Identifying factors associated with the direction and significance of microRNA tumor-normal expression differences in colorectal cancer

Dysregulation of microRNAs, which are small non-protein-coding RNA molecules that regulate gene expression [1‐3], has been of interest in colorectal cancer patients [4‐6] due to the potential epigenetic role of microRNAs in disease progression and survival. Within the context of colorectal cancer patients, we have previously reported on the prognostic role of various microRNAs in disease stage and colorectal cancer-specific mortality [7], on differential expression of microRNAs between tumor and normal samples [8, 9], on predictive microRNAs for differentiating carcinoma from normal mucosa [10], on site-specific associations of microRNAs and survival [11], and on associations of microRNA expression with cigarette smoking [12] and other diet and lifestyle factors [13].

In this study we focus on microRNAs that, in terms of tumor-normal expression differences, exhibit evidence of being up-regulated in some subjects but down-regulated in others, or that are dysregulated only for a subset of the population. We present and implement an approach to identify factors (lifestyle, tumor molecular phenotype, and survival-related) that are associated with the direction and/or significance of these microRNAs’ tumor-normal expression differences. It is important to note that our interest here does not lie simply in identifying microRNAs that are differentially expressed between tumor and normal tissues. Rather, our interest lies in identifying factors that are associated with the direction and/or significance of microRNA differential expression. Considering additional factors beyond the tumor/normal distinction allows for greater specificity in conclusions regarding differential expression, as microRNA expression seems to be quite dynamic. For example, rather than simply concluding that a given microRNA is significantly dysregulated in tumor compared to normal tissue, we can identify sub-groups of subjects (corresponding to levels of a particular factor) where the dysregulation is no longer significant or even changes direction – with the microRNA tending to be up-regulated in one sub-group but down-regulated in another. This work has the goal of identifying such cases where factors of interest are associated with the direction and significance of microRNA tumor-normal dysregulation in colorectal cancer subjects.

Using expression data for 1394 microRNAs and 1836 colorectal cancer subjects (each with both tumor and normal samples), many microRNAs exhibited multimodal tumor-normal expression differences, as in Fig. 1. The left mode (near −2 in Fig. 1) corresponds to subjects in which the microRNA is down-regulated in tumor, while the right mode (near +2 in Fig. 1) corresponds to subjects in which the microRNA is up-regulated in tumor. The center mode (near 0) can actually be considered two components – one in which the expression differences are exactly 0 (the tall spike in the left panel of Fig. 1), and another in which expression differences are spread around 0 (more easily seen in the right panel of Fig. 1, where expression differences of 0 have been dropped). These center components correspond to subjects in which the microRNA is (either exactly or essentially) not dysregulated in tumor.

When Hartigan’s dip statistic was used to test each microRNA’s tumor-normal difference distribution for unimodality, and the family-wise error rate was controlled at 0.05, this resulted in 122, 123, and 276 microRNAs identified as having multimodal distributions in proximal, distal, and rectal tumor sites, respectively. There were 66 microRNAs exhibiting multimodality in all three tumor sites. After subsequent application of the Wilcoxon Signed Rank test and classification of each microRNA as “Up”, “Down”, or “NS” as described in the “Statistical Analysis” section above, Table 2 summarizes the resulting numbers of microRNAs classified to each outcome (“Up”, “Down”, “NS”) overall and within each factor level, across all site / factor combinations.

For convenience in summarizing results, outcomes of interest in Table 2 are given superscripts corresponding to color names, as reported in Table 2. Representative results for each of these colors (i.e., outcomes of interest) are given in Fig. 2; full results for all colors (i.e., outcomes of interest) are given in Additional files 2, 3, 4, 5 and 6. Each row of plots in Fig. 2 (and each page of plots in Additional files 2, 3, 4, 5 and 6) has the same format, which can be summarized as follows, using the yellow row of Fig. 2 (plots m-o) as an example. The plot titles indicate which microRNA (miR-196a-5p) and site (proximal) are considered, and the left plot is a histogram of the tumor-normal expression difference of the indicated microRNA at the indicated site, using data from all subjects (with sample size reported in the second row of the title). The one-sided p-value from the Wilcoxon Rank Sum test for differential expression, after adjustment to control the false discovery rate, is reported in the third row of the plot’s title. Adjusted one-sided p-values close to 0 (less than 0.025) suggest down-regulation in tumor relative to normal, while those close to 1 (greater than 0.975) suggest up-regulation. The second row of the titles of the center and right plots indicate which factor levels are considered, with histograms representing tumor-normal expression differences for the same indicated microRNA in corresponding subsets of the data. Subset sample sizes and significance test results are reported in the second and third rows of the plot titles, respectively. Taken together, the row of yellow plots in Fig. 2 (plots m-o) indicate significant overall up-regulation of miR-196a-5p in 567 proximal colon cancer patients (left plot), similarly significant up-regulation of the same microRNA in 391 proximal colon cancer patients whose tumors lack the BRAF mutation (center plot), but significant down-regulation of the same microRNA in 73 proximal colon cancer patients whose tumors have the BRAF mutation.

