Author response: Predicting effective microRNA target sites in mammalian mRNAs

Full text Figures and data Side by side Abstract eLife digest Introduction Results Discussion Materials and methods References Decision letter Author response Article and author information Metrics Abstract MicroRNA targets are often recognized through pairing between the miRNA seed region and complementary sites within target mRNAs, but not all of these canonical sites are equally effective, and both computational and in vivo UV-crosslinking approaches suggest that many mRNAs are targeted through non-canonical interactions. Here, we show that recently reported non-canonical sites do not mediate repression despite binding the miRNA, which indicates that the vast majority of functional sites are canonical. Accordingly, we developed an improved quantitative model of canonical targeting, using a compendium of experimental datasets that we pre-processed to minimize confounding biases. This model, which considers site type and another 14 features to predict the most effectively targeted mRNAs, performed significantly better than existing models and was as informative as the best high-throughput in vivo crosslinking approaches. It drives the latest version of TargetScan (v7.0; targetscan.org), thereby providing a valuable resource for placing miRNAs into gene-regulatory networks. https://doi.org/10.7554/eLife.05005.001 eLife digest Proteins are built by using the information contained in molecules of messenger RNA (mRNA). Cells have several ways of controlling the amounts of different proteins they make. For example, a so-called ‘microRNA’ molecule can bind to an mRNA molecule to cause it to be more rapidly degraded and less efficiently used, thereby reducing the amount of protein built from that mRNA. Indeed, microRNAs are thought to help control the amount of protein made from most human genes, and biologists are working to predict the amount of control imparted by each microRNA on each of its mRNA targets. All RNA molecules are made up of a sequence of bases, each commonly known by a single letter—‘A’, ‘U’, ‘C’ or ‘G’. These bases can each pair up with one specific other base—‘A’ pairs with ‘U’, and ‘C’ pairs with ‘G’. To direct the repression of an mRNA molecule, a region of the microRNA known as a ‘seed’ binds to a complementary sequence in the target mRNA. ‘Canonical sites’ are regions in the mRNA that contain the exact sequence of partner bases for the bases in the microRNA seed. Some canonical sites are more effective at mRNA control than others. ‘Non-canonical sites’ also exist in which the pairing between the microRNA seed and mRNA does not completely match. Previous work has suggested that many non-canonical sites can also control mRNA degradation and usage. Agarwal et al. first used large experimental datasets from many sources to investigate microRNA activity in more detail. As expected, when mRNAs had canonical sites that matched the microRNA, mRNA levels and usage tended to drop. However, no effect was observed when the mRNAs only had recently identified non-canonical sites. This suggests that microRNAs primarily bind to canonical sites to control protein production. Based on these results, Agarwal et al. further developed a statistical model that predicts the effects of microRNAs binding to canonical sites. The updated model considers 14 different features of the microRNA, microRNA site, or mRNA—including the mRNA sequence around the site—to predict which sites within mRNAs are most effectively targeted by microRNAs. Tests showed that Agarwal et al.'s model was as good as experimental approaches at identifying the effective target sites, and was better than existing computational models. The model has been used to power the latest version of a freely available resource called TargetScan, and so could prove a valuable resource for researchers investigating the many important roles of microRNAs in controlling protein production. https://doi.org/10.7554/eLife.05005.002 Introduction MicroRNAs (miRNAs) are ∼22-nt RNAs that mediate post-transcriptional gene repression (Bartel, 2004). Bound with an Argonaute protein to form a silencing complex, miRNAs function as sequence-specific guides, directing the silencing complex to transcripts, primarily through Watson–Crick pairing between the miRNA seed (miRNA nucleotides 2–7) and complementary sites within the 3′ untranslated regions (3′ UTRs) of target RNAs (Lewis et al., 2005; Bartel, 2009). The miRNAs conserved to fish have been grouped into 87 families, each with a unique seed region. On average, each of these families has >400 conserved targeting interactions, and together these interactions involve most mammalian mRNAs (Friedman et al., 2009). In addition, many nonconserved interactions also function to reduce mRNA levels and protein output (Farh et al., 2005; Krutzfeldt et al., 2005; Lim et al., 2005; Baek et al., 2008; Selbach et al., 