Pacific Symposium on Biocomputing 13:64-74(2008)

ANALYSIS OF MICRORNA-TARGET INTERACTIONS BY A TARGET STRUCTURE BASED HYBRIDIZATION MODEL
DANG LONG, CHI YU CHAN, YE DING# Wadsworth Center, New York State Department of Health, 150 New Scotland Avenue, Albany, NY 12208 Email: dlong, yding@wadsworth.org, c.clarence@yahoo.com
MicroRNAs (miRNAs) are small non-coding RNAs that repress protein synthesis by binding to target messenger RNAs (mRNAs) in multicellular eukaryotes. The mechanism by which animal miRNAs specifically recognize their targets is not well understood. We recently developed a model for modeling the interaction between a miRNA and a target as a two-step hybridization reaction: nucleation at an accessible target site, followed by hybrid elongation to disrupt local target secondary structure and form the complete miRNA-target duplex. Nucleation potential and hybridization energy are two key energetic characteristics of the model. In this model, the role of target secondary structure on the efficacy of repression by miRNAs is considered, by employing the Sfold program to address the likelihood of a population of structures that co-exist in dynamic equilibrium for a specific mRNA molecule. This model can accurately account for the sensitivity to repression by let-7 of both published and rationally designed mutant forms of the Caenorhabditis elegans lin-41 3 UTR, and for the behavior of many other experimentally-tested miRNA-target interactions in C. elegans and Drosophila melanogaster. The model is particularly effective in accounting for certain false positive predictions obtained by other methods. In this study, we employed this model to analyze a set of miRNA-target interactions that were experimentally tested in mammalian models. These include targets for both mammalian miRNAs and viral miRNAs, and a viral target of a human miRNA. We found that our model can well account for both positive interactions and negative interactions. The model provides a unique explanation for the lack of function of a conserved seed site in the 3 UTR of the viral target, and predicts a strong interaction that cannot be predicted by conservation-based methods. Thus, the findings from this analysis and the previous analysis suggest that target structural accessibility is generally important for miRNA function in a broad class of eukaryotic systems. The model can be combined with other algorithms to improve the specificity of predictions by these algorithms. Because the model does not involve sequence conservation, it is readily applicable to target identification for microRNAs that lack conserved sites, non-conserved human miRNAs, and poorly conserved viral mRNAs. StarMir is a new Sfold application module developed for the implementation of the structure-based model, and is available through Sfold Web server at http://sfold.wadsworth.org.


#

Joint first authors with equal contributions Corresponding author


Pacific Symposium on Biocomputing 13:64-74(2008)

1.

Introduction

MicroRNAs (miRNAs) are endogenous non-coding RNAs (ncRNAs) of ~22 nt, and are among the most abundant regulatory molecules in multicellular organisms. miRNAs typically negatively regulate specific mRNA targets through essentially two mechanisms: 1) when a miRNAs is perfectly or nearly perfectly complementary to mRNA target sites, as is the case for most plant miRNAs, it causes mRNA target cleavage1; and 2) a miRNA with incomplete complementarity to sequences in the 3 untranslated region (3 UTR) of its target (as is the case for most animal miRNAs) can cause translational repression, and/or some degree of mRNA turnover2. miRNAs regulate diverse developmental and physiological processes in animals and plants2-6. Besides animals and plants, miRNAs have also been discovered in viruses7. The targets and functions of plant miRNAs are relatively easy to identify due to the near-perfect complementarity1. By contrast, the incomplete target complementarity typical of animal miRNAs implies a huge regulatory potential, but also presents a challenge for target identification. A number of algorithms have been developed for predicting animal miRNA targets. A common approach relies on a "seed" assumption, wherein the target site is assumed to form strictly Watson-Crick (WC) pairs with bases at positions 2 through 7 or 8 of the 5 end of the miRNA. In the stricter, "conserved seed" formulation of the model, perfect conservation of the 5 seed match in the target is required across multiple species8,9. One well-known exception to the seed model is interaction between let-7 on lin-41, for which G-U pair and unpaired base(s) are present in the seed regions of two binding sites with experimental support10. While the seed model is supported as a basis for identifying many well-conserved miRNA targets11, two studies suggest that G-U or mismatches in the seed region can be well tolerated, and that conserved seed match does not guarantee repression12,13. These suggest that the seed model may represent only a subset of functional target sites, and that additional factors are involved in further defining target specificity at least for some cases with conserved seed matches. Recently, a number of features of site context have been proposed for enhancing targeting specificity14. For posttranscriptional gene modulation by mRNA-targeting nucleic acids, the importance of target structure and accessibility has long been established for antisense oligonucleotides and ribozymes15,16, and evidence for this has also emerged for siRNAs17,18; and more recently for miRNAs19-22. These suggest that target accessibility can be an important parameter for target specificity. We recently developed a model for modeling the interaction between a miRNA and a target as a two-step hybridization reaction: nucleation at an


