Imported: 17 Feb '17 | Published: 23 Sep '14
USPTO - Utility Patents
The present invention relates to a method for identifying a binding site on an RNA transcript, wherein the binding site binds to one or more binding moieties. The method includes, among other things, introducing a photoreactive nucleoside into living cells wherein the living cells incorporate the photoreactive nucleoside into RNA transcripts during transcription thereby producing modified RNA transcripts; reverse transcribing the RNA of isolated cross-linked segments thereby generating cDNA transcripts with one mutation wherein the photoreactive nucleoside is transcribed to a mismatched deoxynucleoside; amplifying the cDNA transcripts thereby generating amplicons; and analyzing the sequences of the amplicons aligned against the reference sequence so as to identify the binding site, wherein the sequences of each amplicon having a mutation resulting from the introduction of the photoreactive nucleoside is considered to be a valid amplicon comprising at least a portion of a binding site on the RNA transcript.
This application asserts priority of U.S. Provisional Application Ser. No. 61/137,265 filed on Jul. 28, 2008. The specification of U.S. Provisional Application Ser. No. 61/137,265 is hereby incorporated by reference in its entirety.
Understanding global gene expression at the level of the whole cell requires detailed knowledge of the contributions of transcription, pre-mRNA processing, mRNA turnover, and translation. Although the sum total of these regulatory processes in each cell accounts for its unique expression profile, few methods are available to independently assess each process en masse. DNA arrays are well suited for profiling the steady-state levels of mRNA globally (i.e., the transcriptome). However, because of posttranscriptional events affecting mRNA stability and translation, the expression levels of many cellular proteins do not directly correlate with steady-state levels of mRNAs.
RNA binding proteins (RBPs) and ribonucleoprotein complexes (RNPs), such as microRNA-containing RNPs, are essential regulators of virtually all cellular activities, ranging from development, metabolism and migration to reaction to cellular stress. These proteins do so by binding to coding and non-coding RNAs at specific regions on an RNA transcript. The proteins regulate the rate of transcription, modification, splicing, nuclear export, transport, stability and translation. RNA binding proteins and RNPs recognize canonical binding motifs on a given transcript and cooperate and compete with other RBPs and RNPs in controlling its fate or metabolic rate.
A number of diseases are associated with, or caused by, deregulation or mutations in these proteins. Notable examples among autoimmune disease include systemic lupus erythematosis, primary biliary cirrhosis (PBC) and Sjogren's syndrome, and among neurologic disease include the paraneoplastic neurologic antigens Nova and Hu, and the Fragile X mental retardation FMR1 protein, the spinal muscular atrophy SMN protein, the myotonic dystrophy CELF proteins, and the spinocerebellar ataxia SCA1 protein.
Understanding the role RBPs and RNPs play in disease and normal biology, particularly in the brain, requires methods to identify the set of RNAs to which the RBPs and RNPs bind in vivo. Identifying binding motifs on the RNAs offer ways for targeted therapy. However, the targets of RBPs and RNPs involved in normal and abhorrent cellular processes and systems, including disease states such as autoimmune and genetic diseases have been difficult to identify.
Accordingly, the present invention provides methods for identifying binding sites on RNA transcripts that interact with RBPs and RNPs.
The inventors have surprisingly discovered a powerful method to identify, on a transcriptome-wide level, the target sites of any given RNA-binding protein (RBP) or ribonucleoprotein complex (RNP) with nucleotide resolution. The discovery is based, in part, on using a photoreactive nucleoside that is incorporated into nascent mRNA and is effectively crosslinked to the interacting moieties at long-range UV light in vivo.
The inventors discovered that the photoreactive nucleoside undergoes a structural change upon crosslinking, and is subsequently identified as a mutation in cDNA that is prepared from the modified mRNA. The mutated cDNA can then be analyzed by exploiting the mutation, thereby providing a means of distinguishing UV-crosslinked target sites from background RNA fragments that were captured but not initially crosslinked to the moiety. Such an analysis dramatically increases the recovery of target sites that were crosslinked, reduces the risk of scoring false positives of target sites, and allows for extraction of sequence information of the target site. The inventive methods described herein are also useful for determining the regulation of interactions on the target sites and for binding assays using recombinantly expressed proteins.
In one embodiment, the present invention includes methods for identifying RNA transcripts that associate with a given binding moiety, and also for identifying the region(s) on RNA transcripts which recognize, interact and/or bind to binding moieties. Such a region on an RNA transcript is termed an “RNA protein binding site.” The present invention particularly includes methods for identifying the region(s) on RNA transcripts which recognize, interact and/or bind to binding moieties in vivo. In this specification, “in vivo” refers to a living cultured cell as well as cells in intact organs and intact animals.
As used herein the term “binding moiety” refers to any substantially protein entity that binds to an “RNA protein binding site.” Examples of binding moieties include, but are not limited to, proteins, protein complexes, or portions or fragments thereof, including protein domains, regions, sections and the like. Proteins include one or more RNA-binding proteins (RBP), RNA-associated proteins or combinations thereof. In addition to protein, a protein complex may comprise, for example, nucleic acid components in ribonucleoprotein complexes (RNP), e.g., miRNA, piRNA, siRNA, endo-siRNA, snoRNA, snRNA, tRNA, rRNA or combinations thereof. In RNP complexes, RNA guides and participates in target RNA binding. Binding moieties also include RNA helicases, e.g. MOV10, and proteins containing nuclease motifs, e.g. SND1.
As used herein, the term “RNA protein binding site” or “binding site” refers to that portion, region, position or location on an RNA transcript in which at least one interaction with a binding moiety occurs. Such interaction may include at least one direct base-pairing between the RNA transcript and the binding moiety. A binding site or sites of an RNA transcript may be found at a structured or unstructured region of the RNA transcript. It is also contemplated that more than one binding site may exist for any one RNA transcript. Further, binding sites of RNA transcripts may involve noncontiguous nucleotides of the RNA transcript. Such binding sites are contemplated when structure, such as, for example, a stem loop, is involved in binding.
The method includes introducing a photoreactive nucleoside into living cells wherein the living cells incorporate the photoreactive nucleoside into RNA transcripts during transcription. The RNA transcripts, into which a photoreactive nucleoside is incorporated, are termed “modified RNA transcripts.”
A “photoreactive nucleoside” refers to a modified nucleoside that contains a photochromophore and is capable of photocrosslinking with a binding moiety. Preferably, the photoreactive group will absorb light in a spectrum of the wavelength that is not absorbed by the binding moiety or the non-modified portions of the RNA.
In one embodiment, the photoreactive nucleoside is a thiouridine analog. Thiouridine analogs include 2-thiouridine; 4-thiouridine; and 2,4-di-thiouridine. The thiouridine analogs can optionally be substituted at the 5 and/or 6 position. The substituents can be, for example, methyl, ethyl, halo, nitro, NR1R2 and OR3. R1, R2 and R3 independently represent hydrogen, methyl or ethyl.
In another embodiment, the photoreactive nucleoside is a thioguanosine analog. Thioguanosine analogs include 6-thioguanosine.
In yet another embodiment, the photoreactive nucleoside includes a detectable “label” covalently attached to the nucleoside in order to facilitate use of detection systems, such as luminescence-based systems. Examples of labeled nucleosides include nucleosides modified with biotin and nucleosides modified with suitable haptens such as digoxigenin. Preferred nucleoside analogs for incorporation of a detectable label into RNA include biotin-16-UTP (biotin-16-uridine-5′triphosphate, Roche) and digoxigenin-11-dUTP (an analog of dTTP, 2′-deoxyuridine-5′-triphosphate, coupled to digoxigenin via an 11-atom spacer arm). Fluorescein, Cy3, and Cy5 can be linked to dUTP for direct labeling. Cy3.5 and Cy7 are available as avidin or anti-digoxigenin conjugates for secondary detection of biotin- or digoxigenin-labelled probes. Suitable fluorescence-labeled nucleosides include fluorescein-isothiocyanate-dUTP, cyanine-3-dUTP, and cyanine-5-dUTP.
Accordingly, in one embodiment, the modified nucleoside can be detected by fluorescence (Fluorescein, Rhodamin-green or Cy5). In another example, the nucleoside is detected by an antibody reaction (Digoxigenin, Fluorescein). In yet another example, the nucleoside can be detected by the specific interaction with Streptavidin (Biotin) and a nucleoside that carries a reactive group which can be linked to a label chemically (Aminopentinyl-C7-deaza-dATP).
Any label that can be attached to the nucleoside are suitable. Such labels include radioactive labels, enzymes (that need a reaction with a substrate to be detected), specific binding pair components (such as avidin, streptavidin, and/or biotin), biocytin, iminobiotin, colloidal dye substances, fluorochromes (rhodamin, etc.), reducing substances (eosin, erythrosine, etc.), digoxigenin, metals (ruthenium), metal sols or other particulate sols (selenium, carbon, etc.), dansyl lysine, infra red dyes, coumarines (amino methyl coumarine), antibodies, protein A, protein G, etc. Preferably, the label is biotin, avidin, streptavidin, digoxigenin, or a functional equivalent thereof.
As referred to herein, the “living cell or cells” may be part of a cell culture, a cell extract, cell line, whole tissue, a whole organ, tissue extract, or tissue sample, such as, for example, a biopsy or progenitor cells as from bone marrow or stem cells. The living cell can be from a healthy source or from a diseased source, such as, for example, a tumor, a tumor cell, a cell mass, diseased tissue, tumor cell extract, a pre-cancerous lesion, polyp, or cyst or taken from fluids of such sources. The cells can be any kind of cells, for example, cells from bacteria and yeast, animals, especially mammalian cells, and plants.
Once RNA transcripts have been produced, or at a time at which transcription should have produced transcripts within the living cell or cells, the living cell or cells comprising the modified RNA transcripts are then irradiated. The irradiation is at a wavelength which is significantly absorbed by the photoreactive nucleoside such that covalent cross-links are formed between the modified RNA transcript and a binding moiety and the RNA is not damaged. The minimum wavelength can be 300 nm, preferably 320 nm, and more preferably 340 nm. The maximum wavelength can be 410 nm, preferably 390 nm, and more preferably 380 nm. Any combination of minimum and maximum wavelength values can be used to describe a suitable range. The optimal wavelength is approximately 365 nm for a thiouridine analog. The optimal wavelength for a thioguanosine analog is approximately 310 nm.
Irradiation forms covalent cross-links between the modified RNA transcript and a binding moiety spatially located close enough to said modified RNA transcript to undergo cross-linking. The part or parts of a modified RNA transcript which are in close enough contact to have undergone cross-linking with a binding moiety can be considered binding sites. Thus, binding sites are covalently cross-linked to binding moieties. (For example, see FIG. 1.)
Covalent cross-linking allows the use, in some embodiments of the present invention, of rigorous purification schemes, such as, for example, immuno-precipitation and separating complexes on SDS-PAGE. In some embodiments, the covalent bond enables partial cleavage of RNA molecules without affecting their protein binding by the use of nucleases.
Next, modified RNA transcripts, or portions thereof, which are not covalently cross-linked upon irradiation to one or more binding moieties are removed. The resulting constructs are termed “cross-linked segments.” These “cross-linked segments” include the portion of the modified transcript that comprises the binding site as well as at least the portion of the binding moiety that was subject to cross linking. The binding site therefore contains at least one photoreactive nucleoside through which the binding site is cross-linked to the binding moiety. The “cross-linked segments” also may include additional nucleotides of the modified RNA transcript that are not bound to the binding moiety. These portions are termed “flanking segments.”
In one embodiment, the “cross-linked segments” are formed by harvesting and lysing the cells to form a soluble extract of the cells; and treating the extract with a nuclease. The nuclease trims the regions of the modified transcripts that are not cross-linked to binding moieties. It is contemplated, in one embodiment, that the nuclease would remove, or trim, all of the portion of a modified transcript that is not cross-linked to a binding moiety. However, since trimming can occur in various places on a modified RNA transcript which are not cross-linked to binding moieties, the population of “cross-linked segments” may include “cross-linked segments” with various species of “flanking segments.”