2008). Accordingly, miRNAs have been implicated in a wide range of biological processes in worms, flies, and mammals (Kloosterman and Plasterk, 2006; Bushati and Cohen, 2007; Stefani and Slack, 2008). Critical for understanding miRNA biology is the accurate prediction of miRNA–target interactions. Although numerous advances have been made, accurate and specific target predictions remain a challenge. Analysis of preferentially conserved miRNA-pairing motifs within 3′ UTRs has led to the identification of several classes of target sites (Bartel, 2009). The most effective canonical site types, listed in order of decreasing preferential conservation and efficacy, are the 8mer site (Watson–Crick match to miRNA positions 2–8 with an A opposite position 1 [Lewis et al., 2005]), 7mer-m8 site (position 2–8 match [Brennecke et al., 2005; Krek et al., 2005; Lewis et al., 2005]), and 7mer-A1 site (position 2–7 match with an A opposite position 1 [Lewis et al., 2005]). Experiments have confirmed that the preference for an adenosine opposite position 1 is independent of the miRNA nucleotide identity (Grimson et al., 2007; Nielsen et al., 2007; Baek et al., 2008) and due to the specific recognition of the target adenosine within a binding pocket of Argonaute (Schirle et al., 2014). Two other canonical site types, each associated with weaker preferential conservation and much lower efficacy (Friedman et al., 2009), are the 6mer (position 2–7 match [Lewis et al., 2005]) and offset-6mer (position 3–8 match [Friedman et al., 2009]). Pairing to the 3′ end of the miRNA can supplement canonical sites, although evidence for the use of this 3′-supplementary pairing is observed for no more than 5% of the seed-matched sites (Brennecke et al., 2005; Lewis et al., 2005; Grimson et al., 2007; Friedman et al., 2009). Some effective sites lack canonical seed pairing. For example, very extensive pairing to the 3′ region of the miRNA can compensate for a wobble or mismatch to one of the seed positions (Doench and Sharp, 2004; Brennecke et al., 2005; Bartel, 2009), as exemplified by the two let-7 sites within the 3′ UTR of Caenorhabditis elegans lin-41 (Reinhart et al., 2000). Although these 3′-supplementary sites can be detected above background when searching for preferentially conserved pairing configurations, they are exceedingly rare, with conserved miRNA families in mammals and nematodes each averaging <1 preferentially conserved 3′-supplementary site (Friedman et al., 2009). Other relatively rare, yet effective sites include centered sites, which have 11–12 contiguous Watson–Crick pairs to the center of the miRNA (Shin et al., 2010), and cleavage sites, which have the very extensive pairing required for Argonaute-catalyzed slicing of the mRNA (Yekta et al., 2004; Davis et al., 2005; Karginov et al., 2010; Shin et al., 2010). The existence of additional, still-to-be-characterized types of non-canonical sites is suggested by the large number of mRNA regions that crosslink to the silencing complex in vivo yet lack known site types matching the cognate miRNA (Chi et al., 2012; Loeb et al., 2012; Helwak et al., 2013; Khorshid et al., 2013; Grosswendt et al., 2014). With the prediction of hundreds of conserved targets for most mammalian miRNAs (and even more nonconserved targets), knowing which targets are expected to be most responsive to each miRNA provides important information for both large-scale network analyses and detailed experimental follow-up. As previously mentioned, the type of site (e.g., whether the site is an 8mer or a 7mer-A1) strongly influences the efficacy of repression. The number of sites also influences efficacy, with each additional site typically acting independently to impart additional repression (Grimson et al., 2007; Nielsen et al., 2007), although sites between 8–40 nt of each other tend to act cooperatively, and those < 8 nt of each other act competitively (Grimson et al., 2007). Additional features of site context help explain why a given site (e.g., a 7mer-m8 site to miR-1) can be more effective in one 3′ UTR than it is in another. These features include the positioning of the site outside of the path of the ribosome (which includes the first 15 nt of the 3′ UTR [Grimson et al., 2007]) and the positioning of the site within 3′-UTR segments that are more accessible to the silencing complex, as measured by either high local AU content (Grimson et al., 2007; Nielsen et al., 2007), high AU content of the entire 3′ UTR (Robins and Press, 2005; Hausser et al., 2009), shorter distance from a 3′-UTR terminus (Gaidatzis et al., 2007; Grimson et al., 2007; Majoros and Ohler, 2007), shorter 3′-UTR length (Hausser et al., 2009; Betel et al., 2010; Wen et al., 2011; Reczko et al., 2012), or less stable predicted competing secondary structure (Robins et al., 2005; Ameres et al., 2007; Kertesz et al., 2007; Long et al., 2007; Tafer et al., 2008). Conserved sites are also more effective, in part because they