Pacific Symposium on Biocomputing 13:64-74(2008)

accessible target site, followed by hybrid elongation to disrupt local target secondary structure and form the complete miRNA-target duplex19. Nucleation potential and hybridization energy are two key energetic characteristics of the model. In this model, the role of target secondary structure on the efficacy of repression by miRNAs is taken into account, by employing the Sfold program to address the likelihood of a population of structures that co-exist in dynamic equilibrium for a specific mRNA molecule. This model can accurately account for the sensitivity to repression by let-7 of both published and rationally designed mutant forms of the Caenorhabditis elegans lin-41 3 UTR, and for the behavior of many other experimentally-tested miRNA-target interactions in C. elegans and Drosophila melanogaster. The model is particularly effective in accounting for certain false positive predictions obtained by other methods. In this study, we employed this model to analyze a set of miRNA-target interactions that were experimentally tested in mammalian models. We here report the results of the analysis and discuss implications of the findings. 2. Methods 2.1 mRNA Secondary Structure Prediction The secondary structure of an mRNA molecule can influence the accessibility of that mRNA to a nucleic acid molecule that can bind to the mRNA by complementary base-pairing. Determination of mRNA secondary structure presents theoretical and experimental challenges. One major impediment to the accurate prediction of mRNA structures stems from the likelihood that a particular mRNA may not exist as a single structure, but in a population of structures in thermodynamic equilibrium23-25. Thus, the computational prediction of secondary structure based on free energy minimization is not well suited to the task of providing a realistic representation of mRNA structures. An alternative to free energy minimization for charactering the ensemble of probable structures for a given RNA molecule has been developed26. In this approach, a statistically representative sample is drawn from the Boltzmannweighted ensemble of RNA secondary structures for the RNA. Such samples can faithfully and reproducibly characterize structure ensembles of enormous sizes. In particular, in comparison to energy minimization, this method has been shown to make better structural predictions27 and to better represent the likely population of mRNA structures28, and to yield a significant correlation between predictions and data for gene inhibition by antisense oligos29, gene knockdown by RNAi30 and target cleavage by hammerhead ribozymes (unpublished data), and translational repression by miRNAs19. A sample size of 1,000 structures is sufficient to guarantee statistical reproducibility in sampling statistics and


Pacific Symposium on Biocomputing 13:64-74(2008)