In Fig. 2a-c (and Additional file 2; the “blue” outcomes of interest), the left plot indicates an overall tendency of significant differential expression, while the center and right plots disagree on the statistical significance. The same direction and statistical significance of the left plot (the overall test) is reflected in only one of the factor level subsets (as in Fig. 2c). For some microRNA / factor / site combinations this may be due to a smaller effective sample size (and consequent loss of statistical power) in one of the factor levels, particularly for factors whose levels are greatly unbalanced (such as BMI_extreme; see Table 1). However, for most microRNA / factor / site combinations, this outcome can be seen in the shapes of the tumor-normal expression difference distributions – such as the more pronounced negative mode in Fig. 2c resulting in a statistical conclusion of down-regulation, but the more balanced (if not entirely symmetric) modes in Fig. 2b failing to provide overwhelming evidence of any differential expression. Such a “blue” outcome can generally be interpreted as a microRNA that is overall significantly dysregulated in tumor vs. normal, but only for one of the factor’s levels.

In Fig. 2d-f (and Additional file 3; the “green” outcomes of interest), the left plot indicates a lack of evidence of differential expression, usually due to a relative balance between the numbers of negative and positive tumor-normal expression differences (as in Fig. 2d). Such a balance (and corresponding lack of statistical significance) is also seen in one of the factor levels (as in Fig. 2e), but not in the other factor level which has a more pronounced mode on one side or the other (as in the negative mode of Fig. 2f). This is indicative of a microRNA (such as Fig. 2d-f miR-640 in proximal tumor) that is significantly dysregulated in only one of the factor’s levels (here, down-regulated in subjects who regularly consume any wine).

Figure 2g-i (and Additional file 4; the “orange” outcomes of interest) present an interesting scenario where a microRNA is overall not significantly dysregulated in tumor vs. normal, but upon consideration of subject sub-groups it is determined that the microRNA (miR-4461 in Fig. 2g-i) tends to be significantly down-regulated in one factor level (distant or regional SEER summary stage rectal tumors) but significantly up-regulated in the other factor level (local SEER summary stage rectal tumors). Such outcomes are rare (see Table 2), but interesting.

In Fig. 2 (and Additional files 2, 3, 4, 5 and 6), the sample sizes of the subsets (center and right plots) within each row do not necessarily add up to the total sample size (left plot). This occurs here because of missing values in some factors defined in Table 1. In the overall test of differential expression (left plots in Fig. 2 and Additional files 2, 3, 4, 5 and 6), all subjects (with tumors in the indicated site) are used in the test of differential expression, and this sample size is reported in the second row of the plot title. In the tests of differential expression within factor level subsets (center and right plots in Fig. 2 and Additional files 2, 3, 4, 5 and 6), only subjects with recorded values for the indicated factor (and with tumors in the indicated site) were used in the test of differential expression, and these sample sizes are reported in the second row of the plot title. The widespread presence of missing values in several factors here contributes to an effective loss of statistical power for many of these subset tests of differential expression, which is the most likely explanation for the “purple” outcomes of Fig. 2 and Table 2, where all but one such outcome involved a microRNA being significantly dysregulated overall with a larger sample size, but not significantly dysregulated in either factor level subset (where the sample size was much smaller). Consequently, the “purple” outcomes are of lesser interest than the others, which are summarized in greater detail for specific factors by site in Table 3. All outcomes of interest from Tables 2 and 3 are summarized in greater detail in Additional file 7.

