As our knowledge of the topology of the genome has evolved, a new class of non-coding RNAs has emerged called microRNAs (miRNAs). The latest release of the miRBase database has catalogued 721 human miRNAs. Smaller than protein-coding genes, miRNAs can regulate the translation of hundreds of genes through sequence-specific binding to mRNA⁴, and depending on the degree of sequence complimentary will result in the inhibition of translation and/or degradation of target mRNAs^4,5. Interestingly, a recent report shows that miR-369-3p can upregulate the expression of its target, tumour necrosis factor-α (TNFα)⁶.

Our knowledge and understanding of miRNA biogenesis has evolved in recent years, and is thoroughly described elsewhere^4,7 (FIG. 1). Briefly, mature miRNAs are short RNA molecules of between 19 and 22 nucleotides in length. Nucleotides 2–7 of the mature miRNA sequence create the ‘seed region’ (REFS ^8–11) that primarily specifies the specific mRNA that the miRNA will bind. The degree of specificity conferred by the seed region is comparable to that of the DNA sites recognized by transcription factors¹². Although the binding between the seed region is (mostly) in perfect Watson–Crick complementarity, flanking regions do not have to bind with equal precision. In an additional analogy to transcription factors, it is now apparent that base pairing outside the seed region provides a further layer of specificity, just as chromatin structure limits the potential for transcription factor binding. As multiple transcription factors work cooperatively to ignite gene expression, so too can multiple miRNAs bind to cognate sites in the 3′ untranslated region (UTR) of target mRNAs. Indeed, the complexity of translation can be further extended through heterotypic miRNA–mRNA interactions, as genes can harbour binding sites for several miRNAs^12–14.

To date, miRNAs have been linked to the aetiology, progression and prognosis of cancer¹⁵, and miRNA expression profiles can uniquely identify cancer types^16,17. The gain or loss of specific miRNAs can function as an oncogene or tumour suppressor^18,19, the archetypical examples of this being miR-21 and Let-7, respectively. It should also be noted that some miRNAs can have dual oncogenic and tumour suppressive roles in cancer depending on the cell type and pattern of gene expression²⁰. In addition, approximately 50% of all annotated human miRNA genes are located in fragile sites or areas of the genome that are associated with cancer^21–23.

Their functional association with cancer, small gene size and potential to simultaneously affect a multitude of genes makes them unique candidate loci for conferring cancer susceptibility, as a small genetic change in an miRNA sequence can theoretically lead to widespread phenotypic effects^24,25.

The initial demonstration that miRNA-related SNPs can affect phenotype was elegantly depicted by Abelson et al.²⁶ who found that a mutation in the miR-189 binding site of SLITRK1 was associated with Tourette’s syndrome. Since then, several studies have used systematic sequencing or in silico approaches to identify SNPs in miRNA-related genes, catalogues of which have been created and made public^27–29. These reports provide fertile ground for follow-up case–control studies to determine the association between these genetic markers and cancer risk.

SNPs in miRNA genes are thought to affect function in one of three ways: first, through the transcription of the primary transcript; second, through pri-miRNA and pre-miRNA processing; and third, through effects on miRNA–mRNA interactions (FIGS 2,​,3).3). See Supplementary information S1 (table) for a list of SNPs in all currently known pre-miRNAs and mature miRNAs, which has been created using build 130 of dbSNP (see Further information). Most studies that have followed a biologically based candidate gene approach to search for SNPs in miRNAs that might confer cancer susceptibility rely on knowledge of a functional link between a particular miRNA and gene target.

The first evidence that point mutations in miRNA genes can have a functional effect and confer cancer susceptibility comes from a seminal study by Carlo Croce’s group³² in which a germline mutation in pri-mir-16-1 was found in a kindred with familial chronic lymphocytic leukaemia (CLL) and resulted in low levels of miR-16-1 expression (FIG. 2). Moreover, the mutation was subsequently discovered in New Zealand black mice that naturally develop CLL-like disease³³. However, this mutation was not detected in a large panel of other tumour types, thereby showing specificity for cancers of a particular origin³⁴. Notably, the oncogenic mir-17–92 cluster contains two miRNAs that harbour three SNPs (see Supplementary information S1 (table)), but the relevance to inherited cancer risk and carcinogenesis remains to be investigated.