tend to reside in more favorable contexts (Grimson et al., 2007; Nielsen et al., 2007). Features of the miRNA can also influence site efficacy, with sites being more effective if the miRNA has lower target-site abundance (TA) within the transcriptome (Arvey et al., 2010; Garcia et al., 2011) and stronger predicted seed-pairing stability (SPS) (Garcia et al., 2011). Multiple features can be considered together to build quantitative models of targeting efficacy (Grimson et al., 2007; Nielsen et al., 2007; Wang and El Naqa, 2008; Betel et al., 2010; Liu et al., 2010; Garcia et al., 2011; Wen et al., 2011; Reczko et al., 2012; Vejnar and Zdobnov, 2012; Marin et al., 2013; Gumienny and Zavolan, 2015). Our recent model, called the context-plus (context+) model, considers the features of our original context scores (i.e., site type, 3′-supplementary pairing, local AU content, and distance from the closest 3′-UTR end [Grimson et al., 2007]), plus two miRNA features (TA and SPS [Garcia et al., 2011]). Although the context+ model was trained using multiple regression on 74 high-throughput datasets, the features used to distinguish effective sites (the three features of the original context scores) were identified using only 11 datasets, implying that additional features might be identified through analysis of the additional datasets. Here, we examined the function of non-canonical binding sites identified in recent studies and found that mRNAs with these sites are not more repressed than mRNAs without sites, despite compelling evidence that many of these noncanocial sites bind the silencing complex in vivo. This finding justified a focus on the statistical modeling of canonical, seed-matched sites within 3′ UTRs, which mediate the vast majority of repression that can be predicted with current methods. To this end, we pre-processed the 74 datasets to minimize confounding biases and then used stepwise regression to identify the most informative features from a large set of potential targeting features. This approach unbiasedly selected 14 features, which were combined to develop the context++ model of miRNA targeting efficacy. The context++ model was more predictive than any published model and at least as predictive as the most informative in vivo crosslinking approaches. As the engine powering the latest version of TargetScan (v7.0; targetscan.org), this model provides a valuable resource for placing the miRNAs of human, mouse, zebrafish, and other vertebrate species into their respective gene-regulatory networks. Results Inefficacy of recently reported non-canonical binding sites Several high-throughput crosslinking-immunoprecipitation (CLIP) approaches have been applied to identify sites that bind Argonaute in vivo (Chi et al., 2009; Hafner et al., 2010; Helwak et al., 2013; Grosswendt et al., 2014). These experiments all observe significant enrichment for cognate seed-matched sites in the vicinity of the crosslinks, which validates their ability to detect authentic sites. Despite this enrichment, some crosslinks do not correspond to canonical sites to the relevant miRNAs, raising the prospect that these results might reveal novel types of non-canonical binding that could mediate repression. Indeed, five studies have reported crosslinking to non-canonical binding sites proposed to mediate repression (Chi et al., 2012; Loeb et al., 2012; Helwak et al., 2013; Khorshid et al., 2013; Grosswendt et al., 2014). In addition, another biochemical study has reported the identification of non-canonical sites without using any crosslinking (Tan et al., 2014). Reasoning that these experimental datasets might provide a resource for defining of novel types of sites to be used in target prediction, we re-examined the functionality of these sites in mediating target mRNA repression. We first examined the efficacy of ‘nucleation-bulge’ sites (Chi et al., 2012), which were identified from analysis of differential CLIP (dCLIP) results reporting the clusters that appear in the presence of miR-124 (Chi et al., 2009). Nucleation-bulge sites consist of 8 nt motifs paired to positions 2–8 of their cognate miRNA seed, with the nucleotide opposing position 6 protruding as a bulge but sharing Watson-Crick complementarity to miRNA position 6. Meta-analyses of miRNA and small-RNA transfection datasets revealed significant repression of mRNAs with the canonical site types but found no evidence for repression of mRNAs that contain nucleation-bulge sites but lack perfectly paired seed-matched sites in their 3′ UTRs (Figure 1—figure supplement 1A,B). Reasoning that the nucleation-bulge site might be only marginally effective, we examined the early zebrafish embryo with and without Dicer, analyzing the targeting by miR-430, the most highly expressed miRNA of the early embryo. Even in this system, one of the most sensitive systems for detecting the effects of targeting (where a robust repression is observed for mRNAs with