This study has revealed that a number of miRNAs are strongly differentially expressed during the development of CC, including miRNAs not previously reported. Using the Illumina platform, 39 miRNAs were identified that show highly statistically significant and meaningful fold change alterations in CC tumors. The core of this response showed very similar patterns in all colon tumor types studied. In addition, six miRNAs were identified that are significantly differentially expressed in dMMR tumors compared to pMMR tumors.
As previously described, high-level correlations in the miRNA data derived from both technical and biological replicate experiments were observed over a wide range of detection [21, 33]. In addition, the impact of 2-fold fluctuations in the input RNA population was less than the impact of biological variability between normal samples. As expected, therefore, distinct expression signatures were observed in the current study between normal and colon tumor tissues. Furthermore, unsupervised hierarchical clustering and SVM results showed that this was robust and highly predictive in nature.
Several groups have published miRNA profiles of colon tumors using different platforms with or without an amplification step: RT-PCR for 150 miRNAs , microRNA microarray for 389 miRNAs , and LNA-based oligonucleotide arrays for ~315 miRNAs . The present study is thus the most comprehensive, evaluating 735 miRNAs, the most statistically rigorous, using a Bonferroni correction to the multiple testing problem requiring large fold changes and used a macrodissected set of tumor tissues, therefore minimizing the effect from non-malignant cells. Most notably, 18 of the 39 miRNAs we found significantly altered in CC, have been previously reported by systematic RT-PCR .
Previous work with spotted arrays and qRT-PCR revealed the induction of miR-21 [17, 19] and the repression of miR-143  in CC compared to normal colon samples. Surprisingly, these miRNA did not show differential expression on the Illumina platform. We conducted separate qRT-PCR for these miRNAs and were able to see induction of miR-21 and repression of miR-143 in our tumor set compared to normal tissue. miR-21 showed a significant P-value (P = 2.436e-5 but had an average fold change of only 1.02 between groups. Non-responsive probes for miR-21 and miR-143 all had very high expression levels and very low variance, while responsive probes had higher variance and lower average values.
In order to understand the rationale for this discrepancy, we plotted the miRNA expression profiles relative to the average of normal tissue, alongside the absolute raw expression levels for each probe [see Additional file 6]. Our results demonstrate that miRNAs with high raw intensity values as well as a negative control, show a tight distribution of values whereas that distribution is much more variable in responsive probes. Further comparison reveals that non-responsive probes have very high expression levels and very low variance, while responsive probes have higher variance and lower average values. These effects are also evident following normalization. Filtering the expression dataset for probes with high average expression and low variance reveals 28 probes with these features [see Additional file 7].
Our observation of high expression "silent" probes is consistent with a probe cDNA hybridization model where all probe-binding sites are occupied. According to this model, further increases or decreases in miRNA levels will not be visible by array analyses due to binding saturation. This creates a ceiling above which any change, either increase or decrease in miRNA level, will remain undetectable. Additional evidence in support of this model can be found by looking at the relative concentrations of these miRNA in deep sequencing. As an example, in some cases miR-21 made up 50% of the pool in deep sequencing .
Comparison of arrays generated with both the Illumina platform and spotted arrays for 4 different commonly used cell lines showed that following removal of high average expression low variance "silent" Illumina probes, and low signal probes from the cDNA arrays, the remaining probes cluster together independently of the array platform [see Additional File 8]. This information coupled with our qRT-PCR verification of responsive miRNA, indicated that the Illumina arrays have very low false positive rates; but are potentially susceptible to false negative signal events.
dMMR and pMMR tumors showed differential expression of a small set of miRNA. The molecular etiology of those tumors involving dMMR is very heterogeneous, involving several different genes and numerous mechanisms of gene inactivation, including epigenetic, somatic and germline alterations. Among sporadic CC, the vast majority of cases with dMMR are due to inactivation of MLH1 (~95%), with MSH2 and MSH6 accounting for ~5% and < 1%, respectively . For MLH1, the most common mechanism (~90% of cases) of gene inactivation is promoter hypermethylation . In this study, analyses of dMMR cases were specifically confined to those with loss of MLH1 due to promoter hypermethylation. This was done to achieve a homogeneous group. Thus, the results derived from this study primarily reflect the biology of sporadic MLH1 CC.
Furthermore, hMLH1 methylation-associated microsatellite instability has also been strongly associated with tumors that express the CpG Island Methylator Phenotype (CIMP) . As hMLH1 methylation-associated microsatellite instability generally does not occur among sporadic cases outside the context of CIMP, it appears that the underlying basis for mismatch repair deficiency among this select group of sporadic colon cancer is a broader epigenetic control defect that affects hMLH1 in some, but not all CIMP tumors. CIMP tumors represent another subset of all CC. Thus, the few pMMR samples that closely resemble the dMMR subset as a group, with respect to the miRNA profile, may have an underlying CIMP phenotype, which would be common to both groups. In fact, all but one of the cases, for which the CIMP phenotype was available, cluster within a single group containing both dMMR and pMMR.