To date, the literature suggests that the functional consequences of SNPs in pri-miRNAs relate to processing and levels of the mature miRNA. For example, SNPs in the pri regions of let-7e and mir-16 lead to decreased mature miRNA levels^29,32. Indeed, several studies show an association between pri-mir SNPs and cancer risk (TABLE 1). Specifically, rs7372209 in mir-26a-1 was associated with a 64% decreased risk of bladder cancer in females³⁵, and a twofold increased risk of premalignant oral lesions³⁶. The rs531564 SNP in mir-124-1 is associated with an increased risk of bladder cancer³⁵ and oesophageal cancer in males³⁷, and the pri-miR SNP rs213210 in mir-219 increased the risk of oesophageal cancer³⁷. Another pri-miR SNP, rs2292932 in mir-149, has been tested in several cancers but has not been associated with cancer risk^38,39. This suggests that the molecular mechanisms underlying the genetic associations of mir-SNPs with cancer are complex and vary by cancer site.

Case–control studies have also provided evidence for an association of pre-mir SNPs and cancer risk. Rs11614913 in pre-mir-196a-2 contributes to the risk of developing breast⁴⁰, lung³⁹ and gastric cancers⁴¹ in the Chinese population (TABLE 1). In each case the rs11614913 variant homozygote CC was associated with increased cancer risk. Risk of developing oesophageal cancer in Caucasian males and never-smokers was significantly associated with the rs11614913 variant homozygote TT, the minor allele in this population³⁷. Rs11614913 is located in the 3′ passenger (3p) strand mature sequence of mir-196a-2, thereby possibly affecting both maturation and the repertoire of target mRNAs with which it interacts. Indeed, previous studies have shown that sequence variations in mature and precursor miRNA sequences affect miRNA biogenesis^28,42, and levels of mature miR-196a-2 were lower in CC carriers than in TT carriers³⁹. Notably, this SNP has also been associated with poor survival in patients with lung cancer⁴³. Indeed, this was the first demonstration that miRNA-related SNPs could be related to cancer prognosis.

An SNP in the terminal loop of pre-mir-27a, rs895819, confers a reduced risk of developing breast cancer in families with a history of non-BRCA-related disease⁴⁴ and in families with mutant BRCA2 (REF. ⁴⁵). It has been shown that artificial mutations in the terminal loop of miRNAs such as mir-21 and mir-30 can block miRNA maturation⁴⁶, so it is conceivable that the variant allele of rs895819 might impair the maturation of oncogenic mir-27a, thus explaining the protective effect of the SNP. However, no alterations of free energy or conformation of the miR-27a–mRNA duplex were predicted in silico, thereby leading to the assumption that the SNP would not affect mir-27a maturation or targeting⁴⁵. Nevertheless, SNPs in pre-miRNAs may affect expression even in the absence of apparent effects on its secondary structure. Such is the case for an SNP in let-7e (rs41275792), which leads to reduced levels of the mature miRNA in vivo even though its secondary structure is not predicted to change²⁹. The location of the rs895819 SNP in the centre of the terminal loop of mir-27a is likely to decrease the size of the loop and affect the binding of DROSHA, thus decreasing miRNA maturation^29,46. Alternatively, the SNP might influence the binding affinity of several DROSHA inhibitors, such as Lin28 (REFS ^47,48).

Another pre-miRNA SNP, rs6505162 in pre-mir-423, is associated with an increased risk of bladder cancer³⁵ and ovarian cancer in carriers of mutant BRCA2 (REF. ⁴⁵), and decreased risk of oesophageal cancer in Caucasians³⁷. There is no clear explanation for the opposing effects of this SNP in different cancer types within the same population. Modulations of mature levels of miR-423 have not yet been functionally linked to this SNP, and the RNAFold software (see Further information) does not predict a change in secondary structure⁴⁵. An A/C SNP in mir-30c-2 was predicted to cause the greatest change in target gene identification, and as such was postulated to affect cancer risk³¹, although results from a study in hepatocellular carcinoma (HCC) did not support this hypothesis⁴⁹.