In presenting these results, we report all subject-level factors that we considered, acknowledging that some overlap, redundancy, or even superiority between factors may be possible. For example, while survival at five years (SURV5YRS) may be a better indicator for overall survival, there may also be additional value to some researchers in considering the status of the patient at last follow-up (VITAL_ALIVE), so the results for both factors are reported here. Also for example, the degree of concordance between SEER and AJCC staging is approximately reflected in the results – for example, AJCC_3 = 0 and STAGE_D = 1 both refer to patients with distant metastasis (Table 1); of the 28 (AJCC_3) and 32 (STAGE_D) “blue” outcomes in proximal colon reported for these factors in Table 3, an examination of Additional file 7 reveals that 26 microRNA outcomes are in common. (These two factors’ results are not identical because the original data are actually slightly different – of the 102 AJCC_3 = 0 proximal colon patients in Table 1, only 92 were STAGE_D = 1.) Although such overlap, redundancy, or even superiority between factors reported here may be noted by some researchers, we have chosen to be broad in the reporting of our results, in the interest of providing more information.

A disproportionate number of outcomes of interest in Table 3 occur for the rectal site, particularly for the “blue” outcomes. In other words, while there are many microRNAs that are significantly differentially expressed in the tumor vs. normal comparison, but that are only differentially expressed in one of the levels of some factor of interest, such outcomes are especially common in rectal site comparisons. Additionally, more than half of the microRNAs with multimodal tumor-normal expression differences in rectal cancer have their significance associated with MSI (186 of 276) or BRAF tumor status (176 of 276).

Differential expression of microRNAs in colorectal cancer is a multi-faceted phenomenon, with multiple factors sometimes being associated with the direction and significance of differential expression of the same microRNA at a given site. For example, Table 3 reports that for each of the factors MSI, CIMP, and BRAF, there is one microRNA that is significantly dysregulated in proximal colon tumor relative to proximal colon normal mucosa, but that is significantly dysregulated in the opposite direction for one of the factors’ levels (i.e., a “yellow” outcome). In fact, this is the same microRNA for all three factors, as represented in Fig. 3. Figure 3a indicates that overall, microRNA miR-196a-5p tends to be significantly up-regulated in tumor vs. normal. Figure 3b, d, and f demonstrate that in the absence of MSI (i.e., for MSS), for CIMP status low, or in BRAF-mutated tumors, respectively, (i.e., at factor levels 0) this microRNA tends to be significantly up-regulated, with a bimodal tumor-normal expression difference distribution with major node favoring positive values. However, Fig. 3c, e, and g show that in the presence of MSI, for CIMP status high, or in BRAF-mutated tumors, respectively, (i.e., at factor levels 1) this microRNA tends to be significantly down-regulated, with bimodal distributions whose modes corresponding to negative values are more pronounced.

It is important to note that all of the conclusions of this study (and resulting classifications of outcomes in interest in Tables 2 and 3) are reached after controlling the overall false discovery rate at 0.05. This means that only as much as 5% of the significant findings in this paper can be expected to be false positives. While alternative error rate thresholds could have been selected, it is encouraging to have so many significant results after controlling for multiple comparisons across so many microRNAs, sites, and factors of interest.

While the results reported in Table 3 (and full results in Additional file 7) involve too many microRNAs to discuss at length individually in this manuscript, we can demonstrate the potential clarifying utility of these results (particularly Additional file 7) by referring to the following few representative examples.

MicroRNAs miR-1266 and miR-4727-3p were classified as the two “orange” outcomes for distal colon tumors in Table 3, being not significantly differentially expressed overall, but differentially expressed in different directions for levels of the SURV5YRS factor. Additional file 7 shows that miR-1266 and miR-4727-3p did not show strong evidence of overall tumor-normal differential expression (respective one-sided FDR-adjusted p-values 0.8111 and 0.6329), but were significantly down-regulated (one-sided FDR-adjusted p-values 0.0077 and 0.0153) in subjects that did not survive five years beyond diagnosis, and were significantly up-regulated (one-sided FDR-adjusted p-values 0.9996 and 0.981) in subjects that survived beyond five years. These findings are consistent with those previously reported in the literature. miR-1266 has been shown to be significantly down-regulated in gastric cancer tissues [21], with higher expression values correlating with longer patient survival times [22]. miR-4727-3p has been shown to bind with the BUB1 gene [23], lower expression levels of which have previously been shown to be associated with shorter relapse-free survival after surgery for colon carcinoma [24].

