|Home | About | Journals | Submit | Contact Us | Français|
MicroRNAs (miRNAs) are a widespread class of regulatory noncoding RNAs with key roles in physiology and development, conferring robustness to noise in regulatory networks. Consistent with this buffering function, it was recently suggested that human miRNAs coevolve with genes in copy number regions (copy number variation [CNV] genes) to reduce dosage imbalance. Here, I compare miRNA regulation between CNV and non-CNV genes in four model organisms. miRNA regulation of CNV genes is elevated in human and fly but reduced in nematode and zebrafish. By analyzing 31 human CNV data sets, careful analysis of human and chimpanzee orthologs, resampling genes within species and comparing structural variant types, I show that the apparent coevolution between CNV genes and miRNAs is due to the strong dependency between 3′-untranslated region length and miRNA target prediction. Deciphering the interplay between CNVs and miRNAs will likely require a deeper understanding of how miRNAs are embedded in regulatory circuits.
MicroRNAs (miRNAs) are small noncoding regulatory RNAs playing essential roles by controlling gene expression and protein output (Bartel 2004). The functional characterization of miRNAs has fallen behind their discovery as a widespread class of regulators since miRNAs were identified in the nematode Caenorhabditis elegans by forward genetic screens (Lee et al. 1993; Reinhart et al. 2000). Few miRNAs have a mutant phenotype despite pervasive purifying selection, suggesting functional redundancy and/or that miRNA functions become apparent when organisms are subject to environmental and genetic perturbations (Miska et al. 2007; Li et al. 2009; Alvarez-Saavedra and Horvitz 2010; Brenner et al. 2010; Meunier et al. 2013; Jovelin and Cutter 2014). Indeed, regulatory circuits involving miRNAs may canalize phenotypes by reducing stochasticity inherent to gene expression and by using noise to create thresholds and stable switches (Hornstein and Shomron 2006; Herranz and Cohen 2010; Ebert and Sharp 2012; Siciliano et al. 2013).
Although the origin of miRNAs in eukaryote lineages is still controversial (Tarver et al. 2012; Moran et al. 2013; Robinson et al. 2013), their function in tissue identity evolved early during animal history (Christodoulou et al. 2010). Yet, a recent study suggests that miRNAs may have evolved as a response to dosage imbalance due to structural variation (Felekkis et al. 2011). Human genes located in copy number regions (copy number variation [CNV] genes) have more miRNA regulators and corresponding sites than non-CNV genes, suggesting that miRNAs coevolve with CNVs (Felekkis et al. 2011). This result is consistent with the finding that miRNAs can buffer phenotypic variation against genomic diversity (Cassidy et al. 2013). Nevertheless, the relationship between structural variation and miRNA regulation needs to be investigated in multiple taxa to determine whether it represents an ancient evolutionary interaction or a derived function.
To address this issue, I first used recent predictions of miRNA target sites in TargetScanHuman 6.2 (Garcia et al. 2011) and CNV annotations in the Database of Genomic Variants (MacDonald et al. 2014) to compare miRNA regulation between human CNV and non-CNV genes. On average, human CNV genes are regulated by 18% more miRNAs and have 23% more binding sites than non-CNV genes (fig. 1), consistent with published results (Felekkis et al. 2011). Second, I investigated the interaction between CNVs and miRNAs using predicted miRNA target sites from TargetScan 6.2 and CNV annotations in three other model organisms: C. elegans, Danio rerio, and Drosophila melanogaster (Ruby et al. 2007; Emerson et al. 2008; Maydan et al. 2010; Jan et al. 2011; Brown et al. 2012; Ulitsky et al. 2012). Similar to human, CNV genes in the fruit fly have greater miRNA regulation than non-CNV genes (fig. 1). However, worm and zebrafish show the opposite pattern with significantly more miRNAs and target sites per non-CNV genes (fig. 1). Similar results are obtained with predicted sites from miRanda (Betel et al. 2008) available in human, fly, and nematode (Supplementary Table S1). A potential drawback with miRNA target site predictors is the rate of false positives. Nevertheless, consistent differences among species are observed when using all predicted sites or a more stringent set of sites filtered by phylogenetic conservation or quality scores (Supplementary Table S1) and when using experimentally validated miRNA-target interactions from miRTarbase (Hsu et al. 2014) in human and worm (Supplementary Table S2). These results indicate that the relationship between structural genomic variation and miRNA regulation is complex and does not necessarily lead to increased miRNA target sites for genes in CNV regions. Moreover, the opposite patterns observed within two protostomes and within two deuterostomes argue against the hypothesis that miRNAs may have evolved under selective pressure to accommodate the fluidity of genomes (Felekkis et al. 2011), or that the hypothesized evolutionary interaction is a unique derived function.