A unique example of a functional miRNA SNP is rs2910164, which is located in the 3p strand of mir-146a (FIG. 3). This polymorphism involves a mispairing in the hairpin of the precursor, which leads to altered processing, lower expression of the mature sequence and predisposition to papillary thyroid carcinoma⁵⁰. Intriguingly, individuals with a heterozygous genotype have a greater risk of developing papillary thyroid carcinoma than homozygous individuals⁵⁰. Heterozygosity as a genetic risk factor is rare, and as postulated by the authors may be a form of genetic epistasis in which the phenotype of the heterozygote differs from the sum of the phenotypes of both alleles. Indeed, in a follow-up study, the same authors found that the SNP fell within the seed of the 3p strand, and would give rise to three mature miRNAs (one from the leading strand and two from the 3p strand) instead of the expected two observed in homozygous individuals⁵¹. The resulting complexity is underscored by the finding that each mature miRNA will target a specific repertoire of mRNAs⁵¹. Interestingly, there is evidence of homozygous to heterozygous somatic mutations at rs2910164 in several patients with papillary thyroid tumours⁵⁰. These seminal studies elegantly decipher how a small genetic change can influence the gene expression profile and show that somatic mutations in miRNAs can be an oncogenic event. In separate studies, rs2910164 was associated with an increased risk of developing prostate cancer⁵² and HCC⁵³ in males through reduced expression levels of miR-146a⁵³. Although not directly associated with breast cancer risk, rs2910164 was associated with a younger age of breast cancer diagnosis in familial breast cancer after adjustment for BRCA1 and BRCA2 mutation status⁵⁴. The variant allele of rs2910164 leads to increased levels of mature miR-146a and binds with greater affinity to BRCA1. Predisposition may therefore develop through the downregulation of BRCA1 (REF. ⁵⁴). Alternatively, rs2910146 could disrupt the well-documented role of miR-146a as a mediator of the pro-apoptotic transcription factor nuclear factor-κB (NF-κB)^55,56. In support of this possibility, genes involved in the regulation of apoptosis were differentially transcribed in heterozygote rs2910146 carriers⁵¹.

Although studies have started to reveal the nature of the association between miRNA SNPs and cancer risk, several considerations remain: most of the studies used a candidate gene approach; of those that used a systematic approach, their lists are outdated owing to enhanced screening techniques that have identified new miRNA genes and updated builds of genome-wide SNP repositories. According to the most recent updates of miRNAs, 202 pre-miR genes have 283 SNPs (Supplementary information S1 (table)). It is likely that as new builds of dbSNP are formulated more SNPs with potential relevance to cancer risk will be identified. In addition, the minor allele frequencies of many of the mir-SNPs already identified have not been determined. Therefore, population studies should be conducted that will ascertain whether or not these SNPs are polymorphic and if so in what populations. This is an important consideration, as data are emerging to suggest that some mir-SNPs have evolved to a high level of variance in distinct populations. For example, the variant alleles of several SNPs occur only in populations of African descent⁵⁷. These epidemiological associations need to be validated in independent populations and functionally tested⁵⁸. Furthermore, the inclusion of mir-SNPs in future genome-wide association studies (GWAs) will help to unveil low-penetrance susceptibility mutations. Most mir-SNPs are not included in current GWAS designs, and as such there is a paucity of information in this regard. The identification of tag sNPs for miRNA-related SNPs will also be a useful endeavour. Clarification of the extent of the pri region of miRNA genes is also needed to more accurately assess miRNA-related genetic variation. Many studies currently limit their analysis of mir-SNPs to the pre and mature regions, as these are clearly defined.

As illustrated in this example, several elements have to converge for an miRNA binding site SNP to be considered functional: the SNP must have a proven association with cancer, both the miRNA and its predicted target must be expressed in the tissue, and the allelic changes must result in differential binding of the miRNA and affect expression of the target gene. Choosing candidate genes to analyse in specific cancer types has been one favoured approach to find such interactions. In this manner, rs17281995 in the 3′ UTR of CD86 was predicted to disrupt the binding sites for five miRNAs and was found to be associated with an increased risk of colorectal cancer⁷⁹. miR-582, which is expressed in normal colon tissue, bound less tightly to the variant allele of CD86, thus increasing its expression level. CD86 functions as a co-stimulatory molecule and increases the production of the pro-inflammatory cytokine interleukin-4 (IL-4)⁸⁰, which might explain the contribution of this SNP to colorectal cancer risk.