Preferably, the nuclease is ribonuclease-T1 (Aspergillus). Ribonuclease-T1 cleaves the modified RNA transcript at 3′ phosphates of guanine residues, producing terminal guanosine 3′ phosphates. The characteristic of cleaving the modified RNA transcript at 3′ phosphates of guanine residues assists in identifying valid binding sites from noise. This characteristic is especially important for RNA segments below a size of 20-nt, when the frequency for mapping shorter RNA segments to multiple genomic location increases dramatically. It is contemplated that other nucleases may also be employed either alone or in combination with Ribonuclease T1, as well as any other method known in the art that is suited to digest protected RNAs, for example, such as RNase A (which may be isolated from bovine pancreas or from pancreas of other mammalian species, for example); RNase I (which may be isolated from E. coli); or RNase T2 (which is the broad specificity RNase from Aeromonas).
Next, the cross-linked segments are isolated to generate “isolated cross-linked segments.” In one embodiment, isolation is effected by immunoprecipitation (“IP”). Other suitable isolation means with which immunoprecipitation may be combined with or substituted by include use of SDS-PAGE, protein tags, gel filtration, sucrose gradients, etc., with a caveat that the purification scheme preferably does not include conditions that would damage nucleic acids (such conditions as alkaline hydrolysis of RNA).
In IP, a biological sample (e.g., the soluble extract) is contacted with a specific binding molecule, e.g., a molecule that interacts specifically with the part of the “cross-linked segments” and attaches or adheres the molecules to a substrate. The examples of specific binding molecules include antibodies and fragments of antibodies that comprise at least one hypervariable region.
IP is usually carried out with the specific binding molecule on a solid support. Solid supports are known in the art. Examples include any type of solid support, such as any type of bead, plate, column, fiber or array. The molecule that specifically interacts with the “cross-linked segments” may be attached, in one embodiment, to the substrate using any known method, including chemical or physical attachment in some embodiments, as known in the art. Examples of IP methods include protein A/sepharose beads, protein G/sepharose beads, and magnetic beads such as Dynabeads. Techniques for performing IP are known to those skilled in the art (see, for example, Current Protocols in Molecular Biology (1998) Ausubel, et al, eds.).
In one embodiment, the IP is effected by epitope-tagging the binding moiety. A variety of epitopes may be used. Such epitopes may be naturally-occurring amino acid sequences found in nature, artificially constructed sequences, or modified natural sequences. In one embodiment, an artificial epitope sequence with the eight amino acid FLAG marker peptide (Asp-Tyr-Lys-Asp-Asp-Asp-Asp-Lys) (SEQ ID NO: 1) can be used with antibodies recognizing the epitope readily available (Brewer et al., Bioprocess Technol. 2:239-266 (1991); Kunz et al., J. Biol. Chem. 267:9101-9106 (1991)).
Additional artificial epitope tags include an improved FLAG tag having the sequence Asp-Tyr-Lys-Asp-Glu-Asp-Asp-Lys (SEQ ID NO: 2), a nine amino acid peptide sequence Ala-Trp-Arg-His-Pro-Gln-Phe-Gly-Gly (SEQ ID NO: 3) referred to as the “Strep tag” (Schmidt et al, J. Chromatography 676:337-345), poly-histidine sequences, e.g., a poly-His of six residues which is sufficient for binding to IMAC beads, an eleven amino acid sequence from human c-myc recognized by monoclonal antibody 9E10, or an epitope represented by the sequence Tyr-Pro-Tyr-Asp-Val-Pro-Asp-Tyr-Ala-Ile-Glu-Gly-Arg (SEQ ID NO: 4) derived from an influenza virus hemagglutinin (HA) subtype, recognized by the monoclonal antibody 12CA5. Also, the Glu-Glu-Phe sequence recognized by the anti-alpha-tubulin monoclonal antibody YL1/2 has been used as an affinity tag for purification of recombinant proteins (Stammers et al., FEBS Lett. 283:298-302 (1991)).
Another commonly used artificial epitope is a poly-His sequence having six histidine residues (His-His-His-His-His-His) (SEQ ID NO: 5). Naturally occurring epitopes include the eleven amino acid sequence from human c-myc recognized by the monoclonal antibody 9E10 (Glu-Gln-Lys-Leu-Leu-Ser-Glu-Glu-Asp-Leu-Asn) (SEQ ID NO: 6) (Manstein et al. (1995) Gene 162:129-134). Another useful epitope is the tripeptide Glu-Glu-Phe which is recognized by the monoclonal antibody YL 1/2 against alpha-tubulin.
Preferably, magnetic beads, such as Dynabeads, are used as the substrate. The beads can be easily collected by a magnet. Preferably, precipitate, i.e., the isolated “cross-linked segments,” are washed.
As before isolation, the population of “isolated cross-linked segments” comprises “cross-linked segments” of various lengths having differing portions of modified RNA transcripts which are not cross-linked to binding moieties. In one embodiment, the “cross-linked segments” may optionally be treated again with a nuclease to further trim the regions of the modified transcripts that are not cross-linked to binding moieties. In one embodiment, the nuclease would cut off any remaining piece of a modified transcript that is not cross-linked to a binding moiety. However, typically, a population of “isolated cross-linked segments” of various lengths remains.
In one embodiment, the “isolated cross-linked segments” can be separated on the basis of length from other species in a sample. For example, the “isolated cross-linked segments” can be radiolabeled on the 5′ end of the modified transcript and separated by SDS-PAGE gel electrophoresis, and/or further separated by Western blotting. Such separation means are useful for separating the isolated cross-linked segments from isolated segments that were not crosslinked to the binding moiety. The use of these separation techniques is known to those skilled in the art, and is described in, for example, Methods in Enzymology: Guide to Molecular Cloning Techniques (1987) Berger and Kimmel, eds. A protein of optimal length is then eluted from the gel. The optimal length is that length of a known binding moiety.
In a preferred embodiment, the binding moieties are removed from the “isolated cross-linked segments” to generate “isolated segments.” The protein components of the binding moieties are removed by digesting the binding moieties with a protease. Preferably, digestion is effected by Proteinase K or a homologous enzyme. Proteinase K is capable of efficiently digesting protein binding moieties, liberating RNA and yielding products that can be used for ligation and amplification.
Other examples of classes of proteases or their homologues include: Aspartyl proteases, caspases, thiol proteases, Insulinase family proteases, zinc binding proteases, Cytosol Aminopeptidase family proteases, Zinc carboxypeptidases Neutral Zinc Metallopeptidases, extracellular matrix metalloproteinases, matrixins, Prolyl oligopeptidases, Aminopeptidases, Proline Dipeptidases, Methionine aminopeptidases, Serine Carboxypeptidases, Cathepsins, Subtilases, Proteasome A-type Proteases, Proteosome B-type Proteases, Trypsin Family Serine Proteases, Subtilase Family Serine Proteases, Peptidases, and Ubiquitin carboxyl-terminal hydrolases.
The “isolated cross-linked segments” and/or the “isolated segments” are then reverse transcribed to generate cDNA transcripts. Note that although it is preferred to remove the binding moiety before reverse transcription (i.e., to reverse transcribe the isolated segments), it is also possible to reverse transcribe the isolated cross-linked segments (i.e., the segments to which a whole or partial binding moiety is attached).
The introduction of the photoreactive nucleoside yields a mutation in the cDNA transcript when the isolated crosslinked segment is reverse transcribed. For example, the thiouridine analog is reverse transcribed to a deoxyguanosine instead of the deoxyadenosine that is normally incorporated into the reverse transcribed cDNA by Watson-Crick base pairing.
The thioguanosine analog is reverse transcribed to a deoxythymidine instead of the deoxycytidine normally incorporated by Watson-Crick base-pairing. Therefore, the mutation within the cDNA transcript is located within a binding site.
The cDNA transcripts are then amplified, thereby generating cDNA amplicons. When the thiouridine analog is reverse transcribed to produce the mutation of a deoxyguanosine instead of the deoxyadenosine, as described above, the respective cDNA transcripts, when amplified, will include a mutation wherein the expected deoxythymidine is replaced with a deoxycytidine in the amplicons.
When the thioguanosine analog is reverse transcribed to produce the mutation of a deoxythymidine instead of the deoxycytidine, as described above, the respective cDNA transcripts, when amplified, will include a mutation wherein the expected deoxyguanosine is replaced by a deoxyadenosine in the amplicons.
The reverse transcription and amplification can be performed by methods known in the art. For example, the reverse transcription to generate cDNA transcripts and amplification can be achieved using linker ligation and RT-PCR thereby generating amplified cDNA transcripts.
In one embodiment, to prepare cDNA from the “isolated cross-linked segments” and/or the “isolated segments” (i.e., the isolated small RNAs), first synthetic oligonucleotide adapters of known sequence are ligated to the 3′ and 5′ ends of the small RNA pool using T4 RNA ligases. The adapters introduce primer-binding sites for reverse transcription and PCR amplification.
Along with the “isolated cross-linked segments” and/or the “isolated segments,” the small RNA pool typically comprises contaminants resulting from the nuclease digests of very abundant transcripts and non-coding RNAs such as ribosomal RNAs. If desired, non-palindromic restriction sites present within the adapter/primer sequences can be used for generation of concatamers to increase the read length for conventional sequencing or longer size range 454 sequencing.
As will be appreciated by those in the art, the attachment, or joining, of the adapter sequence to the “isolated cross-linked segments” and/or the “isolated segments” can be done in a variety of ways. For example, the adapter sequence can be attached either on the 3′ or 5′ ends, or in an internal position of “isolated cross-linked segments” and/or the “isolated segments.”
In one embodiment, precautions can be taken to prevent circularization of 5′ phosphate/3′ hydroxyl small RNAs during adapter ligation. For example, chemically pre-adenylated 3′ adapter deoxyoligonucleotides, which are blocked at their 3′ ends to avoid their circularization, can be used. The use of pre-adenylated adapters eliminates the need for ATP during ligation, and thus minimizes the problem of adenylation of the pool RNA 5′ phosphate that leads to circularization. Additionally, a truncated form of T4 RNA ligase 2, Rnl2(1-249), or an improved mutant, Rnl2(1-249)K227Q, can be used to minimize adenylate transfer from the 3′ adapter 5′ phosphate to the 5′ phosphate of the small RNA pool and subsequent pool RNA circularization. See also International Patent Application No. PCT/US2008/001227, published as WO 2008/094599, which is incorporated herein by reference in its entirety.
The length of the adapter sequences will vary. In a preferred embodiment, adapter sequences range from about 6 to about 500 nucleotides in length, preferably from about 8 to about 100, and most preferably from about 10 to about 25 nucleotides in length.
The cDNA amplicons are then sequenced. The sequencing can be performed by any known means. In a preferred embodiment, the sequencing method will generate sequences of amplicons of at least about 20 nucleotides in length.
For example, the amplicons can be sequenced using Solexa massive parallel sequencing which yields 30 million sequences of 32, 36 or 72 nucleotides in length per library and sequencing reaction. Solexa sequencing can also be carried out conveniently at a smaller scale processing a larger sample number, i.e. yielding about 1.5-4 million reads per sample. The larger sets are obtained, if a full sequencing plate is used. (See M. Hafner, P. Landgraf, J. Ludwig, A. Rice, T. Ojo, C. Lin, D. Holoch, C. Lim, T. Tuschl, Identification of microRNAs and other small regulatory RNAs using cDNA library sequencing, Methods, 2008, 44:3-12.) Alternatively, the amplicons can be sequenced using pyrosequencing (454 sequencing, Roche), which provides up to 400,000 sequences of up to 250 nt in length for a single read. Data management and sequence analysis from small RNA cDNA libraries is best carried out in collaboration with an experienced computational biology laboratory.
Next, the amplicons are assessed in order to identify those that include the portion of the RNA transcript that binds to the binding moiety in vivo.
In one embodiment, first unique sequences (i.e., nonredundant sequences) are identified and counted. Preferably, by various steps, the amplicons are filtered to remove irrelevant sequences (i.e., irrelevant amplicons). For example, the amplicon sequences can be filtered in accordance with any or all for the following rules:
The selected amplicons should have sufficient length to enable identification by means of sequencing or hybridization.
The selected amplicons should not have highly repetitive portion(s) within their sequence.
The selected amplicons should avoid sequences that may interfere with the manipulation of RNA and DNA while performing the invention (e.g. they should not have recognition sites for restriction endonuleases used during the manipulation process).