Collectively, our data supports a model for colon tumorigenesis encompassing miRNA::mRNA interactions. In support of this model, the predicted mRNA target lists compiled from the 39 altered miRNAs in CC are enriched in tumorigenesis and cancer-related signaling pathways. The simplest explanation of this phenomenon is that up-regulated miRNAs directly or indirectly decrease expression of tumor suppressor proteins in contrast to down-regulated miRNAs that may lead to increased oncogene expression.
The predicted interactions were further explored between up-regulated miRNAs and a set of "drivers" of CC as defined by a forward genetic screen in mice . Many of the induced miRNA found in this study were predicted to interact with the tumor suppressors. In addition, PTEN , SMAD4 , and NOTCH1 , are known tumor suppressors whose transcript or protein levels are decreased in tumors. miRNA-mediated decreases in tumor suppressors provide an elegant explanation for the observed tumor expression patterns.
In this analysis, highly significant increases were observed in miR-135b in CC and the interaction between miR-135b and APC was identified as relevant. It has recently been shown that increased levels of miR-135a/b lead directly to decreased protein expression of the CC tumor suppressor APC via a direct binding interaction between miR-135b and APC mRNA 3' UTR . Although APC mutations are found in a majority of CC, deregulation of miR-135b may have an adverse effect on APC in the remainder of cases.
Inadvertent expression and activation of tumor suppressors could easily disrupt normal growth. High levels of intestinal cell proliferation are required to offset the very high turnover rate of intestinal tissue . We propose that miRNA mediated attenuation of transcription/translation of tumor suppressors is a necessary step in normal intestinal development allowing for cell proliferation. With differentiation, the intestinal epithelial cells no longer replicate, which coincides with a change in miRNA expression and increased levels of tumor suppressors. miRNA expression profiles found in CC show striking similarities with those miRNA profiles found in undifferentiated embryonic stem cells relative to differentiated stem cells (Figure 6A) and cancer cell lines . Embryonic stem cells and colon tumor cells are both capable of almost unlimited mitotic division. A potential interpretation of this observation is that the intestinal crypt cells fail to properly differentiate and may instead continue to actively divide leading to tumor formation (Figure 6B). This is further supported by the observed decreases in miR-1 and miR-133a, which are involved in maintaining the differentiation status of muscle cells . In addition, miR-1 over-expression has been shown to cause expression of differentiated muscle cell mRNA . The decreased levels of miR-1 and miR-133a in dMMR tumors relative to pMMR tumors may also explain why sporadic CC with dMMR show poor differentiation.
Temporal patterns of gene expression during cell-cell adhesion-initiated polarization of cultured human Caco-2 cells found similar transcript changes to those observed during migration and differentiation of intestinal epithelial cells in vivo, despite the absence of morphogen gradients and interactions with stromal cells characteristic of enterocyte differentiation in situ . Here, we propose that miRNAs may be involved in these changes and that the miRNA profile within each cell is modified as cells stop dividing, undergo chromatin remodeling and become further differentiated with limited ability to replicate.
Taken together, these findings suggest that the inability to properly differentiate due to loss of epigenetic control may be an important factor in the development of colon cancer. The role of induced miRNA found in tumors may be to (i) drive tumorigenesis by attenuating the translation of tumor suppressors thereby maintaining a state capable of cell division; or (ii) act as passengers whose levels reflect that they are locked into a state of uncontrolled cell replication.
Several basic questions arise from these findings. How and why are miRNA expression levels altered and how can they be modified? Our analyses show that miR-135b was highly up-regulated in colon tumors. PDRM5 is a tumor suppressor and a target of epigenetic silencing in CC . Of interest, PDRM5 has been shown to bind to the promoter of miR-135b and PDRM5 silencing leads to increased levels of miR-135b. Mechanistically, PRDM5 recruits HDAC1 and G9a to the miR-135b promoter, and this chromatin remodeling results in decreased miR-135b expression. Decreases in PDRM5 expression levels lead to increased expression of miR-135b presumably due to loss of recruitment of this chromatin-remodeling complex. Intriguingly, PDRM5 was also found by CHIP-CHIP to be present at the promoter regions of miR-9, miR-378, miR-196b, miR-96-182-183, as well as additional colon cancer relevant miRNA miR-21 and miR-143-145 . Decreases in PDRM5 were associated with both increases and decreases in miRNA expression. miR-9 was also down-regulated in invasive breast cancers via promoter hypermethylation . Therefore, PDRM5 -mediated chromatin remodeling may be a general epigenetic mechanism related to a wide range of tumors. Drugs that modify chromatin structure and methylation status in a sequence specific fashion may be useful in colon cancer treatment due to the reversible nature of histone modification and DNA methylation. Those that modify miRNA transcript level or mimic/inhibit miRNA function may also be useful.