What may be causing the observed differences among species? A simple explanation is that more target sites are predicted in longer 3′-untranslated regions (UTRs). Indeed, the number of miRNAs and the number of sites per gene are strongly correlated with the length of the 3′-UTR in all four species (human: ρmiRNAs = 0.987, ρsites = 0.989; fly: ρmiRNAs = 0.852, ρsites = 0.886; zebrafish: ρmiRNAs = 0.905, ρsites = 0.937; worm: ρmiRNAs = 0.844, ρsites = 0.871; Spearman’s rank correlation, P < 0.0001) (fig. 2). Importantly, the 3′-UTR length of CNV genes in human and fly is on average, respectively, 27% and 39% greater than the 3′-UTR length of non-CNV genes (fig. 2). In contrast, the 3′-UTR of CNV genes in worm and zebrafish is, respectively, 30% and 8% shorter than the 3′-UTR of non-CNV genes (fig. 2). Including predicted sites located in the coding sequence (Schnall-Levin et al. 2010; Liu et al. 2015) gave consistent results (Supplementary Tables S3 and S4). Thus, the apparent coevolution between miRNAs and CNV genes in human and fly may be explained by the strong dependency between 3′-UTR length and miRNA target prediction.
Transcripts with more intense posttranscriptional miRNA regulation may have longer 3′-UTRs, and so differences among species could result from functional differences between CNV and non-CNV genes. To test this possibility, I compared human CNV and non-CNV genes with their non-CNV orthologs using chimpanzee CNVs from Perry et al. (2008). Because human miRNAs benefit from more in-depth annotation (1,267 miRNA families in human vs. 423 miRNA families in chimp), both human CNV genes and non-CNV genes have greater miRNA regulation than their chimpanzee orthologs in non-CNV regions, despite no significant 3′-UTR length differences between orthologs (Supplementary fig. S1A and B). When the analysis is restricted to 406 conserved miRNA families (588 human miRNAs and 507 chimpanzee miRNAs), human CNV and non-CNV genes have 16% more miRNAs per gene than their non-CNV orthologs, but differences in target sites are very small (<1.2%) and not significant (Supplementary fig. S1C). Results are similar when both human and chimp CNV annotations are derived from Perry et al. (2008), although human CNV genes have significantly less miRNAs and target sites than human non-CNV genes in this study (not shown). Thus, the hypothesis that miRNA regulation increases following CNV formation (Felekkis et al. 2011) is not supported when orthologous and presumably functionally conserved genes are compared, and when controlling for biased miRNA annotation between species.
Differences in miRNA regulation of CNV genes among species could result from differential abundance of structural variant subtypes. For instance, miRNAs may preferentially regulate dosage sensitive genes located in regions of increased copy numbers to reduce gene expression levels, or may preferentially buffer stochastic variation of dosage sensitive genes with low expression in regions of decreased copy numbers. I tested this possibility by sorting CNVs that result exclusively in gain or loss of DNA, using information on structural variant types available for human, Drosophila and Caenorhabditis. The majority of miRNA CNV targets is located in CNVs with loss of DNA in human (4,762 CNV loss genes, 80.5 %) and worm (637 CNV loss genes, 87.14%) and in CNVs with gain of DNA in fly (977 CNV gain genes, 67.85%). Nevertheless, there is no clear relationship between miRNA regulation and the type of structural alteration. In fly, non-CNV genes have lower miRNA regulation than both CNV loss and CNV gain genes, whereas non-CNV genes in worm have a larger number of miRNA regulators and target sites than genes in either CNV subtype (P < 0.05). And non-CNV genes in human are less targeted than CNV loss genes (P < 0.0001) but more targeted than CNV gain genes (P < 0.05). In addition, CNV loss genes have more miRNA regulators and binding sites than CNV gain genes in human and in fly, but miRNAs and target sites are more abundant for CNV gain genes than for CNV loss genes in worm (Supplementary Table S5). Moreover, miRNA targeting differences between CNV loss and CNV gain genes are fully consistent with differences in 3′-UTR lengths (Supplementary Table S5). In summary, differential abundance of distinct CNV subtypes cannot explain the observed differences between CNV and non-CNV genes among species.