For example, the amplicons are narrowed to those more likely to include the portion of the RNA transcript that binds to the binding moiety in vivo. For example, in one embodiment, amplicons which are shorter than a certain number are removed, for example, less than 20 nucleotides or less than 15 nucleotides. Additionally, amplicons that do not map to a portion of the reference sequence being studied and/or amplicons that do not map to a portion of a known RNA sequence can be removed. Further, amplicons which contain highly repetitive portion(s) within their sequence (e.g., many multiples of TATA or GCGC) can be removed. Such sequences are referred to as “low entropic sequences.”
A “reference sequence” refers to any known sequence with which to compare an amplicon sequence. The reference sequence may be derived from a genomic sequence, a transcriptome sequence, an expressed sequence tags (EST) database, a sequence from which the RNA transcript was extracted, a known sequence library, a synthetic nucleotide sequence, a randomized RNA sequence, or a known RNA sequence. Typically, the human genomic sequence is being studied.
Next, the amplicons with overlapping sequences are “clustered.” “Clustering” refers to grouping together and aligning overlapping sequences.
In one embodiment, the quantities of amplicons in a particular cluster are then counted. For example, overlapping amplicon sequences, which differ by length simply because of a different point of digestion by a nuclease, can be counted as a cluster. For example, if ribonuclease-T1 is used, then a set of overlapping amplicons that differ by length because they extend to the next guanosine site are considered a cluster. Accordingly, in one embodiment, aligning the sequences of the amplicons includes determining which amplicons have the characteristic that, when aligned with the reference sequence, the reference sequence has a guanosine one nucleotide upstream from the 5′ end of the amplicons.
In another embodiment, aligning sequences occurs without narrowing down the amplicons in quantity before analyzing the amplicons.
The greater the quantity of amplicons in a particular cluster, the more likely that those amplicons include an RNA sequence expressed in vivo as opposed to being merely noise. (For example, see FIG. 2.) (See P. Berninger, D. Gaidatzis, E. van Nimwegen, M. Zavolan, Computational analysis of small RNA cloning data, Methods, 2008, 44, 13-21.)
Noise is the low frequency amplicon counts that are due to random degradation or RNA turnover products present as background in cross-linked RNA recovered from IP or gels. In one embodiment, noise is detected by the absence of a deoxythymidine to deoxycytidine mutation when using a thiouridine analog, such as 4-thiouridine, as the photoreactive nucleoside or by the absence of a deoxyguanosine to deoxyadenosine mutation when using a thioguanosine analog, e.g., 6-thioguanosine, as the photoreactive nucleoside. Noise can also be detected by the absence of very sharp “peaks” on a given transcript. Noise is seen as a random distribution of amplicons along a transcript.
In a further embodiment, aligning the sequences of the amplicons includes determining which amplicons have a mutation (preferably, a mismatch mutation) when compared to the reference sequence. For example, aligning the sequences of the amplicons may include determining which amplicons have a mutation wherein a deoxythymidine of the reference sequence is replaced by a deoxycytidine in the amplicons, when a thiouridine analog, such as 4-thiouridine, is used as the photoreactive nucleoside.
As another example, aligning the sequences of the amplicons may include determining which amplicons have a mutation wherein a deoxyguanosine of the reference sequence is replaced by a deoxyadenosine in the amplicons when using a thioguanosine analog, e.g., 6-thioguanosine, as photoreactive nucleoside. In one embodiment, such amplicons that are determined to have a mismatch mutation when compared to the reference sequence are considered “valid amplicons.”
In a preferred embodiment, the aligning the sequences of the amplicons includes determining which amplicons have at least one mismatch mutation when compared to the reference sequence. In another preferred embodiment, the step of aligning the sequences of the amplicons includes determining which amplicons have only one mismatch mutation when compared to the reference sequence.
A “mismatch” as used herein refers to a nucleic acid base that is any other nucleic acid base located on an amplicon at a specific position compared to the nucleic acid base that is aligned to the reference sequence. For example, at Position 1 on the amplicon is a thymidine, and on the reference sequence that is aligned, at Position 1, the mismatch can be Adenosine, Guanosine, or Cytosine. The mismatch between the amplicon and reference sequence may be due to deletions, insertions, substitutions, or frameshift mutations in the amplicon or reference sequence.
The sequences of the amplicons are then analyzed to determine the specific location on an RNA transcript that a given binding moiety binds in vivo, i.e., to determine the binding site. In this method, the amplicons are further narrowed down to find “valid amplicons.” A “valid amplicon” as used herein refers to an amplicon that is not noise, as described above.
A “valid amplicon” includes those having a mutation resulting from the introduction of the photoreactive nucleoside. For example, one method by which to find “valid amplicons” is to use the deoxythymidine to deoxycytidine mutation. Clustered amplicons with only a single mutation with respect to the “reference sequence,” i.e., the deoxythymidine to deoxycytidine mutation, are located. It is considered that the mutation occurred upon reverse transcription as described above. Such amplicons are considered to be “valid.”
Another method by which to find “valid amplicons” is to use the deoxyguanosine to deoxyadenosine mutation. Clustered amplicons with only a single mutation with respect to the “reference sequence,” i.e., the deoxyguanosine to deoxyadenosine mutation, are located. It is considered that the mutation occurred upon reverse transcription, as described above. Such amplicons are also considered to be “valid.”
Preferably, these “valid amplicons” are assessed in view of the total number of sequences that aligned to the region at issue, i.e., the total amplicons in a particular cluster. The total number of aligned sequences includes those sequences that have the mutation and those that do not have the mutation. The greater the percentage of the total aligned amplicons that show the mutation, the greater is the probability that the amplicons showing the mutation are “valid amplicons.”
When assessing the percentage, it is preferable to take into account the quantity of total aligned amplicons i.e., the total amplicons in a particular cluster. For example, a low percentage (e.g., 1% to 49%) is adequate to demonstrate a “valid amplicon” if the total quantity of aligned sequences is large (20 amplicons or more); and a high percentage (e.g., 50% to 100%) is adequate to demonstrate a “valid amplicon” if the total quantity of aligned sequences is small (19 amplicons or less. At least 10% of the sequences have to show the mutation to indicate a “valid amplicon.”
Another method by which to further validate “valid amplicons” is to use the property of ribonuclease-T1 to cleave the modified RNA transcript at 3′ phosphates of guanine residues. The modified transcripts were treated with ribonuclease-T1 to remove all or part of the modified RNA transcripts which was not covalently cross-linked to the binding moiety. In particular, it is determined which amplicon sequences have the characteristic that, when aligned with the “reference sequence” (i.e., in a particular cluster), the “reference sequence” has a guanosine one nucleotide upstream from the 5′ end of the amplicons. Such amplicons are also considered to be “valid amplicons.”
Preferably, an amplicon is considered to be a “valid amplicon” if it is considered to be valid by both the method of using the mutation and the method of using the cleaving property of ribonuclease-T1.
Once “valid amplicons” have been identified, they are further analyzed in view of the “reference sequence” to determine the presence of a consensus motif or sequence within a binding site. The binding site can be part of coding transcript or non-coding transcript of RNA.
For example, the deoxythymidine to deoxycytidine mutation and/or the deoxyguanosine to deoxyadenosine mutation in the amplicon is used as an anchor for comparing the sequence surrounding the mutation to the “reference sequence.” Such surrounding sequence is termed “sequence window.”
In one embodiment, the “sequence window” includes the mutation plus at least one nucleotide on either side of the mutation. Preferably, the number of nucleotides on either side of the mutation ranges from about 5 to about 20 nucleotides. In another embodiment, the mutation is at the center of the sequence window.
As an illustration, see FIG. 3 for the alignment of amplicons against Pum2-targeted transcript. The Pum2 motif contains the mutated deoxythymidine in the recognition element TGTANATA (the underlined T is mutated in every valid amplicon).
As is known in the art, a number of different programs and algorithms may be used to identify whether an amplicon has sequence identity or similarity to a known sequence. Sequence identity and/or similarity is determined using standard techniques known in the art, including, but not limited to, the local sequence identity algorithm of Smith & Waterman, Adv. Appl. Math., 2:482 (1981), by the sequence identity alignment algorithm of Needleman & Wunsch, J. Mol. Biol., 48:443 (1970), by the search for similarity method of Pearson & Lipman, Proc. Natl. Acad. Sci. U.S.A., 85:2444 (1988), by computerized implementations of these algorithms (GAP, BESTFIT, FASTA, and TFASTA in the Wisconsin Genetics Software Package, Genetics Computer Group, 575 Science Drive, Madison, Wis.), the Best Fit sequence program described by Devereux et al., Nucl. Acid Res., 12:387-395 (1984), preferably using the default settings, or by inspection. All references cited in this paragraph are incorporated by reference in their entirety.
In one embodiment, motif searches are conducted for the extracted sequences by computational means known in the art. Examples of methods used in conducting motif searches (i.e., consensus sequence searches) include CONSENSUS, multiple expectation maximization for motif elicitations (MEME) program, Gibbs sampling, PhyloGibbs sampling, Motif Discovery scan program (MDScan), or AlignACE (Roth, F. P., Hughes, J. D., Estep, P. W. & Church, G. M. Finding DNA regulatory motifs within unaligned noncoding sequences clustered by whole-genome mRNA quantitation. Nat Biotechnol 16, 939-45 (1998)). For example, the MEME program finds conserved ungapped short motifs within a group of related, unaligned sequences (Bailey and Gribskov, 1998, J Comput Biol, 5:211-21). MDScan, for example, is used to identify sequence motifs from a set of identified genomic regions (Liu X S et al. (2002) Nat. Biotechnol., 20(8):835-9). In another embodiment, more than one algorithm may be used to identify motifs for the extracted sequences.
In one embodiment, the analysis of the amplicon sequences can further include identifying a feature required for interaction of the binding site and the binding moiety. For example, evaluation of the consensus sequence of the binding site can reveal a structure, such as a stem loop, that may be required or involved in binding to the binding moiety. Additionally, the property of the ribonuclease-T1 cleavage can further be used to validate the sequences, as described above.
Once the consensus motif of the binding site has been identified using the methods described above, it can be utilized for various clinical or research applications. For example, the binding site can be sequenced using patient DNA to identify mutations, deletion or insertions that may link a genetic alteration in an important, regulatory RNA segment to a disease condition. It is known that RNA binding proteins are essential regulators of proteins by binding to coding and non-coding RNAs and regulating their transcription, modification, splicing, nuclear export, transport and translation. Consequently, understanding the binding site on the RNA and the identity of the bound RNA binding proteins offer ways to targeted therapy. For example, an RNA binding protein known to affect the stability or translation of a gene can be utilized as a drug target for the regulation of the targets of the gene.
In one embodiment, the present invention is kit for identifying a binding site on an RNA transcript. The kit comprises a thiouridine analog, as described above, or thiouridine analog-supplemented tissue culture medium; reagents for nuclease digestion, as described above; reagents for immunoprecipitation, as described above; and adapters and primers for small RNA cloning.
The following oligodeoxynucleotides were used for PCR and cloning of the listed genes (restriction site are underlined);
Plasmids pENTR4 IGF2BP1, -2, -3, QKI, and PUM2 were generated by PCR amplification of respective coding sequences (CDS) followed by restriction digest and ligation into pENTR4 (Invitrogen). Primer sequences are listed under DNA oligonucleotides in Suppl. Material. pENTR4 IGF2BP1, -2, and -3 were recombined into pFRT/TO/FLAG/HA-DEST destination vector using GATEWAY LR recombinase according to manufacturer's protocol (Invitrogen), which allow for inducible expression of stably transfected FLAG/HA-tagged protein in Flp-In T-REx HEK293 cells (Invitrogen) from the TO/CMV promoter. Whereas pENTR4 QKI and pENTR4 QKI were recombined into pFRT/FLAG/HA-DEST, which only allows constitutive expression in Flp-In T-REx HEK293 cells. Plasmids for overexpression of N-terminally His-tagged IGF2BP1, -2, and -3 were generated by ligation of CDS into pET16 (Novagen). The plasmids described in this study can be obtained from Addgene.