Among the 120,000 known SNPs that occur in 3′ UTRs, ~17% destroy putative conserved or non-conserved miRNA binding sites. Furthermore, 8.6% create new predicted target sites according to the Patrocles database⁸¹ (see Further information; Supplementary Information S2 (box)). Yu et al.⁸² searched the 3′ UTRs of all genes in the genome and cross-referenced this with dbSNP 126 and the Targetscan database (see Further information) for possible overlap. Interestingly, they observed that 12 of these SNPs were differentially expressed in cancer EST databases⁸². Subsequently, one of these SNPs, rs16917496 in SETD8, was associated with an increased risk of breast cancer⁸³. Landi et al.⁸⁴ have now catalogued 79 SNPs in the 3′ UTRs of 129 colorectal cancer-associated genes. One of these, an insertion/deletion (indel) polymorphism (rs3783553) in the 3′ UTR of IL1A led to a 38% decrease in the risk of developing HCC. The TTCA insertion allele for rs3783553 disrupts a binding site for miR-122 and miR-378, thereby increasing transcription of IL1A in vitro and in vivo ⁸⁵.

Saetrom et al.⁸⁶ mapped HapMap SNPs to putative miRNA recognition sites in genes deregulated in oestrogen receptor-stratified breast tumours and used local linkage disequilibrium patterns to identify high-ranking SNPs in the Cancer Genetic Markers of Susceptibility (CGeMS) breast cancer GWAS. Two SNPs, rs1970801 and rs1109745, were in strong linkage disequilibrium with rs1434536, an SNP in an miR-125b target site in the 3′ UTR of bone morphogenetic protein receptor type 1B (BMPR1B). Subsequently, rs1434536 was validated and miR-125b was shown to differentially regulate the C and T alleles. These results suggest that allele-specific regulation of BMPR1B by miR-125b explains the observed disease risk⁸⁶.

Drawing on the list of putative polymorphic miRNA binding sites in the genome provided by Chen et al.⁸⁷, 11 possible candidate SNPs were selected for their potential relevance to breast cancer⁸⁸. Subsequently, rs2747648, which resides in a predicted binding site for three miRNAs in the oestrogen receptor-α (ESR1) gene, was associated with a 27% reduction in breast cancer risk in premenopausal women. When the C allele is present, miR-453 binds with greater affinity to ESR1, thus leading to decreased levels of eRα. Postmenopausal women already have reduced levels of endogenous oestrogen, perhaps explaining why this SNP is relevant only in premeno-pausal women. Indeed, the authors pose the interesting question of whether or not carriers of the ancestral T allele would respond better to endocrine therapy, given that they will naturally express increased levels of the receptor⁸⁸. Moreover, there is evidence that the variant allele of another SNP in ESR1, rs93410170, enhances the binding between miR-206 and the 3′ UTR, thereby decreasing eRα levels⁸⁹. Subsequently, the authors postulated that the lower incidences of breast cancer in Hispanic and european populations could be partially associated with the increased prevalence of the variant T allele.

The idea that an SNP in an miRNA binding site could specifically affect pharmacokinetics is illustrated in a study that described the effect of a C–T SNP near the miR-24 binding site in the 3′ UTR of human dihydrofolate reductase (DHFR) on its translation. The variant allele interferes with miR-24 binding to the 3′ UTR and leads to a twofold increase in the mRNA half-life, DHFR overexpression and methotrexate resistance⁹⁰. Moreover, a recent follow-up study found that DHFR imparts a selective growth advantage and neoplastic transformation in immortalized cells⁹¹. However, an important caveat of this study, and one that introduces an additional level of complexity when studying miRNA-related SNPs, is that the SNP that affected drug resistance was not located in the miRNA binding site. Rather, it was located further downstream. The relationship between miRNAs and pharmacogenomics was recently discussed⁹². The concept of ‘integrative epidemiology’ (REF. ⁹³) proposes several models that depict how SNPs may overlap in their relationships to therapeutic response, susceptibility and prognosis⁹⁴. An intriguing observation reminds us that multiple SNPs in the miRNA network may interact in a similar manner in carcinogenesis. DHFR is one of the predicted targets of miR-196a-2, the SNP of which, rs11614913, is associated with cancer risk⁴³. Thymidylate synthase, the target of the cancer drug 5-fluorouracil, is another predicted target of miR-196a-2 (REF. ⁴³). Indeed, recent studies have found that genetic variations in the 3′ UTR of thymidylate synthase are associated with resistance^95,96. Whether rs11614913 affects expression of these genes and therapeutic efficacy is an intriguing question.

