Stumbling on elusive cargo: how isomiRs challenge microRNA detection and quantification, the case of extracellular vesicles

One reason that makes extracellular vesicles (EVs) so exciting is the large variety of cargo molecules associated with them. The studies suggesting that specific molecules are sorted and assembled into EVs uncovered their great potential in diagnostics – even small changes in specific cargo abundances have been associated with a variety of metabolic conditions. Among all cargo molecules contained in EV-enriched samples, microRNAs have been intensively scrutinized over the last 10 years because their dysregulation has been associated with diseases, and traces of their expression are often present in biofluids – either associated with EVs or in EV-free manner [[1]-[3]]. Furthermore, because of their exceptional stability in biofluids, microRNAs are being intensively tested as potential biomarkers [[2],[4]-[6]]. Since the launch of the Journal of Extracellular Vesicles (JEV), there has been considerable interest in reporting the association of microRNAs with EVs collected from human biofluids or collected from cultured cells: close to one-third of all JEV publications so far deals with some aspect of the detection, delivery or packaging of microRNAs; this proportion is comparable to the literature in the field of EVs published elsewhere. Circulating microRNAs are not always the most abundant class of small RNAs in all biofluids; other classes, like tRNA fragments, might be more abundant on occasions [[7]]. Yet, the popularity of microRNAs prevails. The seminal reports that showed the functional delivery of microRNAs via EVs and that tumour-derived EVs may promote metastasis or prepare the niche for metastatic cell invasion [[8]-[14]], sparked the interest to identify the microRNA expression patterns found in solid and liquid biopsies from individuals affected not only by cancer but by many other diseases. So far, there are two diagnostic tests clinically applicable in the market – ThyraMIR (Interpace Biosciences) and mir-THYpe (ONKOS Molecular Diagnostics); both are based on a microRNA detection panel to test and classify thyroid cancer. However, the first diagnostic panel targeting microRNAs as analytes from EV samples, or biofluids in general, is yet to get to market. One of the challenges faced by the community in detecting microRNAs (or any other class of small RNAs) contained in EV-enriched samples is the limited number of molecules that each vesicle may contain. Estimates of the microRNA content per vesicle resulted in an average of one molecule per 100 individual EVs (i.e. exosomes) [[15]]. (For a thorough explanation about the use and nomenclature of the term "exosomes", refer to Witwer and Théry [[16]]). Clearly, the starting material of biofluid to obtain enough quantities of EVs per sample is essential. Thus, the development of commercial assays to measure microRNA expression has played a crucial role in this matter. Fortunately, the ease in conducting high-throughput microRNA expression profiling is nowadays more sensitive compared to the early years of microRNA research. A typical workflow adopted by many research groups in order to look for candidate biomarkers within a small RNA sample consists of conducting first a high-throughput profiling of microRNAs, to then "validate" the observed expression pattern by using another technique directed to specific candidates with suspected dysregulation [[17]-[20]]. Because of their commercial availability, high sensitivity and ease of use, reverse transcription quantitative real-time PCR (RT-qPCR) assays are by far the most widely used approach to validate high-throughput microRNA profiling experiments [[21]-[24]]. Although this strategy is widely employed and has led to several publications that support the use of microRNAs as early biomarkers, on many occasions, the results observed by using one quantification platform may not be reproducible by another, yielding data challenging to interpret. This workflow has given rise to inconsistencies not only in those microRNAs whose elevated expression pose potential as biomarkers, but even in those expected to remain constant and unrelated to disease [[25]-[27]]. Recent reports addressed this issue and propose that its origin may lie with the inherent variability of microRNA sequences in combination with the use of different detection platforms to validate microRNA expression data. In this editorial, we aim at raising awareness about the lessons learned on the sensitivity and specificity of the most widely used platforms to quantify the abundance of microRNAs. As not all platforms are created equal, their detection mechanism – and in some cases – the detected molecule, may not be equal. We highlight the studies testing how microRNA variants might affect expression profiling experiments and therefore lead to inconsistencies in data – especially when compared across platforms. We believe that in understanding the limitations each platform presents the research conducted to characterize EV-carried and circulating microRNAs may be particularly benefited because of the scarce material available for experimentation. We provide a simple diagram that summarizes the general strategy followed by the few research groups that have managed to overcome these limitations, which could be used as a workflow in cases of low data correlation when