differentially connected partners are shown. Each gene is represented by a blue oval and the top 5 are indicated by the boxes. For each of the top 5 genes also indicated is the number of differential correlations between AN and TU in total (DiffConn) and the numbers for gain (GOC) and loss (LOC) of correlation in each case. Lines connecting genes indicates that that pair was differentially correlated between AN and TU where both LOC (blue lines) and GOC (red lines) are indicated. Differential connections between the top 5 genes and any other gene are shown, as well as differential correlations between genes differentially connected to the top 5. The top 5 genes were found to be differentially correlated to a highly overlapping set of partners. Shown in the insert table (top right) is the Fishers Exact Test p value for overlap between differentially correlated gene partners for each of the top 5 genes (lower left of table). Also shown (upper right) is the fold enrichment for the overlaps relative to what would be expected by chance. As shown a complex web of differential correlations resulting from tumorigenesis is revealed.

Shown is a heat map of copy number aberrations (sCNV) for tumor derived samples (Y axis) clustered by K means into 10 groups, versus the linear positions through chromosome 1 (X axis). sCNV was estimated as described in methods and is indicated as a continuum of color from red (amplification) to black (no change) to green (deletion). A scale for the sCNV data is indicated on the right hand side of the heat-map (logR ratio from 1 to −1). As can be seen the majority of aberrations appeared to involve large chromosomal sections on the scale of whole chromosome arms.

A. Shown is the significance of association (as negative log10 of the Cox regression p value) found between all 37,585 genes in AN (X axis) and TU (Y axis) and survival. Genes found to be significantly associated with survival (FDR<0.1, Methods) are indicated in AN (green dots), TU (red dots) or both AN and TU (purple dots). As described in the text most genes predictive in one tissue were not predictive in the other. B. Shown is a representation of the network transformations associated with HCC tumorigenesis (transition from pre-tumor state, upper box, to the tumor state, lower box), where predictive genes in AN (green) largely lose their association to survival in TU following association to non-predictive sCNV. In contrast genes predictive of survival in TU (red) are largely not predictive in AN, and were preferentially associated with sCNV markers that were also predictive. Not shown are genes predictive in both AN and TU, and genes not predictive in either tissue. C. Shown diagrammatically is the “Field-effect” hypothesis as proposed [26] (upper box), where adjacent normal genes predict patient survival because they reflect a milieu (field-effect) in which future tumors are more or less likely to arise. In this model the current tumors do not have a large impact on outcome whereas future tumors do. A modification of this hypothesis is proposed here (lower box), where the adjacent normal genes represent a state that directly affected the probability of the current tumors arising and progressing. In this modified hypothesis survival or death is mediated by the current tumors. See text for additional discussion

Shown is a high level summary of various relationships described in this work. The graph was assembled using the Circos software (http://mkweb.bcgsc.ca/circos/) and data in Tables S2 and S4. Genomic positions are indicated on the outer circle with both chromosome (large numbers) and nucleotide position (smaller numbers, x1,000,000 nt) as indicated. The bars on the outer circle indicate the positions of cytobands. The 2^nd (green background) and 3^rd (red background) circles indicate the genes (by chromosomal location) that predict survival in AN or TU tissues respectively where a longer line indicates greater significance (see Table S2 and text). The 4^th circle (yellow background) indicates genes found to be differentially correlated between AN and TU tissue where the length of the line is proportional to the log10 of the number of differential correlations for that gene (see Table S2 and text). The 5^th circle (dark blue background) shows sCNV markers that predict patient survival where a longer bar is more significant (see Table S4 and text). The 6^th and innermost circle (light blue background) shows the number of genes correlated to each sCNV marker (cis and trans correlations, see Table S4 and text).