clustering features26,28. The structure sampling method has been implemented in the Sfold software package31 and is used here for mRNA folding. The entire target transcript is used for folding if its length is under 7000 nts. For two targets in this study with transcript lengths over 9000 nt, we only used the UTRs (HCV and THRAP1, Table 1), so the folding could be efficiently managed. 2.2 Two-step Hybridization Model We recently introduced a target-structure based hybridization model for prediction of miRNA-target interaction19. Here, we briefly describe this model and summarize its energetic characteristics. In vitro hybridization studies using antisense oligonucleotides suggested that hybridization of an oligonucleotide to a target RNA requires an accessible local target structure32. This requirement has been supported by various in vivo studies33-35. Such a local structure includes a site of unpaired bases for nucleation, and duplex formation progresses from the nucleation site and stops when it meets an energy barrier. In a kinetic study, it was suggested that the nucleation step is rate-limiting, and that it involves formation of four or five base pairs between the interacting nucleic acids36. Based on these and other related studies37,38, we model the miRNA-target hybridization as a two-step process: 1) nucleation, involving four consecutive complementary nucleotides in the two RNAs (Fig. 1A), and 2) the elongation of the hybrid to form a stable intermolecular duplex (Fig. 1B).

Figure 1. Two-step model of hybridization between a small (partially) complementary nucleic acid molecule and a structured mRNA: 1) nucleation at an accessible site of at least 4 or 5 unpaired bases (A); 2) elongation through "unzipping" of the nearby helix, resulting in altered local target structure (B).

The model is characterized by several energetic parameters. For a given predicted target structure, the nucleation potential, GN, is the stability of the particular single-stranded 4-bp block within the a potential mRNA binding site


Pacific Symposium on Biocomputing 13:64-74(2008)

that would form the most stable 4-bp duplex with the miRNA (In Fig. 1, there are two 4-bp blocks for the 5-bp helix formed between the miRNA and the target). For the sample of 1000 structures predicted by Sfold for the target mRNA, the final GN is the average over the sample. The initiation energy threshold, Ginitiation, is the energy cost for initiation of the interaction between two nucleic acid molecule. For two published values of Ginitiation36,39, 4.09 kcal/ mol appeared to perform somewhat better in our previous study19. Nucleation for a potential site is considered favorable if the nucleation potential can overcome the initiation energy threshold, i.e., GN + Ginitiation < 0 kcal/mol. For a site with favorable nucleation potential, we next compute Gtotal, the total energy change for the hybridization, by Gtotal = Ghybrid - Gdisruption, where Ghybrid is the stability of the miRNA-target hybrid as computed by the RNAhybrid program40, and Gdisruption is the energy cost for the disruption of the local target structure (Fig. 1B), and is computed using structure sample predicted by Sfold for the target mRNA. These calculations have been incorporated into STarMir, a new application model for the Sfold package. To model the cooperative effects of multiple sites on the same 3 UTR for either a single miRNA or multiple miRNAs, we assume energetic additivity and compute Gtotal, where the sum is over multiple sites. 2.3 Dataset of MicroRNA-Target Interactions We tried to assemble a set of high-quality and representative miRNA-target pairs in mammals. We selected reported miRNA-target interactions that were supported by at least two experimental testing using either human cells or mouse or rat models. These interactions play important roles in various biological processes. The targets also include a viral target for a cellular miRNA, and cellular targets for a viral miRNA family. The complete mRNA target sequences were typically retrieved from the Reference Sequence (RefSeq) database from the National Center for Biotechnology Information (NCBI) (http://www.ncbi.nlm.nih.gov/RefSeq). Information for these miRNA-target pairs and the references is given in Table 1. For a few reported interactions in these references, the complete transcripts were not available from the GenBank databases and thus these interactions were not included in this study. 3. Results 3.1 Analysis of Interaction between Mammalian miRNAs and Viral Genomes An intriguing case worthy of particular note is the regulation of Hepatitis C virus (HCV) by miR -12241. In the viral RNA genome, there are a seed site in