assessing the abundance of microRNAs with different detection platforms. The most widely used methods to explore microRNA expression are microarrays, RT-qPCR and small RNA-seq (sRNA-seq) [[17],[18],[22],[24],[28]]. Although the technologic approach of microarrays and RT-qPCR operates by different mechanisms, both meet at one concurrent principle: specific oligonucleotide probes hybridize to microRNA molecules to produce a signal. This condition inexorably requires previous knowledge on the RNA sequences in order to design and synthesize the hybridizing probes (the molecular approach conducted by each platform may be revisited in Yin et al., 2008; and Pritchard et al., 2012). sRNA-seq, on the contrary, does not require previous knowledge on the sequence to detect, and hence has been typically employed in studies aimed to discover novel sequences. As the commercial availability of these three approaches has increased and their cost has reduced over the last 10 years, many laboratories use them routinely to explore microRNA expression in several experimental setups and disease conditions. Even though each type of platform is very robust in detecting microRNAs, it is common to stumble on puzzling results when their expression is validated by using different platforms sequentially. For example, while some microRNAs correlate robustly and present similar expression trends, others show low correlations between platforms – this can be observed for the same microRNA and sample [[18],[21],[22],[29],[30]]. For this reason, microRNAs with acute fold-changes relative to their controls, and therefore attractive as potential biomarkers must be discarded from further analyses because the platform chosen to validate their expression did not reproduce the trend observed with the first one. The expression of those microRNAs with low correlation is typically ignored due to the suspicion of biases in data collection. This situation is detrimental to biomarker discovery. To understand this phenomenon in more detail, one research group explored in vitro how microRNA sequence variants contribute to such discrepancies [[31]-[33]]. First, Nejad et al. reported confounding results in the detection of miR-222 in human fibroblasts: by RT-qPCR, the authors failed to detect the 70% decrease of miR-222-3p in response to IFN initially observed by sRNA-seq. This observation was later confirmed and extended by Pillman et al., who reported that the use of predesigned LNA-based RT-qPCR assays employed in their study – and that of Nejad et al. – targeted the miRBase entry sequence of miR-222-3p, but failed to detect other of its abundant variants, which were only discernable by computational analysis of the sRNA-seq data. Intriguingly, Pillman et al. also noticed that this situation was not unique to their study, and it applied as well for other published data, but in some cases went unnoticed [[31]]. Finally, Magee et al. reported that the Cq values for miR-21 obtained by using one RT-qPCR platform were affected by changing the stoichiometry between its miRBase entry sequence with either synthetic spiked-in variants or cell-extracted RNA [[30]]. Importantly, the Cq values were affected differently by using another RT-qPCR platform. Reports exemplifying the lack of correlation in microRNA analyses are increasing (reviewed by Witwer [[27]]), and despite numerous review articles highlighting the strengths and limitations of each approach [[17],[18],[20],[26],[27],[34]-[36]], the lack of validation is still a recurring theme. Since the first reports of hundreds of conserved microRNAs in animals, their existence has been accompanied by the presence of sequence variants to what is collectively known as mature sequences [[37],[38]]. Historically, except for few exceptions (like let-7 and bantam), mature microRNAs were given a number, and therefore a name, after their discovery in a progressive list (e.g. miR-1, miR-2, and so on). Most mature microRNAs arise from the processing of longer stem-loop precursors in a series of cleavage events by the RNases type III Drosha and Dicer [[39]-[41]]. The intense study of biogenesis and function of microRNAs in the first decade after their discovery led to assume that in general, the variability observed in their mature sequences originated during their biogenesis, being the most frequently varying location the nucleotide positions near their 3′ ends. These positions were assumed to have little to no effect on microRNA function [[38],[42]]. The variants were primarily observed by direct cloning of microRNA-containing DNA concatamers, and later corroborated by northern blot assays; the combination of these two methods made evident that the variability observed for any given microRNA could be in sequence, in size, or a combination of both. With the advent of deep sequencing technologies, our understanding of microRNA variants evolved in the last decade [[43]-[45]], and several laboratories have demonstrated that they are not merely byproducts of imprecise cleavages by Drosha or Dicer [[46]-[48]], but also that play a biological role and might influence mRNA target selection [[49]-[53]], and their presence associates with diseases as well [[54]-[57]]. These variants are referred to as microRNA isoforms, or isomiRs [[58]]. IsomiRs not only appear when there is a high-depth coverage in sequencing reads, but they were also observed since the first massively