only a single 6mer or offset-6mer sites to miR-430), we observed no evidence for repression of mRNAs with nucleation-bulge sites to miR-430 (Figure 1A, Figure 1—figure supplement 1C, and Figure 1—figure supplement 4A). Because the nucleation-bulge sites were originally identified and characterized as sites to miR-124, we next tried focusing on only miR-124–mediated repression. However, even in this more limited context, the mRNAs with nucleation-bulge sites were no more repressed than mRNAs without sites (Figure 1—figure supplement 1D–F). Figure 1 with 5 supplements see all Download asset Open asset Inefficacy of recently reported non-canonical sites. (A) Response of mRNAs to the loss of miRNAs, comparing mRNAs that contain either a canonical or nucleation-bulge site to miR-430 to those that do not contain a miR-430 site. Plotted are cumulative distributions of mRNA fold changes observed when comparing embryos that lack miRNAs (MZDicer) to those that have miRNAs (WT), focusing on mRNAs possessing a single site of the indicated type in their 3′ UTR. Similarity of site-containing distributions to the no-site distribution was tested (one-sided Kolmogorov–Smirnov [K–S] test, P values); the number of mRNAs analyzed in each category is listed in parentheses. See also Figure 1—figure supplement 1C and Figure 1—figure supplement 4A. (B and C) Response of mRNAs to the loss of miR-155, focusing on mRNAs that contain either a single canonical or ≥1 CLIP-supported non-canonical site to miR-155. These panels are as in (A), but compare fold changes for mRNAs with the indicated site type following genetic ablation of mir-155 in either T cells (B) or Th1 cells (C). See also Figure 1—figure supplement 2. (D and E) Response of mRNAs to the knockdown of miR-92a, focusing on mRNAs that contain either a single canonical or ≥1 CLASH-identified non-canonical site to miR-92a. These panels are as in (A), except CLASH-supported non-canonical sites were the same as those defined previously (Helwak et al., 2013) and thus were permitted to reside in any region of the mature mRNA, and these panels compare fold changes for mRNAs with the indicated site type following either knockdown of miR-92a (D) or combined knockdown of miR-92a and 24 other miRNAs (E) in HEK293 cells. See also Figure 1—figure supplement 3A,B. (F) As in (D), but focusing on mRNAs that contain ≥1 chimera-identified site. See also Figure 1—figure supplement 3C–E and Figure 1—figure supplement 4B. (G) Response of mRNAs to the transfection of 16 miRNAs, focusing on mRNAs that contain either a canonical or MIRZA-predicted non-canonical site. This panel is as in (A), but compares the fold changes for mRNAs with the indicated site type after introducing miRNAs, aggregating results from 16 individual transfection datasets. Fold changes are plotted for the top 100 non-canonical predictions for each of 16 miRNAs compiled either before (MIRZA, top 100) or after (MIRZA, no 6mers) removing mRNAs containing 6mer or offset-6mer 3′-UTR sites. (H) Response of mRNAs to a transfection of miR-522, focusing on mRNAs that contain either a single canonical or ≥1 IMPACT-seq–supported non-canonical site to miR-522. These panels are as in (A), except IMPACT-seq–supported non-canonical sites were the same as those defined previously (Tan et al., 2014) and thus were permitted in any region of the mature mRNA. (I) Response of ribosomes to the loss of miR-155, focusing on mRNAs that contain either a single canonical or ≥1 CLIP-supported non-canonical site to miR-155. This panel is as in (B and C) but compares the response of mRNAs using ribosome-footprint profiling (Eichhorn et al., 2014) after genetic ablation of mir-155 in B cells. Ribosome-footprint profiling captures changes in both mRNA stability and translational efficiency through the high-throughput sequencing of ribosome-protected mRNA fragments (RPFs). https://doi.org/10.7554/eLife.05005.003 Another study examined the response of 32 mRNAs that lack canonical miR-155 sites yet crosslink to Argonaute in wild-type T cells but not T cells isolated from miR-155 knockout mice (Loeb et al., 2012). As previously observed, we found that the levels of these mRNAs tended to increase in T cells lacking miR-155 (Figure 1B). However, a closer look at the distribution of mRNA fold changes between wild-type and knockout cells revealed a pattern not normally observed for mRNAs with a functional site type. As illustrated for the mRNAs with canonical sites (including those supported by CLIP), when a miRNA is knocked out, the cumulative distribution of fold changes for mRNAs with functional site types diverges most from the no-site distribution at the top of the curve, which represents the most strongly derepressed mRNAs (Figure 1B). However, for the mRNAs harboring non-canonical miR-155 sites, the distribution of fold changes converged with the no-site distribution at the top of the curve (Figure 1B), raising doubt as to whether