4.1 Construction of miRNA-regulated PINs
Before constructing miRNA-regulated PINs, we should first find out significantly expressed genes and miRNAs. From the results of SAM analysis (Table 1), we could see some well-known breast cancer related miRNAs. For example, hsa-miR-214-3p [47]
and hsa-miR-335-5p [48] were known to be down-regulated in previous literatures.
Hsa-let-7c, one of let-7 families [49], was found to be down-regulated in previous work.
We also found that miR-21-5p, the sole up-regulated miRNAs in our list, was also found to be up-regulated previously [18, 50]. However, most of miRNAs in our down-regulated list were not known to be changed in previous literature. Those miRNAs might be false positives, however, we could not rule out the possibility that they were novel breast cancer-related miRNAs.
The construction step relied on not only differentially expressed miRNAs and genes but also information in the target prediction databases. Also, the preprocessing step on the target prediction database would play a role in our result. Our current preprocessing parameters, using only the predictions presented in at least 2 of 3 databases, were to reduce some possible false positives in the target prediction databases. This setting would also remove some novel miRNAs, since they were too new to be included in the prediction dataset, thus not suitable for functionary discovery of novel miRNAs. However, this problem could be settled by using only the target prediction database with novel miRNAs instead of requiring that a prediction should exist in at least 2 databases.
After construction of the 18 PINs was completed (Figure 6 to Figure 18), we found that size of the PINs were not similar: some of the miRNAs seems to regulate
large size of PINs, while other miRNAs affect only small number of genes. Small miRNA-regulated PINs may be caused by the strict q-value threshold set on SAM analysis, by the processing steps performed on the target prediction databases discussed before, and possibly by lack of protein-protein interaction data of some proteins in HPRD. Although HPRD was known to be the most comprehensive source of protein-protein interaction data [51], there were some proteins not considered and researched by other investigators, and interaction data for those proteins would not be in HPRD. However, it may be true that some of the miRNA-regulated PINs did be small in breast cancer, since the construction of the PINs were based on differentially expressed miRNAs and genes between normal tissues and tumor samples, and those miRNAs with small PINs may not be as important as others with larger PINs.
We further used activity analysis to check if the miRNA-regulated PINs were activated in certain conditions. In the process of threshold determination, we found that the shape of PCC distributions were different between normal and tumor samples. This may be related to the difference of number of samples (which can affect shape of PCC distribution): we had 11 normal samples, 110 tumor samples in our microarray data (GSE29174). For those miRNAs which is down-regulated in tumor tissue, the PIN it regulates would be activated, due to the repression of the miRNAs became lower, and vice versa. However, only 3 (total 17) PINs of down-regulated miRNAs were activated in tumor. In addition, 7 (total 17) PINs of down-regulated miRNAs were activated in normal condition. This result may suggest participation of other biomolecules, such as transcription factors, which could affect the expression of the PINs, thus we could not observe PIN activation in tumor condition for down-regulated miRNAs.
4.2 Functional analysis of miRNA-regulated PINs
To elucidate functions of miRNA-regulated PINs, we conducted GO enrichment analysis on the PINs (for a list of cancer-associated enriched GO terms, see Table 6).
The first thing we noticed was the miRNA-regulated PINs enriched with apoptosis and / or cell proliferation related GO terms. Such miRNAs were miR-520d-3p, miR-497-5p, miR-125b-5p, miR-21-5p, miR-31-5p, let-7c, and miR-125-5p. Most of the miRNAs were discussed and researched before and known to be related to breast cancer.
mir-125b-5p, which was down-regulated in tumor tissue, was previously known as a tumor suppressor [22], and such finding was consistent to our functional enrichment results. mir-125a-5p-regulated PIN was found to have anti-apoptosis activity and the ability to regulate epithelial cell proliferation, and the miRNA was known to repress cell growth [52].
We also found that some breast cancer specific functions were enriched in our results. For example, in miR-497-5p-regulated PIN, function “androgen receptor signaling pathway” was enriched. Although it is not clear whether androgens were related to breast cancer, androgen receptors were known to be up-regulated in breast cancer and relate to node invasiveness [53].
To further verify the findings of enriched cancer-related functions, we used GOBO for survival analysis. Our hypothesis is that expression of genes annotated with enriched cancer-related terms may be related to survival outcome of the patients. Except some cell death / proliferation related terms, we knew that some pathways or functions are also related to clinical outcomes. For example, cell proliferation-related GO terms have high probability to affect survival of cancer tissue. The patient may become worse if cancer tissue survives. In addition, some signaling pathways were known to enhance
invasiveness, migration abilities, or even decrease the survivability of the patient. For example, BMP signaling pathway was known to confer various tumor cells with migration and invasion ability [54]. NGFR was found to be associated with overall survival of breast cancer [54]. Furthermore, Toll-like receptor 4 was known to promote adhesion and invasive migration in breast cancer [55]. Finally, cytoskeleton is an important part for every kind of cells in cell mobility. Actin filament was known to be participated in the process of invasive migration of cancer cells [56]. Since we had some of these functions in our enriched terms, we wished to test if the expression of gene set annotated to the cancer related enriched terms in the PIN would be related to clinical outcome of patients.
Our results showed in Table 9 and Figure 21. As one can see, only some of the enriched terms were associated significantly with clinical outcome. It may be that the changes of these key genes occurred on protein level, such as protein expression or post-translational modification, so association of such genes to clinical outcomes cannot be explored by tools like GOBO. It may also because the miRNA did not regulate whole pathway, or the miRNA did not target the key part of the pathway directly, thus we cannot find the clinical outcome of gene sets of the enriched GO terms in such condition.
However, we did find some functions associated with clinical outcomes. For example, enriched term “regulation of epithelial cell proliferation” of both miR-125a-5p and miR-125b-5p were found to be associated with overall 10 years survival of patients.
This result acted as evidence that further supported GO enrichment analysis and the discussion before.
Since the miRNAs of the miRNA-regulated PINs were differentially expressed between normal and tumor tissue, and we found some cancer related functions in our functional enrichment analysis, the miRNAs may be used as diagnostic marker for
breast cancer. To verify this, we applied ROC curve analysis on the miRNA expression profile which was obtained from NTU Hospital and was not used in construction process of miRNA-regulated PINs. Notably, our results showed that miR-497-5p, miR-125b-5p, along with some other miRNAs of miRNA-regulated PINs, had good performance when used as diagnostic marker. The results were not beyond our expectation, since miR-497-5p and miR-125b-5p were found to have cancer-related functions in this work and in literature (discussed before).