cloning studies in animals from Profs. Tuschl's and Ambros' laboratories in 2001 [[37],[38]]. In those studies, due to the method employed for the cloning of small RNAs (concatamerization), the depth and coverage of the sequences obtained did not go beyond a few dozen thousands of "reads". It is currently known that isomiRs exist in every sample, regardless of the abundance of their canonical microRNAs, and also that some isomiRs could be the most abundant representative of a microRNA name within a sample. The evidence obtained from experiments aimed to address the inconsistencies in microRNA quantification across platforms suggests that the issue may have roots in the abundance of isomiRs within a sample, and also by how each detection platform deals with them. Notably, this phenomenon is not only constrained to one specific detection platform only, but rather to their sequential combination [[31]-[33],[56]] – e.g. microarray or sRNA-seq followed by individual RT-qPCR validation. To explain how the abundance of isomiRs may affect microRNA detection across platforms we present two hypothetical scenarios: one in which data validation is robust, and another in which is low (Figure 1). Consider a sample that contains several isoforms of a given microRNA of interest (in this case, hsa-let-7a-1); all isoforms are present at various stoichiometries. In the case of robust validation (Figure 1, top), the miRBase entry sequence for hsa-let-7a-1 (red) – also commonly referred to as canonical, reference, archetype, consensus, or |0|0| – is the predominant and most abundant compared to the sum of all other isoforms. If a high-throughput analysis of this sample is conducted by microarray profiling, it is expected that the mature archetype sequence is detected, together with its other isoforms in a bulk signal (checkmarks). A subsequent validation by RT-qPCR for hsa-let-7a-1 would show a high correlation of detection across platforms, as the abundance of all other isoforms has little to no influence in detecting the predominant |0|0| isoform (for isomiR nomenclature refer to [[56],[57],[59]]). The abundance of isomiRs may reveal inconsistencies in the bulk detection of mature microRNAs across detection. Two examples of high (top) and low (bottom) correlation are shown. The archetype miRBase entry sequence (|0|0|) of hsa-let-7a-1 is shown in red, and its isomiRs are shown in black. The hypothetical abundance of each isoform is represented by a horizontal bar (left). NTA, Non-templated nucleotide additions (grey Us). In the case of low validation, however, the miRBase entry sequence might not be the most predominant isoform within the sample (Figure 1, bottom). All platforms will detect the archetype, but the majority of the bulk signal will come from the most predominant isoform, hsa-let-7a-1 |0|–3|. Since some abundant isoforms may not be detected by each platform technology (cross marks), the confounding results will arise because of low correlation. For example, if hsa-let-7a-1 |0|–3| is detected and profiled by microarray, it will not be detected by stem-loop RT-qPCR; conversely, if it is followed by poly-A RT-qPCR, the bulk detection of the different isoforms altogether would yield a higher signal-to-noise. These scenarios are expected because it is known that individual hybridization probes cross-react with highly similar isoforms (e.g. having 1–2 nt difference) [[30]], but the detection tolerance for all microRNA isoforms remains largely unknown, and it is uncertain whether some isoforms might even be detected by each probe (Figure 1, question marks). Therefore, since the bulk detection of each microRNA in a sample is the sum of all its isomiRs detected individually, assays based on the poly-A RT-qPCR approach typically yield higher signal-to-noise ratios compared to those based on the stem-loop (e.g. the hydrolysis-based probes known as Taqman [[60]]) [[22],[29],[33]]. The only platform from which individual isoforms may be deconvoluted from the bulk signal is sRNA-seq. In general, the diversity and relative proportions of all microRNA isoforms in a sample can be assessed quantitatively by computational analyses, regardless of the sequencing technology (Illumina, SeqLL, Invitrogen-Thermo, etc.). Additionally, sRNA-seq offers the advantage to detect non-templated nucleotide additions (NTAs), that do not correspond to the original precursor sequences that give rise to mature microRNAs – NTAs arise from the activity of terminal transferases that typically add extra uridines or adenines at the miR-3′ ends, and their presence is sometimes related to a target-directed microRNA decay mechanism (reviewed in refs. 14 and 35). Despite the current understanding of the importance of isomiRs biology, however, most computational tools used for the analysis of microRNAs expression do not include an isomiR deconvolution feature by default, so the user must specifically look for this option (for a list of tools for isomiR analysis refer to www.tools4mirs.org [[61]]). If there is not an appropriate computational analysis after sRNA-seq profiling and the advantage offered of individual isoform detection with their respective abundances is not harnessed, the detected isomiRs might end up simply portrayed as bulk signals. Therefore, a careful bioinformatic analysis is required to figure out whether the lack of robustness can be explained given the abundance of specific isomiRs