A recent paper provides new insight into mir-SNPs and their relevance to cancer. New miRNAs have been found in epstein–Barr virus, which is a causal factor for several cancers⁹⁷. One of these miRNAs, mir-BART22, harbours a single genetic variant that increases its mature levels and downregulates its target, latent membrane protein 2A (LMP2A), a potent immunogenic viral antigen. As a consequence, it is possible that miR-BART22 could accelerate carcinogenesis through the downregulation of LMP2A and the evasion of the host immune response⁹⁸. Whether similar interactions contribute to other viral-related cancers remains to be determined.

Although much less explored, miRNAs can also bind to target sites in the 5′ UTR and open reading frames⁹⁹. Such binding sites can occasionally be interrupted by introns and therefore require splicing to bind with their complementary miRNAs¹⁰⁰. Several lines of evidence support miRNA binding to the 5′ UTR^101–103 and coding sequences^9,11,104. For example, miR-10a binds to the 5′ UTR of ribosomal proteins to increase their translation¹⁰⁵, and miR-148 regulates DNMT3B expression through a conserved site in the protein coding sequence. Interestingly, the target site is absent in the DNMT3B3 splice variant. Therefore, the expression of miR-148 changes the relative abundance of DNMT3B splice variants¹⁰⁶. Let-7 directly targets DICER1 in its coding sequence, thus establishing a mechanism for a miRNA–DICeR1 negative feedback loop¹⁰⁷. This work suggests that the search for miRNA-related susceptibility loci should be expanded to include both the 5′ UTR and coding regions. Sensitive alleles identified in epidemiological studies, but with obscure functional roles, should perhaps be tested under miRNA prediction algorithms that are not limited to the 3′ UTR of genes, particularly if evidence indicates that altered expression of that gene can be associated with specific phenotypes.

By contrast, the nuclear export proteins XPO5 and RAN have been associated with cancer risk³⁷. XPO5 is responsible for miRNA nuclear export, and knocking down its expression leads to reduced miRNA levels¹¹². XPO5 is downregulated in bronchioloalveolar carcinoma and stage 1 lung cancer¹¹³ but upregulated in high-grade prostate cancer¹¹⁴. The SNPs in XPO5 and RAN occur in their 3′ UTRs, suggesting that they might affect mRNA stability. Our analysis of the RAN SNP, rs14035, in the PolymiRTS database suggests that the ancestral allele lies in a binding site for miR-575, which is disrupted by the derived allele that in addition creates a binding site for miR-182*. Although these are in silico results, they raise the possibility that in addition to affecting cancer risk through the disruption of miRNA nuclear export, a more intricate pathway may be involved that includes miRNA regulation.

DICER1 and transactivation-responsive RNA-binding protein (TRBP) mediate pre-miRNA processing. A recent study indicated that DICeR1 functions as a haploinsufficient tumour suppressor in cancer¹¹⁵. Indeed, lower levels of DICER1 mRNA have been associated with decreased cancer survival¹¹⁰. Interestingly, rs3742330 in the 3′ UTR of DICER1 was associated with an increased risk of premalignant oral lesions in individuals with leukoplakia and/or erythroplakia³⁶. Melo and colleagues¹¹⁶ identified two frameshift mutations in TRBP that introduce premature stop codons, resulting in reduced TRBP expression. One function of TRBP is regulating DICER stability, thus these mutations resulted in reduced DICER expression and lower miRNA production and were associated with higher cellular proliferation levels¹¹⁶.