For several years miR-145 has been of interest in rectal cancers as a possible tumor-suppressor [25, 26], being significantly down-regulated in colorectal carcinoma (with up-regulation in response to neoadjuvant chemotherapy) [27]. Our results are consistent with this literature – miR-145-3p was found to be significantly down-regulated in rectal tumors (one-sided FDR-adjusted p-value <0.0001). However, our results (Additional file 7) provide additional insight, as miR-145-3p was classified as a “blue” outcome in rectal tumors in Table 3, exhibiting significant overall down-regulation in rectal tumor, for MSS subjects (one-sided FDR-adjusted p-values <0.0001 for MSS vs. 0.2267 for MSI subjects), for non-BRAF-mutated tumors (one-sided FDR-adjusted p-values <0.0001 for non-BRAF-mutated vs. 0.5 for BRAF-mutated tumors), for non-wine-drinking subjects (one-sided FDR-adjusted p-values <0.0001 for non-wine-drinking vs. 0.1657 for wine-drinking subjects), and for non-extremely-obese subjects (one-sided FDR-adjusted p-values <0.0001 for non-extremely-obese vs. 0.2267 for extremely-obese subjects). The public availability of Additional file 7 makes such clarifying insights widely available for colorectal cancer researchers.

Previously, miR-130a has been shown to play a complex role in tumorigenesis, being down-regulated in chronic lymphocytic leukemia [28] but up-regulated in nonsmall cell lung cancer and chronic myeloid leukemia [29, 30], and also up-regulated in colon cancer (compared to paired adjacent normal mucosa) [31]. In rectal cancer, up-regulation of miR-130a-3p is significantly associated with better survival [11]. Additional file 7 shows that in rectal cancer, miR-130a-3p is classified as a “blue” outcome, being significantly up-regulated, but only for low-CIMP subjects (one-sided FDR-adjusted p-values 0.999 for low-CIMP vs. 0.6341 for high-CIMP subjects), non-BRAF-mutated tumors (one-sided FDR-adjusted p-values 0.9987 for non-BRAF-mutated vs. 0.7154 for BRAF-mutated tumors), TP53-mutated tumors (one-sided FDR-adjusted p-values 0.9987 for TP53-mutated vs. 0.8521 for non-TP53-mutated tumors), non-KRAS-mutated tumors (one-sided FDR-adjusted p-values 0.9992 for non-KRAS-mutated vs. 0.7466 for KRAS-mutated tumors), non-distant SEER summary stage (one-sided FDR-adjusted p-values 0.9993 for non-distant vs. 0.6568 for distant subjects), subjects with AJCC stage less than 4 (one-sided FDR-adjusted p-values 0.9998 for subjects with AJCC stage less than 4 vs. 0.2258 for subjects with AJCC stage 4), subjects alive at last follow-up (one-sided FDR-adjusted p-values 0.9994 for subjects alive at last follow-up vs. 0.8333 for subjects not alive at last follow-up), subjects reaching at least five years survival (one-sided FDR-adjusted p-values 0.9984 for five-year survival vs. 0.8875 for subjects surviving less than five years), non-current smokers (one-sided FDR-adjusted p-values >0.9999 for non-current smokers vs. 0.4375 for current smokers), subjects with BMI outside normal range (one-sided FDR-adjusted p-values 0.9998 for non-normal BMI range vs. 0.9136 for normal BMI range subjects), or subjects with non-extreme BMI (one-sided FDR-adjusted p-values 0.9994 for non-extreme BMI vs. 0.3929 for extreme BMI subjects). In addition, Additional file 7 shows miR-130a-3p as a “yellow” outcome in rectal cancer cases, being up-regulated in MSS tumors (one-sided FDR-adjusted p-value 0.9998) but down-regulated in MSI tumors (one-sided FDR-adjusted p-value 0.0134). At a minimum, this all suggests the need to account for some of these factors when considering the prognostic role of miR-130a-3p in rectal cancer subjects. It also raises questions regarding the potential roles these factors could play in affecting the survival of rectal cancer subjects.

The results of this study demonstrate the benefit to colorectal cancer researchers to consider multiple subject-level factors when studying dysregulation of microRNAs, whose tumor-related changes in expression can be associated with multiple factors. In instances where microRNAs can be both up and down regulated, depending on specific factor levels, the consequences of not accounting for these factors would most likely be failure to detect any association with the microRNA. In other instances, failure to consider subject-level factors most likely would result in underestimation of the significance of the association. Since direction of regulation could be an important component when considering functionality of microRNAs, this information is important from a potential translational perspective. Our Additional file 7 will serve as a publicly-available resource to provide clarifying information about various factors associated with the direction and significance of tumor-normal differential expression of microRNAs in colorectal cancer.