in each case. A concern that escapes the scope of this text is the ligation and amplification biases that may occur during library construction, and the type of normalization that limits sRNA-seq profiling (for review on this matter, see references [[2],[5],[20],[25],[28],[29],[62]-[64]]). Poor reproducibility across platforms is the result of a complex blend of several factors, originated by the intrinsic mechanisms by which each platform detects microRNAs. It is essential to highlight that within any sRNA-seq dataset, only a handful of microRNAs uniquely correspond to the archetype miRBase entry. For the majority of microRNA names, there is a rich blend of isoforms that may even outnumber their |0|0| counterpart and this applies for several microRNA names [[31]-[33],[65]-[67]]. In 81 colorectal neoplasia samples, for example, isomiRs account for 70% of all microRNA sequencing reads and their archetype sequences account only for 30% of all microRNA reads. Similar archetype:isomiRs ratios are observed from the sRNAseq data collection at miRBase, or directly from any of the SRA datasets referenced at the repository (miRBase v22 [[65]],), or even by a random selection of SRA files from the Gene Expression Omnibus (GEO) from a variety of human and mouse cell lines and tissues, including biofluids and EV-enriched these microRNA names in general microRNA names whose are the only sequence microRNA names whose are more abundant the sum of all their respective microRNA names whose are abundant the sum of all of their and microRNA names for which one or many individual isomiRs are more abundant their The most aspect of such is that is the (in with et al., and et In other the representative of a microRNA name – for which most commercial probes are directed to hybridize – as a unique sequence in a sample, previous reports and the data at miRBase Since the of EVs is typically their cargo is for example, at in so only a few microRNA molecules could be assembled into one In this not all may correspond to the same microRNA name, but most could be The average of the number of microRNA molecules per vesicle in the of is [[15]]. such it is not that RT-qPCR is the platform for the detection of EV-carried microRNAs – it has been demonstrated that as low as 10 molecules per sample can be detected using this platform Notably, for sRNA-seq the of starting material must be higher of small quantities might be used for library construction, but the tests and must be Additionally, higher PCR amplification are required during library construction, but it is known this acute amplification biases Despite the current lack of nomenclature that the of every class of vesicle the proportions of isomiRs detected in cultured are in EV-enriched studies, however, have reported that specific isomiRs may be in EVs in to suggesting that the proportion cell may et al. reported that (i.e. are in the of microRNAs of compared to their are rather the same group later observed this in EV-enriched samples of from cancer et al. reported an in isoforms for in EVs such was not observed in other studies reported at specific nucleotide positions within microRNAs in EV-enriched samples the that a sequence may influence the of specific microRNAs into or also promote their research is required to the that may promote specific isomiR association with We that the changes observed for specific isomiRs in EV-enriched samples discernable only after bioinformatic analyses of sRNA-seq data, like those of Nejad et al. and Pillman et al. referred Since it is at this the sensitivity at which microarrays or RT-qPCR probes between individual may be ignored in the a situation that may for samples with of small So far, we two different but related that arise with the detection of microRNAs with complex isomiR stoichiometries. First, the discrepancies in expression by the inherent molecular employed by each associated with their sensitivity in detecting or not isomiRs as bulk signals. the diversity of isomiRs for each microRNA name within every sample, by experiments that the same detection platform but the isomiR with spiked-in either with synthetic or cell-extracted It is to that the that for in vitro cultured may also occur for samples of different like solid or liquid biopsies – including and EV-free circulating microRNAs. The in vitro studies that managed to discrepancies in the of expression detected between platforms have one in they all sRNA-seq data and the authors to the isomiR based on the general strategy followed in those studies, we present a simple diagram that could in the to to a test across platforms in cases of suspicion of a high abundance of isoforms (Figure We that this workflow applies to microRNA-containing samples from any or For high-throughput microRNA profiling conducted by sRNA-seq, it is to sequencing reads and look for abundant isoforms (Figure For microarrays, however, if there is no previous knowledge of candidate isoforms, it is challenging to the confounding a high correlation is expected to occur the same detection platform is employed to validate previous results with samples (e.g. stem-loop followed by stem-loop individual regardless of isomiR stoichiometries. the isoforms are sorted it is to design specific assays for their detection by using RT-qPCR probes of commercial and methods (Figure technology has been shown to detect specific isoforms in a of quantification no previous