Recombinant Protein Expression
pET16 IGF2BP1, -2, and -3 plasmid were transformed in E. coli STAR(DE3) (Invitrogen). Cells were grown in LB supplemented with 50 μg/ml ampicillin at 37° C. to A600=0.5. Culture was cooled to 25° C. for induction of soluble protein by addition of IPTG to a final concentration of 1 mM. Incubation was continued at 25° C. for 3 h before cells were harvested and resuspended in 8 ml/g cells of lysis buffer (25 mM Tris-HCl, pH 7.5, 0.5 M NaCl, 5 mM imidazole, and complete EDTA-free protease inhibitor cocktail (Roche)) All following steps were carried out at 4° C. Cells resuspended in lysis buffer were sonicated to complete lysis. Insoluble material was removed by centrifugation at 12,000×g. Supernatant was incubated with 250 μl/10 ml cell supernatant HIS-Select Cobalt Affinity Gel (Sigma) for 1 h. The affinity gel was washed three times with 10 gel volumes of lysis buffer and His-tagged proteins were eluted in 4 gel volumes of 25 mM Tris-HCl, pH7.5, 0.5 M NaCl, and 250 mM imidazole.
Cell Lines and Cell Culture
HEK293 T-REx Flp-In cells (Invitrogen) were grown in D-MEM high glucose (1×) with 10% (v/v) fetal bovine serum, 1% (v/v) 2 mM L-glutamine, 1% (v/v) 10,000 U/ml penicillin/10,000 μg/ml streptomycin, 100 μg/ml zeocin and 15 μg/ml blasticidin. Cell lines stably expressing FLAG/HA-tagged proteins were generated by co-transfection of pFRT/TO/FLAG/HA or pFRT/FLAG/HA constructs with pOG44 (Invitrogen). Cells were selected by exchanging zeocin with 100 μg/ml hygromycin. Expression of FLAG/HA-IGF2BP1, -2, -3 was induced by addition of 250 ng/ml doxycycline 15-20 h before crosslink.
UV 254 nm or UV 365 nm Crosslinking
For UV crosslink, cells were washed once with ice-cold PBS while still attached to the plate. PBS was removed completely and cells were irradiated with 254 nm UV light, or 365 nm UV light for cells treated with nucleoside analogues (0.15 J/cm2) using a UV Stratalinker 2400 (Stratagene). After UV treatment, 5 ml of ice-cold PBS was added per plate immediately. Cells were scraped off the plate and collected by centrifugation in 15 ml Falcon tubes (Sorvall legend RT, 2000 rpm, 10 min, 4° C.).
Cell Lysis and Partial RNase T1 Digestion
The pellets of 365 nm UV-crosslinked cells (ca. 3 ml cell pellet for about 20 15 cm cell culture dishes) were resuspended in approximately 3 cell pellet volumes of NP40 lysis buffer (50 mM HEPES, pH 7.5, 150 mM KCl, 2 mM EDTA, 1 mM NaF, 0.5% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and passively incubated on ice for 10 min. The cell lysate was cleared by centrifugation in 13 ml polypropylene tubes (Sorvall SS34, 13,000 rpm, 10 min, 4° C.) and the supernatant was passed through a 0.5 μm Supor membrane syringe filter (Pall). RNase T1 (Fermentas) was added to the cleared cell lysates to a final concentration of 1 U/μl and incubated in a water bath at 22° C. for 15 min and subsequently cooled for 5 min on ice before addition of antibody-conjugated magnetic beads (see below).
Immunoprecipitation and Recovery of Crosslinked Target RNA Fragments
Preparation of Magnetic Beads
Conjugation of antibodies to magnetic beads was performed the same day as the immunoprecipitation (IP). Dynabeads Protein G magnetic particles (Invitrogen) were resuspended by vortexing and an aliquot (10 μl for per ml cell lysate) was transferred to an Eppendorf tube. Beads were collected with the Dynal MPC-S magnetic particle concentrator (Invitrogen) for removal of the supernatant. Beads were washed twice with 1 ml of citrate-phosphate buffer (pH 5.0, 4.7 g/l citric acid, 9.2 g/l Na2HPO4) and resuspended in twice the volume of citrate-phosphate buffer relative to the original volume of bead suspension. 0.25 μg of anti-FLAG M2 monoclonal antibody (Sigma) per ml suspension was added and incubated on a rotating wheel at room temperature for 40 min. Beads were then washed twice with 1 ml of citrate-phosphate buffer to remove unbound antibody and resuspended again in twice the volume of citrate-phosphate buffer relative to the original volume of bead suspension.
Immunoprecipitation (IP), Further RNase T1 Digestion, and Dephosphorylation
A typical IP for an average expressed RNA-binding protein requires about 10 ml crosslinked and partially RNase T1 digested lysate (see above). 10 μl of freshly prepared antibody-conjugated magnetic beads per ml of cell lysate were added and incubate in 15 ml centrifugation tubes on a rotating wheel for 1 h at 4° C. Magnetic beads were collected on a magnetic particle collector (Invitrogen). Manipulations of the following steps were carried out in 1.5 ml centrifugation tubes. The supernatant was removed from the bead-bound material, which was washed 3 times with 1 ml of IP Wash Buffer (50 mM HEPES-KOH, pH 7.5, 300 mM KCl, 0.05% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)). Beads were resuspended in one volume (here and in following steps volume refers to volume of used bead suspension) of IP Wash Buffer. RNase T1 (Fermentas) was added to obtain a final concentration of 100 U/μl, and the bead suspension was incubated in a water bath at 22° C. for 15 min, and subsequently cooled for 5 min on ice. Beads were washed 3 times with 1 ml of High-Salt Wash Buffer (50 mM HEPES-KOH, pH 7.5, 500 mM KCl, 0.05% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and resuspended in one volume of Dephosphorylation Buffer (50 mM Tris-HCl, pH 7.9, 100 mM NaCl, 10 mM MgCl2, 1 mM DTT). Calf Intestinal Alkaline Phosphatase (NEB) was added to obtain a final concentration of 0.5 U/μl, and the suspension was incubated for 10 min at 37° C. Beads were washed twice with 1 ml of Phosphatase Wash Buffer (50 mM Tris-HCl, pH 7.5, 20 mM EGTA, 0.5% (v/v) NP40) and twice with 1 ml of Polynucleotide Kinase (PNK) Buffer (50 mM Tris-HCl, pH 7.5, 50 mM NaCl, 10 mM MgCl2, 5 mM DTT). Beads were resuspended in one volume of Polynucleotide Kinase Buffer.
Radiolabeling of RNA Segments Crosslinked to Immunoprecipitated Proteins
To the bead suspension described above, γ-32P-ATP (GE Healthcare, formerly Amersham Biosciences) was added to a final concentration of 0.5 μCi/μl and T4 Polynucleotide Kinase (NEB) to 1 U/μl. The suspension was incubated for 30 min at 37° C. Thereafter, non-radioactive ATP (Fermentas) was added to obtain a final concentration of 100 μM and the incubation was continued for another 5 min at 37° C. The magnetic beads were then washed 5 times with 800 μl of PNK Buffer and resuspended in 70 μl of SDS-PAGE Loading Buffer (10% glycerol (v/v), 50 mM Tris, pH 6.8, 2 mM EDTA, 2% SDS (w/v), 100 mM DTT, 0.1% Bromophenol blue).
SDS-PAGE and Electroelution of Crosslinked RNA-Protein Compounds from Gel Slices
The radiolabeled bead suspension was denatured for 5 min at 95° C., and the supernatant, which was separated from the magnetic beads, was loaded in 2 wells of a NuPAGE Novex 4-12% BT Midi 1.0 gel (Invitrogen) next to well loaded with 5 μl of Precision Plus Protein Dual Color Standard (BioRad). The gel was run for 1 h at 200 V in NuPAGE MOPS SDS Running Buffer (Invitrogen). To detect radiolabeled bands, the gel was covered in Saran plastic wrap and analyzed by phosphorimaging. The radioactive band migrating at the expected molecular weight of the target protein and any other distinct molecular size bands, were excised from the gel and electroeluted in a D-Tube Dialyzer Midi with a molecular cut-off of 3.5 kD (Novagen). The electroelution was performed at 100 V for 2 h in SDS Running Buffer (25 mM Tris base, 192 mM glycine, 0.1% (w/v) SDS). The eluate (approx. 400 μl) was transferred to an Eppendorf tube.
Proteinase K Digestion
An equal volume of 2× Proteinase K Buffer (100 mM Tris-HCl, pH 7.5, 150 mM NaCl, 12.5 mM EDTA, 2% (w/v) SDS) with respect to the electroeluate was added, followed by the addition of Proteinase K (Roche) to a final concentration of 1.2 mg/ml, and and incubation for 30 min at 55° C. The RNA was recovered by acidic phenol/chloroform extraction. One volume of acidic phenol:chloroform:isoamyl alcohol (25:24:1) was added to the proteinase K digest and vortexed. Phases were separated by centrifugation in a tabletop centrifuge (13,000 rpm, 5 min). The upper aqueous phase was transferred to a new tube, extracted once with an equal volume of chloroform. The RNA was then precipitated from the aqueous phase after adding 1/10 volume of 3 M NaCl, 2 μl of 10 mg/ml GlycoBlue (Ambion) and 3 volumes of absolute ethanol. After incubation of at least 1 h on ice or overnight at −20° C., the precipitated RNA was collected by centrifugation (13,000 rpm, 20 min, 4° C.), and the pellet was dried and resuspended in 30 μl of sterile water.
cDNA Library Preparation and Deep Sequencing
The recovered RNA was carried through a cDNA library preparation protocol originally described for cloning of small regulatory RNA (Hafner, Methods, 2008). The first step, 3′ adapter ligation, was carried out as described on a 20-μl scale using 10.5 μl of the recovered RNA. UV 254 nm crosslinked RNAs were processed using standard adapter sets, followed by PCR to introduce primers compatible with 454 sequencing; UV 365 nm crosslinked sample RNAs were processed using Solexa sequencing adapter sets. Depending on the amount of RNA recovered, fast migrating 5′-adapter-3′-adapter products without inserts may be detected after amplification of the cDNA. In such case, the expected longer PCR products was excised from a 3% NuSieve low-melt agarose, eluted from the gel pieces with the Illustra GFX-PCR purification kit (GE Healthcare) and directly submitted for Solexa sequencing. 454 pyrosequencing was carried out by Agnes Viale at the Sequencing core facility at Memorial Sloan Kettering Cancer Center, New York. Solexa sequencing was performed by Scott Dewell at the Genomics Resource Center of Rockefeller University.
2-D electrophoresis begins with 1-D electrophoresis but then separates the molecules by a second property in a direction 90 degrees from the first. The two dimensions that proteins can be separated into using this technique can be, for example, isoelectric point, protein complex mass in the native state, and protein mass.
Human embryonic kidney (HEK) 293 cells stably expressing FLAG/HA-tagged IGF2BP1-3, QKI, and PUM2 (Landthaler et al., 2008, incorporated herein by reference in its entirety) were grown overnight in medium supplemented with 100 μM 4-thiouridine (4SU). Living cells were irradiated with 365 nm UV light (0.15 J/cm2) n a Stratalinker 2400 (Stratagene). Cells were harvested and lysed in 3 cell pellet volumes of NP40 lysis buffer (50 mM HEPES-KOH, pH 7.5, 150 mM KCl, 2 mM EDTA, 1 mM NaF, 0.5% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)). The cell lysate was cleared by centrification at 13,000×g.
RNase T1 (Fermentas) was added to the cleared cell lysates to a final concentration of 1 U/μl and the reaction mixture was incubated at 22° C. for 15 min and subsequently cooled on ice before addition of antibody-conjugated magnetic beads. FLAG/HA-tagged AGO proteins were immunoprecipitated with anti-FLAG antibodies bound to Protein G Dynabeads. RNase T1 (Fermentas) was added to obtain a final concentration of 100 U/μl to the immunoprecipitate, and the bead suspension was incubated in a water bath at 22° C. for 15 min, and subsequently cooled for 5 min on ice. Beads were washed 3 times with wash buffer (50 mM HEPES-KOH, pH 7.5, 500 mM KCl, 0.05% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and resuspended in one volume of dephosphorylation buffer (50 mM Tris-HCl, pH 7.9, 100 mM NaCl, 10 mM MgCl2, 1 mM DTT). Calf intestinal alkaline phosphatase (NEB) was added to obtain a final concentration of 0.5 U/μl, and the suspension was incubated for 10 min at 37° C.