non-canonical binding of these mRNAs mediates repression. To investigate these mRNAs further, we examined their response to the miR-155 loss in helper T cell subtypes 1 and 2 (Th1 and Th2, respectively) and B cells, which are other lymphocytic cells in which significant derepression of miR-155 targets is observed in cells lacking miR-155 (Rodriguez et al., 2007; Eichhorn et al., 2014). In contrast to mRNAs with canonical sites, the mRNAs with non-canonical sites showed no evidence of derepression in the knockout cells of each of these cell types, which reinforced the conclusion that non-canonical binding of miR-155 does not lead to repression of these mRNAs (Figure 1C and Figure 1—figure supplement 2). We next probed the functionality of non-canonical interactions identified by CLASH (crosslinking, ligation, and sequencing of hybrids), a high-throughput technique that generates miRNA–mRNA chimeras, which each identify a miRNA and the mRNA region that it binds (Helwak et al., 2013). As previously observed, mRNAs with CLASH-identified non-canonical interactions involving miR-92 tended to be slightly up-regulated upon knockdown of miR-92 in HEK293 cells (Figure 1D). However, a closer look at the mRNA fold-change distributions again revealed a pattern not typically observed for mRNAs with a functional site type, with convergence with the no-site distribution in the region expected to be most divergent. Therefore, we examined a second dataset monitoring mRNA changes after knocking down miR-92 and other miRNAs in HEK293 cells (Hafner et al., 2010). As reported recently (Wang, 2014), the slight up-regulation observed for mRNAs with CLASH-identified non-canonical interactions in the original dataset was not reproducible in the second dataset (Figure 1E). Moreover, mRNAs with non-canonical interactions to other miRNAs showed no sign of derepression when the cognate miRNAs were knocked down (Figure 1—figure supplement 3A). To mirror the original analyses of CLASH-identified interactions (Helwak et al., 2013), our analyses included sites located in any region of the mature mRNA (Figure 1D,E and Figure 1—figure supplement 3A). No significant difference from the no-site control distribution was observed when restricting our analysis to mRNAs with CLASH-identified non-canonical sites in their 3′ UTRs (Figure 1—figure supplement 3B). Many miRNA–mRNA chimeras can also be found in standard AGO CLIP datasets, presumably generated by an endogenous ligase acting in cell lysates during workup (Grosswendt et al., 2014). Global experiments examining function of these interactions group the mRNAs with non-canonical interactions together with those with canonical interactions (Grosswendt et al., 2014), and thus the signal for function might arise from only canonical interactions. Indeed, when we re-examined the response of these mRNAs to miRNA knockdown, those with chimera-identified canonical sites tended to be derepressed, whereas those with only chimera-identified non-canonical sites did not (Figure 1F and Figure 1—figure supplement 3C–E). Although at first glance this finding might seem at odds with the elevated evolutionary conservation of chimera-identified non-canonical sites (Grosswendt et al., 2014), we found that this conservation signal was not smaller for the sites of less conserved miRNAs and therefore was not indicative of functional miRNA binding (Figure 1—figure supplement 5). Instead, the reported conservation signal might occur for the same reason that artificial siRNAs tend to target conserved regions of 3′ UTRs (Nielsen et al., 2007). Next, we evaluated the response of non-canonical sites modeled by MIRZA, an algorithm that utilizes CLIP data in conjunction with a biophysical model to predict target sites (Khorshid et al., 2013). As noted by others (Majoros et al., 2013), the definition of non-canonical MIRZA sites was more expansive than that used elsewhere and did not exclude sites with canonical 6mer or offset-6mer seed matches. Indeed, when focusing on only targets without 6mer or offset-6mer seed matches, the top 100 non-canonical MIRZA targets showed no sign of efficacy (Figure 1G). Finally, we examined non-canonical clusters identified by IMPACT-seq (identification of miRNA-responsive elements by pull-down and alignment of captive transcripts—sequencing), a method that sequences mRNA fragments that co-purify with a biotinylated miRNA without crosslinking (Tan et al., 2014). Although the mRNAs with an IMPACT-seq–supported canonical site were down-regulated upon the transfection of the cognate miRNA, those with an IMPACT-seq–supported non-canonical site responded no differently than mRNAs lacking a site (Figure 1H). Collectively, the novel non-canonical sites recently identified in high-throughput CLIP and other biochemical studies imparted no detectable repression when monitoring mRNA changes. However, monitoring of only mRNA changes leaves open the possibility that these sites