knowledge of candidate isoforms is it is to some or all samples by sRNA-seq to then to isomiR analysis (Figure high-throughput microRNA profiling data are typically validated by conducting specific RT-qPCR assays for microRNAs with acute expression over In those cases the RT-qPCR assays do not reproduce the observed for any candidate an analysis may to previous expression profiling results or the of isomiRs in a sample is conducted by the computational analysis that after sRNA-seq specific isoforms can be detected by RT-qPCR probes or the lack of about the presence of confounding isomiRs within samples requires sRNA-seq profiling and the computational analysis research groups have in detecting specific isoforms of microRNAs by using and commercial probes that either with stem-loop or poly-A RT-qPCR assays few research groups have to methods to detect specific isoforms not available assays like RT-qPCR and look and approaches to the low correlation issue directly due to the abundance of sequences by However, further is still to their and specificity of the presence of isomiRs might be the for cases in which the abundance of candidate microRNAs could be associated with the of a disease but there is a low correlation when their expression is with a different detection platform. The reports of isomiRs and the nucleotide variability observed in EV-enriched samples are an to look more in for expression changes within the isomiR Importantly, understanding the different molecular mechanisms by which each microRNA detection platform operates in the type of analysis in with the expected We the that microRNA detection inconsistencies across platforms do not to data as each platform is robust in its However, isomiR within samples – some microRNAs are present as a rich blend of isoforms, while others are unique molecular the sequence variability by each platform may to – but not – are biomarkers because their expression was validated with another platform. those platforms that hybridizing inexorably assume specific sequences and therefore detect specific mature microRNA sequences The that most available probes are directed to hybridize the miRBase entry sequence to and the microRNA field for several years However, our current understanding of microRNAs requires to their detection to confounding results Although microRNA other miRBase have been (for example, the there is no currently to which archetype be primarily or even whether an archetype be at In our this the for microRNA detection using predesigned groups of microRNAs are profiled as highly represented in specific diseases, either from cultured or from solid and liquid biopsies (reviewed in the for biomarkers examples of individual microRNAs associated with it is common to groups of microRNAs from unrelated whose expression can be used as One to however, is miR-21 – microRNAs, is the of biomarkers in cancer. The by et al. in the in not only microRNA names and their archetype but also their all microRNAs, miR-21 is by far the most associated with the term in the Yet, its presence is of a previous by et al. and also that specific isomiRs to other microRNAs could be even more but are as biomarkers in the Since commercial RT-qPCR assays of miR-21 are and directed the sequence referenced at miRBase for most it is expected that such archetype is detected and the miRBase archetype of miR-21 is not the most abundant of its variants in sRNA-seq data from the and many of its isoforms are high compared to other microRNAs – close to isoforms Additionally, Magee et al. reported that two different commercial RT-qPCR assays for miR-21 do not some of its most abundant isoforms and show a either in experiments that used only synthetic spiked-in RNA as target molecules or by the with cell-extracted RNAs [[30]]. of the miR-21 variants was profiled in those studies that reported a high correlation of different of cancer with the mature If the bulk detection of either by stem-loop or poly-A does not its most abundant isoforms, could it be that this of miR-21 probes can beyond and also cross-react with other similar In a different study, et al. that the detection of miR-21 archetype is a biomarker of colorectal neoplasia is the presence of miR-21 so associated with it be that the of miR-21 RT-qPCR probes with the large proportion of its is more to expression fold-changes obtained by using other In support of et al. reported that a miR-21 isoform – not its archetype – is a more biomarker of cancer One that could explain this phenomenon is that the biomarker literature is microRNAs, like but not because of their potential as biomarkers, but rather because the of sequences their of expression fold-changes across different platforms. For example, microRNAs with abundant for which RT-qPCR probes are more sensitive to sequence may not the validation and therefore are discarded from the and further while others that do the validation, highly in the biomarker We by that not for isomiRs due to the limitations of detection, from the of microRNA as biomarkers and from the development of to diseases. In understanding cargo isomiRs will in the our knowledge to synthetic EVs as microRNA delivery the platform for the microRNA content in EVs would be such that could identify the isomiR per EV per sample, in an For we not to yet for We and for in 1, and for on We and for The authors that they have no This was in by the and to and the and to