Beads were washed twice with phosphatase wash buffer (50 mM Tris-HCl, pH 7.5, 20 mM EGTA, 0.5% (v/v) NP40) and twice with polynucleotide kinase (PNK) Buffer (50 mM Tris-HCl, pH 7.5, 50 mM NaCl, 10 mM MgCl2, 5 mM DTT). The crosslinked RNA segments were radiolabeled on the magnetic beads with γ-32P-ATP and 1 U/μl T4 PNK (NEB). After incubation for 30 min at 37° C., non-radioactive ATP (Fermentas) was added to a final concentration of 100 μM and the incubation was continued for another 5 min.
The magnetic beads were then washed 5 times with PNK Buffer, resuspended in 70 μl of SDS-PAGE loading buffer, and incubated for 5 min at 95° C. The protein-RNA complexes were separated on a SDS-PAGE. The gel was analyzed by phosphorimaging. The radioactive band migrating at the expected molecular weight of AGO proteins was excised from the gel and electroeluted. The electroeluate was proteinase K digested. The RNA was recovered by acidic phenol/chloroform extraction followed by a chloroform extraction and an ethanol precipitation. The recovered RNA was carried through a small RNA cDNA library preparation protocol as described for cloning of small RNAs (Hafner et al., 2008) and Solexa sequenced.
The extracted sequence reads were mapped to the human genome (hg18) and human mRNAs. Transfection of siRNAs and mRNA profiling by array analysis were described previously (Landthaler et al., 2008).
Incorporation of modified photoreactive nucleoside analogues into RNA facilitates UV crosslinking to RBPs (Meisenheimer and Koch, 1997), as shown previously for in vitro transcribed site-specifically labeled RNAs (Kirino and Mourelatos, 2008; Moore and Query, 1998). Several of these photoreactive nucleosides are excitable by longer wavelength UV (>300 nm) and are readily taken up by cells, where they are incorporated into nascent transcripts without noticeable toxicity and then can be used for in vivo crosslinking (Bezerra and Favre, 1990; Favre et al., 1986; Wagenmakers et al., 1980). We applied a subset of these nucleoside analogues (FIG. 1A) to cultured cells expressing the FLAG/HA-tagged RBP IGF2BP1. Live cells were UV-irradiated at 365 nm. The crosslinked RNA-protein complexes were isolated by immunoprecipitation, and the covalently bound RNA was partially digested with RNase T1 and radiolabeled. Separation of the radiolabeled RNPs by denaturing SDS gel electrophoresis indicated that 4SU-labeled RNA crosslinked most efficiently to IGF2BP1. Compared to unmodified UV 254 nm crosslinking, the nucleoside analogues enhanced RNA recovery between 100- and 1000-fold (FIG. 1B and FIG. 8). It is important to note that we did not optimize either short or long wavelength UV exposure to define the limits of crosslinking, but used standard laboratory UV crosslinking equipment at their high intensity settings and a time-window selected to minimize overheating and denaturation of cellular proteins.
Based on our initial analysis we selected 4SU as crosslinker. However, before we proceeded with analysis of IGF2BP proteins, for which the consensus binding sites were not known, we studied human Pumilio 2 (PUM2), a member of the well-studied Pufprotein family (FIG. 9A-E) with well defined consensus motifs (Wang et al., 2002; Wickens et al., 2002).
PUM2 protein, similar to IGF2BP1, readily crosslinked to 4SU-containing transcripts (FIG. 2B and FIG. 8B), and the crosslinked segments were converted into a cDNA library (FIG. 1C) and Solexa sequenced to obtain about 10 million reads (Hafner et al., 2008). The sequence reads were annotated by alignment against the human genome and EST databases. Sequence reads mapping to one unique position in the genome with up to one mismatch were used to build sequence clusters (FIG. 2C and Supplementary Methods). We identified 7,523 clusters mapping to about 3,000 unique transcripts, 93% of which were found within the 3′UTR, in agreement with previously published work characterizing the C. elegans members of this protein family (Wickens et al., 2002) (FIG. 9). All sequence clusters with information on the annotation and mapping are available online at the RNA regulatory networks-Zavolan lab website.
The PUM2 RNA recognition element, UGUANAUA (Galgano et al., 2008; Gerber et al., 2006; Gupta et al., 2008; Zamore et al., 1997), was readily identified by a PhyloGibbs analysis (Siddharthan et al., 2005) from the top 100 sequence read clusters annotated as mRNA (FIG. 2D).
About 71% of the clustered sequence reads showed a thymidine (T) to cytidine (C) alteration, with the T corresponding to U7 of the RRE mutated at higher frequency compared to the Ts corresponding to U1 and U3 (FIG. 2E), which are bound tightly in the PUM2 RNA binding site and for which the local amino acid environment probably does not allow efficient crosslinking. The 71% conversion frequency of a uridine residue to cytidine within sequence reads mapping to mRNA transcripts was unexpected, and suggested that crosslinked 4SU residues encountered during reverse transcription caused the sequence change, and that within the consensus motif, certain residues were preferably crosslinking 4SU residues in immediate vicinity of the binding site were also found to participate in crosslinking, indicating that our method can capture PUM2 binding sites that have a nucleotide other than U at position 7.
To further validate our method, we investigated the RBP Quaking (QKI) for which neither the RNA recognition element nor the target mRNAs have been so far defined, though an ACUAAY motif was determined by SELEX (Galarneau and Richard, 2005). The QKI protein contains one heterogeneous nuclear ribonucleoprotein K homology (KH) domain (FIG. 3A), embedded in the context of a larger protein domain involved in signal transduction (Chen and Richard, 1998). Mice with reduced expression of QKI show dysmyelination and develop rapid tremors or “quaking” 10 days after birth (Ebersole et al., 1996). The precise molecular function of QKI and its presumably deregulated mRNA targets responsible for the mutant phenotype are unknown. QKI has been implicated in post-transcriptional regulation including pre-mRNA splicing, mRNA export, mRNA stability and protein translation (Chenard and Richard, 2008). FLAG/HA-tagged QKI was 4SU-crosslinked to RNA (FIG. 3B, FIG. 8B), and the isolated RNA was converted into a cDNA library and Solexa sequenced.
PhyloGibbs analysis of the top 100 clusters aligning to sequences annotated as mRNA (FIG. 3C) yielded the sequence AYUAAY as the RNA recognition element (FIG. 3D), similar to the sequence motif ACUAAY identified by SELEX from randomized RNA sequences (Galarneau and Richard, 2005). We found that close to 75% of the more than 6,000 sequence clusters were derived from intronic sequences, supporting the hypothesis that QKI is a splicing regulator (Wu et al., 2002). 70% of the remaining, exonic, clusters fall into 3′UTRs. Together, all of the identified sites map to about 2,500 protein-coding transcripts. This analysis shows that our protocol was not only able to isolate RNA segments from mature transcripts accumulating in the cytoplasm but also segments present in presumably nuclear-localized pre-mRNAs. T to C mutation analysis of the clustered sequence reads showed that T corresponding to position 2 in AUUAAY was frequently sequenced as C whereas the T corresponding to position 3 in AUUAAY or ACUAAY remained unaltered (FIG. 3E). Crosslinking of 4SU residues in immediate vicinity to the consensus binding site was sufficient for exposing the motif with C at position 2.
To better understand the cytidine sequence conversion after sequencing of crosslinked RNA segments, we subjected synthetic 4SU-containing oligoribonucleotides together with recombinant QKI to UV 365 nm crosslinking (FIGS. 3F, G). Oligoribonucleotide U9 with a 4SU substitution at position 2 of the recognition element crosslinked approximately 50-times more efficiently than oligonucleotide U10 with a modification at position 3 (FIG. 3G). RNA-binding to recombinant QKI protein was unaffected by 4SU incorporation at U9 or U10 as determined by gel-shift analysis (FIG. 11A-B), whereas regular mutation of the recognition element weakened the binding (FIG. 11C). The position-dependent crosslinking results suggest that the local amino acid environment of the protein RNA-binding site influences crosslinking efficiency and that T to C changes in the cDNA identify crosslink sites.
When we sequenced a library prepared from a synthetic 4SU-labeled oligoribonucleotide crosslinked to recombinant QKI protein, approximately 80% of the sequence reads derived from the crosslinked RNA oligonucleotide showed a T to C change at the modified position compared to about 20% of the sequence reads obtained from non-crosslinked RNA (FIG. 3H). A likely explanation for the 4-fold increase in T to C transition is a chemical structural change of 4SU upon crosslinking to an amino acid side chain, resulting in a modified base with altered hydrogen bond donor/acceptor properties preferentially directing the incorporation of deoxyguanosine rather than deoxyadenosine during reverse transcription (FIG. 10D). At the doses of 4SU applied to cultured cells, about 1 out of 50 uridines will be substituted by 4SU, and the average T to C conversion of a 40-nt sequence reads derived from non-crosslinked sequences, the background, is estimated to be near 4%. Clusters of sequence reads with average T to C conversion above this threshold, irrespective of the number of sequence reads, most certainly represent crosslinking sites, unless polymorphisms or errors in the genome assembly occur. Nonetheless, such problems cannot have a high prevalence, because we found that in the case of PUM2 and QKI, the binding motifs are strongly enriched around the positions with high T to C mutation frequency, something that would not be expected if these positions were in fact polymorphic or part of an incorrect assembly (FIG. 12). Therefore, the ability to separate signal from noise by focusing on clusters with a high frequency of T to C mutations rather than clusters with the largest number of reads, represents a major enhancement of our method over conventional CLIP, and we therefore termed our method PURE-CLIP (Photoreactive-Uridine-Enhanced Crosslinking and Immunoprecipitation).
To assess whether the transcripts identified by PURE-CLIP are regulated by QKI, the protein was knocked down using siRNAs and the mRNA levels of QKI knockdown cells were compared to that of mock-transfected cells using microarray analysis. The mRNA levels of transcripts identified by PURE-CLIP were significantly increased, indicating that QKI protein negatively regulated bound mRNAs, consistent with its proposed function as being a repressor protein (FIG. 3I and FIG. 13) (Chenard and Richard, 2008).
We then applied PURE-CLIP to study the insulin-like growth factor 2 mRNA-binding proteins 1, 2, and 3 (IGF2BP1-3), comprising a family of highly conserved proteins that play a role in cell polarity and cell proliferation (Yisraeli, 2005). They are predominantly expressed in the embryo but are re-expressed in various cancers (Boyerinas et al., 2008; Dimitriadis et al., 2007). IGF2BP2 has also recently been associated with type-2 diabetes (Diabetes Genetics Initiative of Broad Institute of Harvard and MIT et al., 2007; Scott et al., 2007). The IGF2BPs regulate mRNA stability, transport and translation (Yisraeli, 2005). They contain six canonical RNA-binding domains, two RNA recognition motifs (RRMs) and four KH domains (FIG. 4A). RRMs recognize several nucleotides of single-stranded RNA (Clery et al., 2008), but target recognition for this six-domaincontaining protein family appears complex, and so far, only a small number of coding and non-coding RNAs have been shown to interact with IGF2BPs without precise localization or definition of its RREs (Yisraeli, 2005).
We crosslinked FLAG/HA-tagged IGF2BP1, 2, and 3 to 4SU-labeled RNA (FIG. 4B and FIG. 8B) and analyzed the bound RNA by Solexa sequencing. The three IGF2BPs recognized a highly similar set of target transcripts (FIG. 14A-E), suggesting similar and redundant functions consistent with the high degree of sequence similarity between family members. PhyloGibbs analysis of the clusters derived from mRNAs (FIG. 4C) yielded the sequence CAUH (H=A, U, or C) as the only consensus recognition element (FIG. 4D). In total, we identified over 100,000 sequence clusters recognized by the IGF2BP family that map to about 8,400 protein-coding transcripts. The annotation of the clusters was predominantly exonic (ca. 90%), and there was a slight preference for binding in the 3′UTR compared to the coding sequence (CDS) (FIG. 9A-E). The mutation frequency of all sequence tags containing the element CAUH showed that the crosslinked residue was positioned inside the motif, or immediately around it (FIG. 4E). The consensus motif CAUH was generally found repeated in targeted transcripts, predominantly within a distance of three to five nucleotides (FIG. 15A-D). Furthermore, in vitro binding assays showed that nucleotide changes of the CAUH motif decreased the binding affinity (FIG. 4F and FIG. 16A-C), emphasizing the importance of the consensus motif. Nevertheless, mutation of an RRE did not fully abolish binding, suggesting that the sequence environment around the RRE, e.g. its high overall CA-content and likely the multiplicity of binding sites, also contribute to binding (FIG. 16A-C).