RISC has a pivotal role in guiding single-stranded mature miRNA sequences to their target mRNA sites. The variant allele of the GEMIN3 nonsynonymous SNP, rs197412, was associated with a reduced risk of premalignant oral lesions, and rs197414, also in GEMIN3, was associated with an increased risk of bladder³⁵ and oesophageal cancer³⁷. Therefore, GEMIN3 variants could alter global miRNA homeostasis and have a major effect on cellular signalling pathways. Two SNPs in GEMIN4, rs2740348 and rs7813, were associated with a decreased risk of renal cell carcinoma¹¹¹ and reduced transformation of Hep3B cells¹¹⁷.

The tumour suppressor p53 was recently implicated in miRNA processing¹¹⁸. Through interaction with p68 and DROSHA, p53 facilitates pri-miRNA to pre-miRNA processing. Given the well-documented relationship between p53 mutations and cancer^119–121, it is possible that there might be p53 mutations or SNPs that affect miRNA processing and so increase or decrease the risk of cancer development. This also raises the possibility that p53-associated conditions, such as Li–Fraumeni syndrome, may relate to a global decrease in miRNA production and function.

Although candidate gene approaches can certainly ascertain the effect of a single SNP on an individual’s risk of cancer, the cumulative effect of the inheritance of multiple SNPs in miRNA-related genes might augment risk. Consistent with this idea, an increased risk of oesophageal and bladder cancer was observed in individuals with SNPs in both miRNAs and miRNA processing genes³⁷. Furthermore, some of those mir-SNPs might be in linkage disequilibrium and therefore inherited together. The polygenic model of inherited breast cancer purports that unfavourable combinations of polymorphic genetic variants in low-penetrance susceptibility genes contribute to the excess familial breast cancer risk and most of these genes have not yet been discovered^139–141. As mir-SNPs are rare^27,30, and their minor allele frequencies are globally low, large studies will be needed to draw out their importance.

The complexity of the miRNA network is further intensified by the discovery of miRNA functions that fall outside their classic range. For example, there is evidence of miRNA-mediated increases in protein translation⁶, nuclear import of miRNAs with distinctive hexanucleotide terminal motifs¹⁴² and the secretion of miRNAs^143,144. Furthermore, an alternative miRNA processing pathway has been uncovered in both Drosophila melanogaster and Caenorhabditis elegans that bypasses DROSHA and instead uses a splicing technique to generate miRNA precursors from short intronic sequences (mirtrons)^145–147. How SNPs in miRNAs affect these pathways remains to be tested.

The data reviewed here explore the strong link between alterations in miRNA structure and function and inherited cancer risk. Although the pathways that mediate this risk have not been fully elucidated, there is a clear suggestion that cancer risk is mediated by changes in miRNA sequence and maturation. mir-SNPs affect cancer susceptibility, response to treatment and prognosis^75,78,148. In addition to broadening our understanding of the astounding complexity of how miRNAs function, the study of genetic variation in miRNA networks has expanded our knowledge of the myriad ways in which miRNAs can affect cancer. The 3′ UTRs of genes are involved in multiple levels of regulation⁸⁷. These oftneglected regions are now known to be prime regulators of the transcriptome, and the importance of SNPs in these regions for human traits is exemplified by the range of phenotypes affected by these mutations (for reviews see REFs 59,149). Historically, these regions were not extensively mined for SNP discovery, something that should now be addressed⁷⁴. Moreover, it is clear that miRNA genetic miscellany can affect the diversity of the genome and is related to cancer susceptibility. Several approaches have been used to apply this genetic basis to cancer, and a polygenic, network-based approach should be adopted in the future¹⁴⁰. The validation of these findings in multiple cohorts and the testing of their applicability to different ethnic populations is also required. Furthermore, the linkage of population-based studies to functional validation is crucial for both basic science and the advancement of these findings to clinical applications — this step should no longer be overlooked in epidemiological studies if we are to unravel the implications of these networks in human disease. Therefore, despite the initial insights covered in this Review, we believe that a vast anthology of knowledge remains to be discovered.

Supp Info

Supp Table

01