Pacific Symposium on Biocomputing 13:64-74(2008)

the 5 non-coding region (NCR) and a seed site in the 3 NCR, both are conserved among the six HCV genotypes. However, the site in the 5 NCR was found to be essential for up-regulation of HCV replication by miR-122, whereas the site in the 3 NCR was not. Current miRNA prediction algorithms that based on seed site conservation, e.g., TargetScan8, PicTar42, cannot explain the lack of function of the 3 NCR seed site. Other algorithms that based only on the alignment and hybridization energy of miRNAs and potential binding sites, e.g., miRanda43, RNAhybrid40, cannot explain the difference between those two sites. We analyzed this miRNA-target pair using our interaction model that takes into account secondary structures of the target sequence. To classify an interaction as functional or nonfunctional, we previously used an empirical threshold of -10.0 kcal/mol for Gtotal19. For this threshold, we predicted a functional interaction between miR-122 and the 5 NCR, but a lack of interaction between miR-122 and the 3 UTR, for which the Gtotal is merely -3.54 kcal/mol. The energetic characteristics for potential binding sites that passed the nucleation threshold are listed below: hsa-miR-122a:HCV 5 NCR interaction Site 1: Target site position in 5 NCR: 21-44 G CUC A AU C Target 21 ACA CACCAU G CACUCC 44 ||| |||||| | |||||| miRNA 23 UGU GUGGUA C GUGAGG 1 UU A AGU U Gtotal = -16.70 kcal/mol; Gdisruption = 6.40 kcal/mol; Ghybrid =-29.10 kcal/mol; GN + Ginitiation = -3.71 kcal/mol. Site 2: Target site position in 5 NCR: 55-70 C CU A Target 55 UACUGU UCACGC 70 |||||| |||||| miRNA 23 GUGGUA AGUGUG 1 UGUUU AC AGGU Gtotal =-9.981 kcal/mol; Gdisruption = 7.619 kcal/mol; Ghybrid =-17.60 kcal/mol; GN + Ginitiation = -2.61 kcal/mol. hsa-miR-122a: HCV 3 NCR Site 1: Target site position in 3 NCR: 9-36 C UG AG GGGUAA G Target 9 CGA A GUUG ACACUCCG ||| | |||| |||||||| miRNA 23 GUU U UAAC UGUGAGGU U UG GG AG

36 1


Pacific Symposium on Biocomputing 13:64-74(2008)

Gtotal =-3.538 kcal/mol; Gdisruption =20.262 kcal/mol; Ghybrid =-23.80 kcal/mol; GN + Ginitiation = -3.71 kcal/mol. The result here suggests that the lack of function for some (conserved) seed sites can be explained by poor target accessibility. In addition, for each of two single-substitution mutations (p3, p6) and a double-substitution mutation (p3-4) of the proposed seed region in the 5 NCR41, the HCV RNAs failed to accumulate. Our predictions for the mutants are consistent with the experimental finding, with Gtotal of -2.057 kcal/mol, -2.013 kcal/mol, and -1.934 kcal/mol, respectively. We note that the more energetically favorable site 1 in the 5 NCR predicted by our model has some overlap with but is substantially different from the published binding site. This suggests an alternative binding conformation for further testing. 3.2 Analysis of Other MicroRNA-Target Interactions We next analyzed 18 other validated interactions listed in Table 1. Our model accounted for 16 of the 18 (thus 17 for 19 including HCV 5 NCR, a sensitivity of 89.5%) positive interactions. Among the two positive cases unaccounted for by our model, the interaction between miR-133a and HCN4 has a Gtotal of -9.5 kcal/mol, which is close to the threshold, and thus could be effective for miRNA-target hybridization. Moreover, the sum of this energy and that for the interaction between miR-1 and HCN4 is -20.304 kcal/mol, which is consistent with the combined effect by miR-133a and miR-1 on HCN4 that was reported44. Because miR-200c is not conserved across five vertebrate genomes, no target prediction can be made by TargetScan8. The regulation of HMGA2 by the let-7 family (all family members sharing the same seed sequence) has been reported by two studies, with let-7a used in one study45, and let-7b, let-7d used in the other46. Data from both studies suggested functionality of multiple target sites identified by conserved seed matches. The rather large value of Gtotal for the interaction between HMGA2 and any of three tested let-7 members is consistent with the understanding that a target can be efficiently regulated through multiple sites for the same miRNA. While convincingly validated mammalian miRNA targets are limited, the functions of viral miRNAs are even less understood. Recently the regulation of several cellular targets by the KSHV-encoded miRNAs has been reported47. We found that our model supports the cooperativity of multiple miRNAs acting on the same target. In particular, for the well-validated target, THBS1, the Gtotal is rather large, a results of many binding sites on this target 3 UTR. The results for both let-7 and KSHV miRNAs suggest that Gtotal presents a promising