To test whether binding of the IGF2BPs affects the stability of their interacting mRNAs, as reported previously for some targets, we simultaneously depleted all three IGF2BP members using siRNAs (FIG. 17) and compared the cellular RNA from knockdown and mock-transfected cells on microarrays. Transcripts identified by CLIP showed a decrease in their mRNA levels in IGF2BP depleted cells, indicating that the IGF2BP proteins stabilize their target mRNAs. In addition, the top mRNAs, ranked according to the total number of T to C mutations were most destabilized (FIG. 4G). This indicates that the ranking criterion that we derived based on the analysis of PUM2 and QKI data generalizes to other RBPs. For comparison to conventional CLIP and deep-sequencing HITS-CLIP (Licatalosi et al., 2008; Ule et al., 2003), we also sequenced cDNA libraries prepared from UV 254 nm crosslinking and then compared the sequence clusters derived from these experiments with those of PURE-CLIP (FIG. 18). UV 254 nm crosslinking identified the identical segments of a target RNA, yet, the crosslinking site was not readily deduced, because the sequence reads making up the cluster had a mutation rate lower than 1% with a weak G to A bias in the mutation frequency that might be more difficult to exploit (FIG. 18). This can lead to problems separating signal from noise and deducing binding sites from analysis of clusters simply ranked by number of sequence reads.
To test our approach on ribonucleoprotein complexes, we selected the protein components mediating miRNA-guided target RNA recognition. In animal cells, miRNAs recognize their target mRNAs through base-pairing interactions between 6-8 nucleotides at the 5′ end of the miRNA, the so called “seed”, and complementary sequences thought to reside mainly in the 3′UTR of mRNAs. Computational methods to predict miRNA binding sites in several organisms mostly rely on the interaction of the miRNA seed region with its target mRNA 3′UTR- and evolutionary conservation of the target site (Gaidatzis et al., 2007; Grimson et al., 2007; Grun et al., 2005; John et al., 2004; Krek et al., 2005; Lewis et al., 2005). miRNA target sites are located in the 3′UTRs of mRNAs (Brennecke et al., 2005; Grimson et al., 2007; Wightman et al., 1993). There are also individual examples of target mRNAs that are efficiently repressed by miRNA-binding sites in the 5′ UTR and the coding sequence (CDS) (Kloosterman et al., 2004; Lytle et al., 2007; Tay et al., 2008). We isolated mRNA fragments bound by miRNPs from HEK293 cell lines stably expressing FLAG/HA-tagged AGO and TNRC6 family proteins (Landthaler et al., 2008).
The AGO IPs revealed two prominent RNA-crosslinked full-length proteins of 100 and 180 kDa size, representing AGO, and likely TNRC6 and/or DICER1 protein, respectively (Landthaler et al., 2008) (FIG. 5A). The TNRC6 IPs showed one prominent RNAcrosslinked protein band of 180 kDa size, and several smaller proteins, one of which likely represents AGO (Landthaler et al., 2008). The RNA recovered from the AGO-band in the AGO-IP and from the TNRC6-band in the TNRC6-IP was converted into cDNA libraries and Solexa sequenced.
We clustered overlapping, uniquely-mapped reads that were at least 20 nts long into clusters, and we retained for further analyses those clusters that contained at least five sequence reads and had a content of more than 20% crosslinked sequences (based on the T to C transition, indicative of crosslinking) We removed all clusters that were present at a tenfold higher level in the cDNA libraries derived from PURE-CLIP for the completely unrelated IGF2BP protein family, and therefore constitute background sequences. Finally, we selected sequence regions that were centered over the predominant crosslinking site from each cluster, and extended them 20 nt in both directions to allow for accommodation of the various positions at which target RNA crosslinking was observed (see below). We refer to those 41-nt sequences as crosslinkcentered regions (CCRs).
The AGO experiments yielded on average about 4,000 clusters that overlapped, supporting our earlier observation that AGO1 through 4 bound similar sets of transcripts (Landthaler et al., 2008). We therefore combined all sequence reads obtained from the individual AGO experiments and identified 17,319 clusters of sequence reads using the cut-offs defined above (FIG. 5B)). Note that a large number of additional clusters that were not present in the individual AGO1-4 cluster sets, because of the condition that a cluster is made up by at least 5 sequence reads, were created. The clusters mapped to 4,647 transcripts with defined GeneIDs, corresponding to 21% of the 22,466 unique HEK293 transcripts we identified by Solexa sequencing of oligoT-purified total RNA (digital gene expression, DGE, Illumina).
The individual TNRC6 experiments yielded on average about 600 clusters that were similar to each other, also consistent with our observation that TNRC6 family proteins bind similar sets of transcripts (Landthaler et al., 2008). We therefore combined all sequence reads obtained by the different TNRC6 experiments, yielding 1,865 clusters. Although they had similar numbers of mRNA-annotated sequence reads, the TNRC6 libraries had a higher proportion of sequence reads that were shorter than 20 nts and were not considered for construction of clusters. For this reason we identified fewer clusters in the TNRC6 dataset compared to the AGO dataset. An analysis of the TNRC6 CCRs showed that the crosslinked positions of 26% of the CCRs were virtually indistinguishable from those obtained for AGO (at least 75% overlap between the 41-nt AGO- and TNRC6-derived CCRs) (FIG. 21). We showed previously that AGO and TNRC6 protein family members bound to a large extent the same transcripts (Landthaler et al., 2008). Here we find that when AGO and TNRC6 bind to the same transcripts they also bind to the same site. We speculate that the TNRC6 protein acts like a lid covering the solvent-accessible side of the AGO miRNA-target RNA complex (Wang et al., 2008a) in order to stabilize the interaction with the target mRNA.
To relate CCRs to the endogenously expressed miRNAs, we determined the miRNA profiles in cell lysate as well as AGO1-4 IPs of non-crosslinked FLAG/HA-AGO expressing HEK293 cells by Solexa sequencing as well as the miRNA profiles from miRNAs in the PURE-CLIP experiments. The non-crosslinked miRNA profiles were highly similar (FIG. 5C) and supported our observation that AGO1 to 4 crosslinked to similar transcript sites. The most abundant miRNAs in our HEK293 cells were miR-103 (7% of miRNA sequence reads), miR-93 (6.5%), and miR-19b (5.5%), showing that HEK293 cells differ from other systems such as muscle, liver or brain, in that they do not have a clearly dominant miRNA, which could yield a strong sequence signature in the mRNAs. The top 25 expressed miRNAs account for 72%, and the top 100 account for 95% of the total of miRNA sequence reads; the residual 5% correspond to 457 distinct miRNAs and miRNA*. We restricted the subsequent computational analysis of miRNA target sites to the top 100 miRNAs and their sequence families, and the experimental inhibition analysis of miRNAs for assessment of endogenous miRNA binding on mRNA stability to the top 25 most abundant miRNAs.
miRNA profiles from combined AGO1- to 4 PURE-CLIP experiments were constructed from a total of 2.4 M sequence reads mapping to the miRNA precursor sequences from the latest version of Rfam (v9.1) (Gardner et al., 2009). The vast majority of sequence reads derived from prototypical miRNAs (Landgraf et al., 2007) displayed T to C conversion near or above 50%. Five of the 100 most abundant miRNAs in HEK293 cells lack uridines at position 8-13, yet only 2 of those miRNAs, miR-374a and b, revealed T to C transitions at background level, because uridines at residues 14 and higher can still be crosslinked. The T to C conversion was predominantly concentrated within positions 8 to 13 (FIG. 5D), residing in the singlestranded region of the AGO protein ternary complex (Wang et al., 2008a). Position 1, though predominantly a U, was not crosslinking, presumably because its binding pocket protein environment was not amenable to crosslinking. The miRNA profile derived from AGO-PURE-CLIP miRNA sequence reads was comparable to non-crosslinked miRNA profiles (FIG. 5C), even though one would expect that the uridine composition would have an influence on the miRNA profile. Very interestingly, we noted that many of the more recent additions of miRNAs to Rfam returned with sequence reads devoid of crosslinking evidence despite their presence of uridines at positions 8 to 13, indicating that those precursor sequences and their corresponding miRNAs are unlikely to represent an AGO protein associated RNAs. AGO-PURE-CLIP therefore might represent a novel means for miRNA functional annotation.
Animal miRNAs were shown to primarily recognize their target mRNAs based on sequence complementarity between the 5′ end (the so-called seed region) of the miRNA (see illustrations in FIG. 22), particularly the 7-nt segment from position 2 to 8 of the miRNA, and the target mRNA (Brennecke et al., 2005; Gaidatzis et al., 2007; Grimson et al., 2007; Grun et al., 2005; Krek et al., 2005; Lai, 2002; Lewis et al., 2005; Stark et al., 2003). Independent of any pairing models, we first determined the enrichment of all 16,384 theoretically possible 7-mers within the 17,319 CCRs from the AGO experiments, relative to random sequences with the same dinucleotide composition. The most significantly enriched 7-mers corresponded to the reverse complement of the seed regions 2-8 of the most abundant HEK293 miRNAs, and they were most frequently positioned 1-2 nt downstream of the predominant crosslinking site within CCRs (FIG. 6A). This places the crosslinking site near the centre of the AGO-miRNA-target RNA ternary complex, where the target RNA is proximal to the PIWI/RNase H domain of AGO (Wang et al., 2008a). To further examine the positional dependence of target RNA crosslinking, we anchored the CCRs containing seed complements (defined as A opposite miRNA pos. 1 and match pos. 2-8) of the top 100 expressed miRNAs and plotted the position-dependent frequency of finding a crosslinked position (FIG. 6B). This identified two additional crosslinking regions, which in structural terms correspond to the unpaired 5′ and 3′ ends of the target RNA exiting from the AGO ternary complex. This indicates that the 41-nt windows centered on the predominant crosslink position cover the miRNA-complementary sites.
We then analyzed the occurrence of miRNA-complementary sequences in CCRs. We found 57 CCRs that had long matches (more than 12 nucleotides) to one of the top 100 miRNAs. The most significant enrichment relative to sequences with the same dinucleotide composition was generally obtained with 8-mers that were complementary to miRNA seed regions (defined as above), and a total of 14,809 of the 17,319 CCRs (85%) contained at least one 6-mer miRNA seed-complementary region (FIG. 23). 6-mers still showed some degree of excess conservation in comparative genomics studies (Gaidatzis et al., 2007; Lewis et al., 2005), and thus these results suggest that the majority of the CCRs represent bona fide miRNA binding sites. Furthermore, the number of miRNA seed complements counted for all known miRNAs over all CCRs correlated well with the expression levels of miRNA families found in HEK293 cells, and less well with miRNA profiles of other tissue samples (FIGS. 22-23). We also evaluated whether uridine sequence biases due to the use of 4SU were detectable for the most frequently isolated mRNA target segments. We found no obvious connection between the nucleotide composition of the top 100 most abundant miRNAs and the number of seed complementary sites found in the CCRs (FIG. 26). While the CCRs themselves had a slightly higher U-content (approximately 30% U) as would be expected around miRNA-binding sites, there was no apparent connection between the probability of finding seed complementary sites and the U content of a CCR (Grimson et al., 2007) (FIG. 27).