Patterns of miRNA CNV gene regulation depend on the accuracy of CNV annotations, and so differential coverage of studies identifying CNVs could mask a potential evolutionary interaction between CNVs and miRNAs. Indeed, although the Database of Genomic Variants compiles CNV regions from 52 studies, CNV annotations in other species were derived from a single study. To evaluate how annotations may affect the inference of coevolution between CNV genes and miRNAs, I separately analyzed human CNV data sets from 31 studies with greater than 500 CNV miRNA target genes. Human CNV genes have greater miRNA regulation than non-CNV genes in 17 data sets (55%), lower regulation in eight data sets (26%), and no significant difference in six data sets (19%) (Supplementary Table S6). To further investigate the effect of CNV annotation, I generated 1,000 data sets for each species with 500 random CNV genes and 500 random non-CNV genes. Patterns of miRNA regulation for CNV and non-CNV genes are similar between the original data sets and those resulting from the resampling procedure, although less than 32% of the replicates in zebrafish show significantly lower miRNA regulation for CNV genes (Supplementary Table S7). Importantly, differential miRNA regulation between human CNV and non-CNV genes depends entirely on 3′-UTR length differences in all 31 studies (Supplementary Table S6), and the probability that miRNA regulation significantly differs given that the 3′-UTR length is significantly different is greater than 0.92 in all four species (Supplementary Table S7). These results do not support that CNV genes have longer 3′-UTRs. Instead, they indicate that greater miRNA regulation does not depend on a gene being in a CNV but on a gene having a longer 3′-UTR.
To test whether miRNA regulation differs when controlling for 3′-UTR length differences, I first compared the number of miRNAs and binding sites normalized by the 3′-UTR length (Supplementary Table S8). Second, I predicted miRNA binding sites with TargetScan using 3′-UTR lengths from 100 bp to 1 kb (Supplementary Table S9). Differences between CNV and non-CNV genes are small (<4%) in all species except C. elegans, although some remain statistically significant after normalizing and when genes have the same 3′-UTR length (Supplementary Tables S8 and S9). Thus, these results do not provide support for increased miRNA regulation or miRNA avoidance for genes in regions of CNV.
In conclusion, the apparent pattern of coevolutionary interactions noted in Felekkis et al. (2011) can be explained by the strong correlation between 3′-UTR length and target sites. Moreover, the hypothesis that natural selection favors a tighter posttranscriptional regulation of CNV genes rests on the assumption that miRNAs reduce expression levels to restore dosage balance (Felekkis et al. 2011). However, the relationship between expression level and CNVs is complex. Most CNV genes are dosage insensitive, whereas expression variation can either follow or be reversed with copy number decrease and increase for dosage sensitive genes (Zhou et al. 2011). In addition, genes encoding protein complexes, prone to dosage imbalance (Birchler and Veitia 2012), do not survive long in CNVs (Dopman and Hartl 2007; Schuster-Bockler et al. 2010; Zhou et al. 2011). The results presented here do not support a systemic and consistent relationship between CNVs and miRNAs. Instead, they suggest that deciphering the interplay between miRNAs and structural variants will likely require a deeper and precise understanding of the function of miRNAs within regulatory networks. For instance, miR-9 a but not miR-7 reduces the effect of genomic diversity on phenotypic variation in fly (Cassidy et al. 2013). Depending on their position within regulatory circuits and the type of loops they form with transcription factors, miRNAs can either attenuate or amplify expression variation (Hornstein and Shomron 2006; Herranz and Cohen 2010; Leung and Sharp 2010; Ebert and Sharp 2012; Siciliano et al. 2013). This may explain why expression variation within and among species is elevated for some miRNA target genes but reduced for others (Cui et al. 2007; Lu and Clark 2012).
The author thank George Bell for providing the Summary Counts table for TargetScanFly and two anonymous reviewers for constructive comments on the manuscript. This work is supported by a grant from the National Health Institutes (R01-GM096008) to P.C.P. and A.D.C.