might still mediate translational repression. To address this possibility, we examined ribosome-profiling and proteomic datasets, which capture repression also occurring at the level of translation, and again we found that the CLIP-identified non-canonical sites imparted no detectable repression (Figure 1I and Figure 1—figure supplement 4). All of our analyses of experimentally identified non-canonical sites examined the ability of the sites to act in mRNAs that had no seed-matched site to the same miRNA in their 3′ UTRs. Any non-canonical site found in a 3′ UTR that also had a seed-matched site to the same miRNA was not considered because any response could be attributed to the canonical site. At first glance, excluding these co-occurring sites might seem to allow for the possibility that the experimentally identified non-canonical sites could contribute to repression when in the same 3′ UTR as a canonical site, even though they are ineffective in 3′ UTRs without canonical sites. However, in mammals, canonical sites to the same miRNA typically act independently (Grimson et al., 2007; Nielsen et al., 2007), and we have no reason to think that non-canonical sites would behave differently. More importantly, although the non-canonical sites examined were in mRNAs that had no seed-matched 3′-UTR site to the same miRNA, most were in mRNAs that had seed-matched 3′-UTR sites to other miRNAs that were highly expressed in the cells. Therefore, even if the non-canonical sites could only function when coupled to a canonical site, we still would have observed a signal for their function in our analyses. Confirmation that miRNAs bind to non-canonical sites despite their inefficacy The inefficacy of recently reported non-canonical sites was surprising when considering evidence that the dCLIP clusters without cognate seed matches are nonetheless enriched for imperfect pairing to the miRNA, which would not be expected if those clusters were merely non-specific background (Chi et al., 2012; Loeb et al., 2012). Indeed, our analysis of motifs within the dCLIP clusters for miR-124 and miR-155 confirmed that those without a canonical site to the miRNA were enriched for miRNA pairing (Figure 2A). Although one of the motifs identified within CLIP clusters that appeared after transfection of miR-124 into HeLa cells yet lacked a canonical miR-124 site did not match the miRNA (Figure 2—figure supplement 1C), the top motif, as identified by MEME (Bailey and Elkan, 1994), had striking complementarity to the miR-124 seed region (Figure 2A). This human miR-124 non-canonical motif matched the ‘nucleation-bulge’ motif originally found for miR-124 in the mouse brain (Chi et al., 2012). Although the top motif identified within the subset of miR-155 dCLIP clusters that lacked a canonical site to miR-155 was not identified with confidence, it had only a single mismatch to the miR-155 seed, which would not have been expected for a motif identified by chance. Figure 2 with 2 supplements see all Download asset Open asset Confirmation of experimentally identified non-canonical miRNA binding sites. (A) Sequence logos corresponding to motifs enriched in dCLIP clusters that either appear following transfection of miR-124 into HeLa cells (Chi et al., 2009) (top) or disappear following knockout of miR-155 in T cells (Loeb et al., 2012) (bottom). Shown to the right of each logo is its E-value among clusters lacking a seed-matched or offset-6mer canonical site and the fraction of these clusters that matched the logo. Shown below each logo are the complementary regions of the cognate miRNA family, highlighting nucleotides 2–8 in capital letters. (B) Position of the top-ranked motif corresponding to non-canonical sites enriched in CLASH (Helwak et al., 2013) (left) or chimera (Grosswendt et al., 2014) (right) data for each human miRNA family supported by at least 50 interactions without a seed-matched or offset-6mer canonical site. For each family the most enriched logo was aligned to the reverse complement of the miRNA. In cases in which a logo mapped to multiple positions along the miRNA, the positions with the best and second best scores are indicated (red and blue, respectively). (C) Sequence logos of motifs enriched in chimera interactions that lack canonical sites. As in (A), but displaying sequence logos identified in the chimera data of panel (B) for a sample of nine human miRNAs. Logos identified for the other human miRNAs are also provided (Figure 2—figure supplement 1B). A nucleotide that differs between miRNA family members is indicated as a black ‘n’. https://doi.org/10.7554/eLife.05005.009 Previous analysis of CLASH-identified interactions shows that the top MEME-identified motifs usually pair to the miRNA, although for many miRNAs this pairing falls outside of the seed region (Helwak et al., 2013). Repeating this analysis, but focusing on only interactions without canonical sites, confirmed this result (Figure 2B).