Structural and biochemical studies of Thermus thermophilus AGO indicated that small bulges and mismatches can be accommodated in the seed pairing region within the target RNA strand (Wang et al., 2008b). We therefore searched for putative non-seed miRNA binding sites, which we defined as binding sites that contained 2 blocks, discontinuous in either the miRNA and/or target, with a total of at least 6 matched nucleotides. We only considered such non-seed putative sites if they were significantly enriched in CCRs compared to dinucleotide randomized sequences. We identified 891 CCRs with mismatches and 256 with bulges in the seed region. Mismatches occurred most frequently across from pos. 5 of the miRNA as G-U or U-G wobbles, U-U mismatches and A-G mismatches, with the A residing in the miRNA. Therefore, it appears that only a small fraction of the miRNA targets that we isolated (less than 6.6%), contained bulges or loops in the seed-region. Another contribution to the specificity of miRNA-target RNA interactions was proposed to be contiguous base-pairing of sequences in the miRNA 3′ half and the target mRNAs (Brennecke et al., 2005; Grimson et al., 2007). We examined the occurrence of 4-nt complementary segments to any of the top 100 miRNAs in the CCRs that contained a match to the 7-mer seed sequences of these miRNAs, and detected a small peak in the frequency for segments corresponding to positions 13-18 of these miRNAs (FIG. 6C).
miRNAs are predominantly localized to the cytoplasm although a smaller fraction can be detected in nuclear extracts (Hwang et al., 2007; Meister et al., 2004; Pena et al., 2009; Robb et al., 2005; Weinmann et al., 2009). Our CCRs correspond to 84% exonic, 14% intronic, and 2% not assigned transcribed regions. This distribution is consistent with the predominant cytoplasmic function of miRNAs. Of the exonic CCRs, 50% distributed to the CDS, 46% to the 3′ UTR, and 4% to the 5′ UTR of mRNAs, with a specific enrichment in the 3′UTRs relative to what is expected based on the relative lengths of different transcript regions (FIG. 6D). The identification of such a large number of miRNA binding sites within the CDS was unexpected, because miRNA are believed to predominantly act on 3′ UTRs (Bartel, 2009), with a few reports providing evidence for miRNA-binding to 5′ UTR or CDS (Forman et al., 2008; Lytle et al., 2007; Orom et al., 2008; Tay et al., 2008). In order to determine if CCR-containing transcripts were indeed subject to miRNA-dependent regulation, we inhibited 25 of the most abundant miRNAs in HEK293 cells (FIG. 5C) with a cocktail of 2′-O-methyl-modified antisense oligoribonucleotides (FIG. 7A and see FIG. 28A-C for the efficiency of the individual antisense oligoribonucleotide). The change in relative mRNA abundance before and after miRNA inhibition was recorded by Affymetrix whole transcriptome array analysis.
Regarding FIG. 28A-C:
2-7 seed let-7a: down p=0.223, up p=0.436
2-8 seed let-7a: down p=0.357, up p=0.364
1-7 seed let-7a: down p=0.292, up p=0.027
1-8 seeds let-7a: down p=0.013, up p=0.076
2-7 seed miR-92a: down p=0.029, up p=0.164
2-8 seed miR-92a: down p=0.010, up p=0.003
1-7 seed miR-92a: down p=0.023, up p=0.242
1-8 seeds miR-92a: down p=0.001, up p=0.000
MiR-27a: Didn't Work.
2-7 seed miR-27a: down p=0.429, up p=0.474
2-8 seed miR-27a: down p=0.281, up p=0.433
1-7 seed miR-27a: down p=0.301, up p=0.396
1-8 seeds miR-27a: down p=0.169, up p=0.392
2-7 seed miR-301b: down p=0.073, up p=0.049
2-8 seed miR-301b: down p=0.032, up p=0.002
1-7 seed miR-301b: down p=0.359, up p=0.378
1-8 seeds miR-301b: down p=0.414, up p=0.042
2-7 seed miR-17: down p=0.227, up p=0.223
2-8 seed miR-17: down p=0.010, up p=0.001
1-7 seed miR-17: down p=0.027, up p=0.006
1-8 seeds miR-17: down p=0.000, up p=0.000
2-7 seed miR-103: down p=0.137, up p=0.244
2-8 seed miR-103: down p=0.067, up p=0.019
1-7 seed miR-103: down p=0.208, up p=0.204
1-8 seeds miR-103: down p=0.337, up p=0.015
2-7 seed miR-21: down p=0.184, up p=0.013
2-8 seed miR-21: down p=0.150, up p=0.010
1-7 seed miR-21: down p=0.280, up p=0.408
1-8 seeds miR-21: down p=0.344, up p=0.013
MiR-186: Didn't Work Well.
2-7 seed miR-186: down p=0.438, up p=0.455
2-8 seed miR-186: down p=0.496, up p=0.258
1-7 seed miR-186: down p=0.330, up p=0.131
1-8 seeds miR-186: down p=0.088, up p=0.335
2-7 seed miR-378: down p=0.008, up p=0.013
2-8 seed miR-378: down p=0.377, up p=0.238
1-7 seed miR-378: down p=0.337, up p=0.261
1-8 seeds miR-378: down p=0.443, up p=0.224
2-7 seed miR-101: down p=0.341, up p=0.014
2-8 seed miR-101: down p=0.281, up p=0.091
1-7 seed miR-101: down p=0.039, up p=0.042
1-8 seeds miR-101: down p=0.094, up p=0.107
2-7 seed miR-15a: down p=0.325, up p=0.292
2-8 seed miR-15a: down p=0.033, up p=0.109
1-7 seed miR-15a: down p=0.370, up p=0.081
1-8 seeds miR-15a: down p=0.108, up p=0.001
2-7 seed miR-18a: down p=0.374, up p=0.223
2-8 seed miR-18a: down p=0.175, up p=0.102
1-7 seed miR-18a: down p=0.093, up p=0.229
1-8 seeds miR-18a: down p=0.099, up p=0.002
MiR-10a: Didn't Work.
2-7 seed miR-10a: down p=0.023, up p=0.202
2-8 seed miR-10a: down p=0.298, up p=0.339
1-7 seed miR-10a: down p=0.328, up p=0.472
1-8 seeds miR-10a: down p=0.345, up p=0.207
2-7 seed miR-30a: down p=0.092, up p=0.013
2-8 seed miR-30a: down p=0.134, up p=0.011
1-7 seed miR-30a: down p=0.037, up p=0.046
1-8 seeds miR-30a: down p=0.012, up p=0.001
2-7 seed miR-19a: down p=0.463, up p=0.010
2-8 seed miR-19a: down p=0.037, up p=0.001
1-7 seed miR-19a: down p=0.001, up p=0.063
1-8 seeds miR-19a: down p=0.009, up p=0.000
First, we examined the change in stability for target RNAs that contain at least one CCR, as a function of the length of the seed-complementary region (FIG. 7B). Consistent with previous studies (Crimson et al., 2007), the magnitude of the destabilization effects dropped from 9-mer, to 8-mer to 7-mer to 6-mer matches. Transcripts harboring CCRs that did not contain a 6-mer match to the antagonized miRNAs were not significantly stabilized, suggesting only limited regulation of transcripts with no canonical seed site.
Second, we examined the change in stability of CCR-containing transcripts as a function of the number of binding sites, and found that transcripts containing more than one CCR were more efficiently destabilized than transcripts containing a single CCR (FIG. 7C). We also found that multiple binding sites for highly expressed miRNAs can co-occur within a single CCR (FIG. 27). Both of these findings were supported by previous observations (Grimson et al., 2007).
Third, we asked whether transcripts harboring CCRs exclusively in the CDS exhibited a similar level of regulation compared to transcripts with CCRs exclusive to the 3′UTRs; the number of 5′UTR-localized seed-match-containing CCRs was too small to allow for a statistically significant analysis of miRNA-dependent regulation. Transcripts with sites exclusively in the CDS were subject to a statistically significant miRNAdependent destabilization, albeit a lot less pronounced compared to the destabilizing effect caused by sites located in the 3′UTR (FIG. 7D), independent of whether they contained extensive seed matches to abundant miRNAs or not (FIG. 30). It was recently reported that miRNA targeting in the CDS was enhanced by inserting rare codons upstream of the miRNA-binding site, presumably by stalling the ribosomes and increasing the lifetime of miRNA-target-RNA interactions (Gu et al., 2009). We therefore examined the codon usage around crosslinked seed matches by calculating the codon adaptation index (Sharp and Li, 1987) and found that it was different from the average codon usage around non-crosslinked seed matching CDS regions. The bias in codon usage extended at least 70 codons up- as well as downstream of the crosslinked seed matches (FIG. 7E). It is conceivable that this bias led to a reduced translational efficiency and higher stability of interaction between the CDS of these transcripts and miRNA-AGO complexes. In summary, the miRNA inhibition studies provided evidence that CCRs represent functionally active sites and that the magnitude of miRNA-induced mRNA destabilization depends on the strength of base-pairing between miRNA and transcript, the number of interaction sites, and the position of these sites within the transcript.
The abundance of mRNA species within cells varies over several orders of magnitude (Bishop et al., 1974). We determined the relative abundance of HEK293 transcripts by DGE profiling. Based on 3.9 M reads annotated as mRNA, we detected 22,466 expressed transcripts with expression levels varying over 5 orders of magnitude. When we related the expression level of CCR-containing transcripts with the magnitude of transcript stabilization after miRNA inhibition, we found that miRNAs preferentially act on transcripts with low and medium expression levels (FIG. 7F). Highly expressed mRNAs appear to avoid miRNA regulation (Farh et al., 2005; Sood et al., 2006; Stark et al., 2005), at least for those miRNAs expressed in HEK293 cells.
Earlier studies that revealed the destabilizing roles of miRNAs were carried out by transfection of individual cell-type specific miRNAs into cellular systems devoid of these particular miRNAs (Baek et al., 2008; Burchard et al., 2009; Lim et al., 2005; Selbach et al., 2008). To allow for a comparison to such studies, we transfected miRNA duplexes corresponding to the deeply conserved miR-7 and miR-124 into HEK293 cells stably expressing FLAG/HA-AGO2 and performed PURE-CLIP. Before transfection, these miRNAs were only present in trace amounts in HEK293 cells, less than 0.2% of total miRNA abundance. We found 1,866 unique sequence read clusters in the miR-7 and 10,929 in the miR-124 transfection experiments. The CCRs isolated from miR-7 and miR-124 transfection experiments were enriched for the seed sequences of the transfected miRNA duplexes. These CCRs were predominantly exonic (87 and 88%, respectively), 28% mapped to the CDS, 57 and 59%, respectively, to the 3′UTR of mRNAs and only 3 and 2%, respectively, to the 5′UTR. miR-7 and miR-124 binding sites that were located in the CDS for miR-7 and -124 were also embedded in regions with unusual codon bias, similar to the binding sites of endogenous miRNA target sites (FIG. 31).
The change in mRNA stability was analyzed after miRNA duplex transfection by Affymetrix whole genome microarrays (Hausser et al., submitted). We observed that transcripts containing miR-7- or miR-124-specific CCRs were destabilized upon miRNA delivery, especially for transcripts with CCRs located in the 3′UTR (FIG. 32-33). Therefore, analysis of data obtained with endogenous or transfected miRNAs yielded similar insights into miRNA targets and their regulation.
miRNA-target interactions are predominantly driven by seed interactions, however, not every seed-complementary sequence in the HEK293 transcriptome represented an experimental miRNA target site. We therefore tested whether CCRs and non-crosslinked seed-complementary sites differ in the ElMMo selection pressure (Gaidatzis 2007), the TargetScanS mRNA stability scores (Grimson et al., 2007), as well as the sequence and structural elements around the miRNA seed complementary site. The feature that distinguished most CCR sites from non-crosslinked sites was the free energy required to resolve local secondary structure involving the miRNA-binding region. The free energy for CCRs was on average 25% lower than that of non-crosslinked sites (FIG. 34). This difference in the stability of the secondary structure for CCRs is also associated with a 6% increase in the A/U content from within 100 nts around the seedpairing site. These properties hold for seed-complementary regions located in the CDS as well as in the 3′UTRs. Compared to non-crosslinked sites, CCRs are under stronger evolutionary selection (ElMMo) and in sequence contexts favoring degradation (TargetScanS).
We also compared our large datasets to computationally predicted interactions that involved at least a 7-mer subsequence (1-7 or 2-8) of miRNA seed regions and 3′UTRs of mRNAs. The rate of crosslinking of high-confidence target sites predicted by ElMMo (Gaidatzis et al., 2007) for the most abundant endogenous miRNA families (miR-15, miR-19, miR-103, miR-20 and let-7, see Supplementary Methods) and miR-7 and -124 was 8 to 15%. On the other hand, similar proportions (8-15%) of the seed-containing CCR were among low-confidence ElMMO predictions. Similar numbers were obtained using the TargetScanS (Grimson et al., 2007) target prediction tool. The apparent over-prediction of miRNA target sites may have several reasons, e.g. binding to other mRNA-binding proteins or the involvement of the putative target sites in RNA secondary structure, interactions that may change as a function of cellular context.