Pacific Symposium on Biocomputing 13:64-74(2008)

measure for modeling the additive effects of multiple binding sites by either single or multiple mammalian or viral miRNAs.
Table 1. Target prediction based on miRNA-target interaction energy computed by Gtotal and local AU content of 19 positive interactions and four negative interactions (shaded) Gtotal Local AU Target name, GenBank Accession, miRNA (kcal/mol) and content b and test system a and references the prediction the prediction miR-200c TCF8 AL831979 H48,49 -109.979 + 0.742 + miR-133 RhoA NM_016802 M50 -17.098 + NA c miR-133 Cdc42 NM_009861 M50 -29.939 + 0.432  miR-133 WHSC2 NM_011914 M50 -18.240 + 0.461  miR-1 GJA1 NM_012567 R51 -55.805 + 0.626 + let-7a HMGA2 NM_003483 H45 -185.195 + 0.628 + let-7b HMGA2 NM_003483 H46 -300.446 + 0.628 + let-7e HMGA2 NM_003483 H46 -218.989 + 0.628 + miR-124a Foxa2 NM_010446 M52 -7.662  0.672 + miR-1 HCN4 NM_021658 R44 -10.798 + NA c miR-133a HCN4 NM_021658 R44 -9.506  NA c 10 KSHV 47 -75.231 + NAd SPP1 NM_000582 H miRNAs 10 KSHV -113.169 + NAd SRGN NM_002727 H47 miRNAs 10 KSHV -325.078 + NAd THBS1 NM_003246 H47 miRNAs NM_001025 miR-155 C-Maf -78.858 + 0.710 + M53 577 miR-208 THRAP1 NM_005121 H5 -45.032 + 0.895 + miR-375 Mtpn NM_145808 H54 -13.084 + 0.715 + miR-122 CAT-1 NM_013111 R55 -136.82 + 0.530  miR-122 HCV 5 NCR NC_004102 H41 -26.681 + 0.478  miR-122 HCV 3 NCR NC_004102 H41 -3.538  0.412  miR-122 HCV 5 NCR p3 NC_004102 H41 -2.057  NA c 41 c miR-122 HCV 5 NCR p6 NC_004102 H -2.013  NA miR-122 HCV 5 NCR p3-4 NC_004102 H41 -1.934  NA c a H: human cells; M: mouse; R: rat. b as defined in Grimson et al., 200714; c no perfect (7- or 8mer) seed sites; d not calculated due to multiple miRNAs; + : predicted effective target,  : predicted ineffective target

We also calculated local AU content of seed sites of the miRNAs and targets following a scoring scheme proposed by Grimson et al.14. When there are multiple seed sites in the same 3 UTR sequence, we report the best local AU content (Table 1). In order to correlate the local AU content to the qualitative information of miRNA activity in our dataset, we select a threshold of 0.6 for the local AU content. miRNA-target pairs having the local AU content is higher or equal 0.6 are predicted functional. This threshold is partly based on the experimental data in Grimson et al.14, where the local AU content of 0.6 correlated to the average fold change of 0.89 in the mRNA level from the


Pacific Symposium on Biocomputing 13:64-74(2008)

microarray experiment. The AU content of 0.6 is also just above the mean AU content of all possible 7-mer sites of the 3 UTR sequences being considered here (data not shown). For this threshold, the local AU content alone can explain the positive interactions for 9 of the 13 miRNA-target pairs. For each of these 13 pairs, there is at least one seed site and only the concerned miRNA is known to be involved in regulation of the target. In comparison, we predict effective interactions for 12 of the 13 cases (Table 1). Furthermore, both of the two conserved seed sites for miRNA-122 in HCV 5 NCR and 3 NCR have comparable low AU content (Table 1). Therefore, the local AU content cannot explain the functional difference between the two seed sites. 4. Conclusion In this study, we employed a recently developed target-structure based hybridization model to analyze a set of miRNA-target interactions. These interactions were experimentally tested in human cells or in animal models (mouse or rat). These include mammalian targets for both cellular miRNAs and viral RNAs, and a viral target for a cellular miRNA. Our model can well account for positive interactions, as well as negative interactions. In particular, the model can explain the difference in the interactions of miR-122 to HCV 5 NCR and HCV3 NCR, which could not be explained by several popular miRNA target prediction programs. In our previous analysis of repression data for worm and fly19, we observed that the model can not only uniquely account for interactions between let-7 and worm lin-41 mutants that cannot be explained by other algorithms, but also explain negative experimental results for 11 of 12 targets with seed matches for lsy-6. These and the findings of this analysis here suggest that target structural accessibility is generally important for miRNA function in a broad class of eukaryotic systems, and that the model can be combined with other algorithms to improve the specificity of predictions by these algorithms. Our comparison of the predictions based on the interaction energies and the ones based on the local AU content suggests that the local AU content does not reflect accurately target sites' accessibility in many cases. Therefore, the interaction model considered here can more accurately account for miRNA activities. Because the model does not involve sequence conservation, it can be particularly valuable for target identification for microRNAs that lack conserved sites56, non- or poorly-conserved human miRNAs57 (e.g., the lack of prediction by TargetScan for miR-200c), and usually poorly conserved viral mRNAs.


Pacific Symposium on Biocomputing 13:64-74(2008)

Acknowledgments The Computational Molecular Biology and Statistics Core at the Wadsworth Center is acknowledged for providing computing resources for this work. This work was supported in part by National Science Foundation grants DMS0200970, DBI-0650991, and National Institutes of Health grant GM068726 (Y.D.). References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. Rhoades, M.W. et al. Cell 110, 513-20 (2002). Ambros, V. Nature 431, 350-5 (2004). Boehm, M. & Slack, F. Science 310, 1954-7 (2005). Dugas, D.V. & Bartel, B. Curr Opin Plant Biol 7, 512-20 (2004). van Rooij, E. et al. Science 316, 575-9 (2007). Calin, G.A. et al. N Engl J Med 353, 1793-801 (2005). Cullen, B.R. Nat Genet 38 Suppl, S25-30 (2006). Lewis, B.P., Burge, C.B. & Bartel, D.P. Cell 120, 15-20 (2005). Lewis, B.P., Shih, I.H., Jones-Rhoades, M.W., Bartel, D.P. & Burge, C.B. Cell 115, 787-98 (2003). Vella, M.C., Choi, E.Y., Lin, S.Y., Reinert, K. & Slack, F.J. Genes Dev 18, 132-7 (2004). Rajewsky, N. Nat Genet 38 Suppl, S8-13 (2006). Didiano, D. & Hobert, O. Nat Struct Mol Biol 13, 849-51 (2006). Miranda, K.C. et al. Cell 126, 1203-17 (2006). Grimson, A. et al. Mol Cell 27, 91-105 (2007). Vickers, T.A., Wyatt, J.R. & Freier, S.M. Nucleic Acids Res 28, 1340-7 (2000). Zhao, J.J. & Lemke, G. Mol Cell Neurosci 11, 92-7 (1998). Overhoff, M. et al. J Mol Biol 348, 871-81 (2005). Schubert, S., Grunweller, A., Erdmann, V.A. & Kurreck, J. J Mol Biol 348, 883-93 (2005). Long, D. et al. Nat Struct Mol Biol 14, 287-294 (2007). Zhao, Y. et al. Cell 129, 303-17 (2007). Zhao, Y., Samal, E. & Srivastava, D. Nature 436, 214-20 (2005). Robins, H., Li, Y. & Padgett, R.W. Proc Natl Acad Sci U S A 102, 4006-9 (2005). Christoffersen, R.E., McSwiggen, J.A. & Konings, D. J. Mol. Structure (Theochem) 311, 273-284 (1994). Altuvia, S., Kornitzer, D., Teff, D. & Oppenheim, A.B. J Mol Biol 210, 265-80 (1989). Betts, L. & Spremulli, L.L. J Biol Chem 269, 26456-63 (1994). Ding, Y. & Lawrence, C.E. Nucleic Acids Res 31, 7280-301 (2003). Ding, Y., Chan, C.Y. & Lawrence, C.E. RNA 11, 1157-66 (2005).


Pacific Symposium on Biocomputing 13:64-74(2008)

28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54. 55. 56. 57.

Ding, Y., Chan, C.Y. & Lawrence, C.E. J Mol Biol 359, 554-71 (2006). Ding, Y. & Lawrence, C.E. Nucleic Acids Res 29, 1034-46 (2001). Shao, Y. et al. RNA (2007). Ding, Y., Chan, C.Y. & Lawrence, C.E. Nucleic Acids Res 32, W135-41 (2004). Milner, N., Mir, K.U. & Southern, E.M. Nat Biotechnol 15, 537-41 (1997). Darnell, J.C. et al. Genes Dev 19, 903-18 (2005). Friebe, P., Boudet, J., Simorre, J.P. & Bartenschlager, R. J Virol 79, 380-92 (2005). Mikkelsen, J.G., Lund, A.H., Duch, M. & Pedersen, F.S. J Virol 74, 600-10 (2000). Hargittai, M.R., Gorelick, R.J., Rouzina, I. & Musier-Forsyth, K. J Mol Biol 337, 951-68 (2004). Paillart, J.C., Skripkin, E., Ehresmann, B., Ehresmann, C. & Marquet, R. Proc Natl Acad Sci U S A 93, 5572-7 (1996). Reynaldo, L.P., Vologodskii, A.V., Neri, B.P. & Lyamichev, V.I. J Mol Biol 297, 511-20 (2000). Xia, T. et al. Biochemistry 37, 14719-35 (1998). Rehmsmeier, M., Steffen, P., Hochsmann, M. & Giegerich, R. RNA 10, 1507-17 (2004). Jopling, C.L., Yi, M., Lancaster, A.M., Lemon, S.M. & Sarnow, P. Science 309, 1577-81 (2005). Krek, A. et al. Nat Genet 37, 495-500 (2005). Enright, A.J. et al. Genome Biol 5, R1 (2003). Xiao, J. et al. J Cell Physiol 212, 285-92 (2007). Mayr, C., Hemann, M.T. & Bartel, D.P. Science 315, 1576-9 (2007). Lee, Y.S. & Dutta, A. Genes Dev 21, 1025-30 (2007). Samols, M.A. et al. PLoS Pathog 3, e65 (2007). Hurteau, G.J., Carlson, J.A., Spivack, S.D. & Brock, G.J. Cancer Res 67, 7972-6 (2007). Hurteau, G.J., Spivack, S.D. & Brock, G.J. Cell Cycle 5, 1951-6 (2006). Care, A. et al. Nat Med 13, 613-8 (2007). Yang, B. et al. Nat Med 13, 486-91 (2007). Baroukh, N. et al. J Biol Chem 282, 19575-88 (2007). Rodriguez, A. et al. Science 316, 608-11 (2007). Poy, M.N. et al. Nature 432, 226-30 (2004). Jopling, C.L., Norman, K.L. & Sarnow, P. Cold Spring Harb Symp Quant Biol 71, 369-76 (2006). Farh, K.K. et al. Science 310, 1817-21 (2005). Bentwich, I. et al. Nat Genet 37, 766-70 (2005).