Consistent with the hypothesis that some of the predicted target sites may function in different cellular contexts that were not covered in our experiments. Consistent with this hypothesis, we found that high-confidence predicted target sites that were crosslinked resided in transcripts with higher expression (25-60% higher average expression depending on the experiment in HEK293 cells compared to high-confidence predicted target sites that were not crosslinked.
Finally, the relative position of CCRs is non-random. The CCRs containing 7-mer or 8-mer sites that are complementary to endogenous miRNAs as well as to miR-7 and miR-124 and residing in transcripts with relatively long 3′UTRs (more than 3 kb) were preferentially located near the stop codon or the polyA tail (FIG. 35). The CCRs located in the CDS showed a bias towards positions near the stop codon for the transfected miR-7 and 124, but not for the endogenous miRNAs (FIG. 36), again possibly reflecting a bias for highly efficient targeting sites in transfection experiments.
Maturation, localization, decay and translational regulation of mRNAs involves RBPs that form RNP complexes with their RNA targets (Komili and Silver, 2008; Martin and Ephrussi, 2009; Moore and Proudfoot, 2009). The human genome encodes several hundred RBPs, many of which consist of multiple repeats of a small set of RNA-binding domains that generate the diverse structural requirements for target RNA recognition (Lunde et al., 2007). Two of the most abundant domains are the RRM and KH domains, both of which recognize about 3- to 5-nucleotide (nt) long single-stranded RNA sequence motifs. Proteins containing these domains are involved in all aspects of posttranscriptional gene regulation and are also linked to a number of diseases (Keene, 2007).
Pumilio proteins represent another conserved group of RBPs repressing translation and/or enhancing mRNA decay and their RRMs comprise 8-10 nt of single-stranded RNA (Wickens et al., 2002). Furthermore miRNAs, hundreds of which are encoded in mammalian genomes (Landgraf et al., 2007), function in the context of AGO and TNRC6 proteins to destabilize target mRNAs and/or repress their translation, and they require 6-8 nt complementarity between the miRNA 5′ end (seed) and their mRNA targets (Bartel, 2009; Filipowicz et al., 2008). Collectively, it is this dynamic association of RBPs and miRNPs with mRNAs that constitutes the code for posttranscriptional gene regulation, which we aim to elucidate. Until now, one of the limitations towards this goal was the lack of methods that directly identify RBP mRNA targets and binding sites with a clear separation of “true” versus “false” sites, while avoiding possible reorganization of RBP27 and RNP-target-RNA networks in the process (Mili et al., 2008).
Utilizing 4SU effectively separates true binding sites from background. We showed that growing cultured cells in tissue culture medium supplemented by a photoreactive nucleoside enabled effective long-wavelength UV crosslinking of RNA to interacting proteins, thereby freezing even transient or low affinity interactions by covalent bond formation. We concentrated on the application of 4SU after it became apparent that the crosslink sites in isolated RNAs were revealed upon sequencing by a prominent transition from T to C in the cDNA prepared from the isolated crosslinked RNA segments. Regular UV 254 crosslinking in the absence of photoreactive nucleosides not only yielded less RNA in our hands, but more importantly, the location of the crosslink was not readily apparent by a mutational pattern. Studies using conventional 254 nm CLIP have not reported exploitable incidences of deletions and mutations (Chi et al., 2009; Licatalosi et al., 2008; Ule et al., 2003). A recent publication that focused on RBPs interacting exclusively with one RNA, the U3 snoRNA, showed that there was an increased incidence of deletions and substitutions around the single binding site (Granneman et al., 2009), however it remains unclear whether these mutations can be exploited on a transcriptome-wide level. Our identification of 4SU crosslink sites by frequent T to C transitions was unexpected and points to a distinct structural change of the crosslinked nucleobase likely resembling a cytidine analog that directs the reverse transcriptase to incorporate its cognate nucleotide partner distinct from the starting nucleotide. Standard UV 254 crosslinking, which occurs by a different mechanism, presumably yields base structures unable to induce a specific sequence change at scorable frequency. Irrespective of the UV crosslinking method applied, one must assume that amino-acid-modified nucleotides represent a severe block for reverse transcription, and that non-crosslinked RNA, always present as background after purification of crosslinked RNA segments, has a significant advantage over crosslinked RNA. It is therefore important to have a method that now provides an easy means to separate signal from noise.
In order to decipher with high-resolution crosslink sites by sequencing, they have to fall into the 32- to 36-nt Solexa sequencing read windows. This was accomplished by stringent RNase T1 digestion of the immunopurified RNPs to recover RBAs within a 20-40 nt RNA size window. Overdigestion of the RNA to size fragments too small to map uniquely to the genome was not a serious issue, because RNase T1 only cleaves after guanosines and the crosslinked RBP or miRNP also protects the bound target RNA. Nevertheless, it is useful to point out that some clusters of sequence reads arising from regions devoid of guanosines over a longer distance from the crosslinking sites must have received T to C transition scores similar to background in the current analysis pipeline. Using less specific nucleases, such as RNase A, may remedy the detection of these subsets of clusters, however, fine-tuning of RNase treatment may be technically more challenging.
Finally, the PURE-CLIP approach allows the definition of binding sites of cytoplasmic as well as nuclear localized proteins and RNPs, as shown by the identification of intronic clusters of sequence reads, reaching up to 70% for QKI. This is an important finding and indicates that the method is suitable to also study emerging new RNA regulatory processes closely linked to transcription (Kapranov et al., 2007).
The crosslinking efficiency of 4SU appears to be context-dependent and affected by the local amino acid environment of the RBP. Therefore, the number of sequence reads for a given binding site is a complicated sum of not only the cellular abundance of target mRNAs and the affinity and association and dissociation rates of the RBP, but also the sequence composition affecting crosslinking efficiency. Nevertheless, by the depth of sequencing one can accomplish today, it is feasible to identify the majority of binding sites at a transcriptome level. An increasing understanding of copy numbers of RNA transcripts and RBPs and miRNAs inside cells, combined with further biochemical binding studies, will eventually yield a more quantitative description of RNA recognition processes.
Certain extreme types of binding sites or sequence environment may be difficult to be fully covered by using solely 4SU. To assess the possible magnitude of this potential problem, we determined that only about 0.4% of 32-nt long sequence segments present in the entire transcriptome are actually devoid of uridines, or in other words, an 8-kb long mRNA would on average contain only one 32 nt long U-less segment (FIG. 19). To provide a means to resolve such unlikely situations, the use of other photoreactive nucleosides, such as 6SG, was explored. We found that IGF2BP1 sites identified by use of 6SG overlapped substantially with those from using 4SU, despite of the fact that the environment of IGF2BP1 binding sites was strongly depleted for guanosine (FIG. 20A-C). Furthermore, the sites of crosslinking using 6SG were enriched for G to A transitions, pointing to an interesting structural change in 6SG upon crosslinking presumably mimicking an adenosine analog. The only drawback for using 6SG is its higher cytotoxicity noticeable above 100 μM at exposure longer than 24 hrs in culture medium (Melvin et al., 1978), which is at the threshold of the concentrations we have used in our studies. We therefore recommend to first use 4SU and then possibly resort to use 6SG when sequence contexts were encountered that are avoiding uridines over long sequence regions.
When applying PURE-CLIP to isolate mRNA segments bound by endogenously expressed or additionally co-transfected miRNAs, we were surprised by the extent of binding sites present in the CDS, approaching nearly 50%. miRNA inhibition experiments showed that miRNA binding at these sites also caused mRNA destabilization, albeit to a lesser extent than binding sites located in the 3′UTR. The CDS miRNA binding sites showed an increased incidence of rare codon usage, presumably reducing translational efficiency and opening a window for transient miRNP binding. Similar observations were made previously using artificially designed reporter systems (Gu et al., 2009). A broad distribution of binding sites across CDS and 3′UTR were also seen for IGF2BP or QKI protein binding sites, however for PUM2 protein, which has a consensus site length similar to miRNAs, the sites were almost exclusively restricted to 3′UTRs. The reason for this striking difference is unclear, presumably different biochemical mechanisms and evolutionary selection of target RNA regulation must be at play.
High-resolution knowledge of the crosslinking site allows us to narrowly define the miRNA binding regions for matching the site with the most likely miRNA endogenously co-expressed with its targets, and assess non-canonical miRNA binding modes. This approach is distinct from relying on matching predicted miRNA binding sites to a broad peak of clusters (Chi et al., 2009). We were able to explain the majority of PURE-CLIP binding sites by conventional miRNA-mRNA seed-pairing interactions (Grimson et al., 2007), yet found that about 6% of miRNA target sites might best be explained by accepting bulges or mismatches in the seed pairing region, similar to the interaction between let-7 and its target lin-41 (Vella et al., 2004) and those recently observed in biochemical and structural studies of T. thermophilus AGO protein (Wang et al., 2008a).
Finally, mutational analysis of miRNAs in AGO PURE-CLIP also showed the expected T to C transition, but not for all sequences currently proposed to be miRNAs. In part, these differences are due to the absence of U within the regions of the miRNA effectively crosslinking to AGO proteins, but others do not show such biases and therefore likely represent RNAs that were misclassified as miRNAs. Consequently, we now have a new method that is able to validate putative miRNAs acquired from deep sequencing approaches. It is important to note that all of our PURE-CLIP libraries, also those unrelated to AGO or TNRC6 proteins contain between 10 and 30% miRNAs, yet only AGO libraries do contain T to C changes in their sequence reads. This is further testament for the importance of separating signal from noise using crosslinking approaches to isolate and identify bound RNA segments.
It was surprising to discover that each of the examined RBP or miRNPs control a substantial (one to seven thousand) fraction of transcripts out of the possible 20,000 detectable in HEK293 cells. These numbers illustrate that transcripts are mostly under combinatorial control, and that simple regulatory networks cannot be deduced from solely determining binding sites for a single RBP. This discovery also implies that highresolution mapping at a nucleotide level of RBP RNA interaction is imperative as many of these sites are adjacent to each other. It will therefore be critical to expand the PURECLIP approaches to all other RBPs and also transiently interacting RNA-binding factors, such as helicases, nuclease or polymerases, to ultimately enable us to understand how the sum or a particular combination of these interactions ultimately controls the fate of each mRNA. To assess the impact of post-transcriptional regulation, we can now begin to interpret the rapidly emerging data on genetic variation between individuals and how these variations contribute to complex genetic diseases.
Oligonucleotides and siRNA Duplexes
The following oligodeoxynucleotides were used for PCR and cDNA cloning into pENTR4 (Invitrogen), restriction site are underlined:
The following oligoribonucleotides were used for QKI protein in vitro binding and crosslinking studies and were purchased from Dharmacon:
The following siRNA duplexes (sense/antisense) were used for knockdown experiments and synthesized on a modified ABI 392 RNA/DNA synthesizer using Dharmacon synthesis reagents.
2′-O-Methyl Oligoribonucleotides and miRNA Duplexes
The following sequences were chemically synthesized on an ABI394 RNA/DNA synthesizer using 5′silyl/2′orthoester chemistry (Dharmacon):
Plasmids pENTR4 IGF2BP1, 2, 3, QKI, AGO1-4, TNRC6A-C and PUM2 were generated by PCR amplification of the respective coding sequences (CDS) followed by restriction digest with SalI and NotI and ligation into pENTR4 (Invitrogen). pENTR4 IGF2BP1, -2, and -3 were recombined into pFRT/TO/FLAG/HA-DEST destination vector (Invitrogen) using GATEWAY LR recombinase according to manufacturer's protocol (Invitrogen) to allow for doxycycline-inducible expression of stably transfected FLAG/HA-tagged protein in Flp-In T-REx HEK293 cells (Invitrogen) from the TO/CMV promoter. pENTR4 QKI and pENTR4 PUM2 were recombined into pFRT/FLAG/HA4 DEST for constitutive expression in Flp-In T-REx HEK293 cells.
Plasmids for bacterial expression of N-terminally His6-tagged IGF2BP1, 2, and 3 in E. coli were generated by ligation of CDS into pET16 (Novagen). The plasmid for bacterial expression of N-terminally His6-tagged QKI was generated by LR recombination of pENTR4 QKI with pDEST17 (Invitrogen). The plasmids described in this study can be obtained from Addgene.
Polyclonal rabbit antibodies against IGF2BP1, 2, and 3 were generated by injection of synthetic peptides corresponding to positions aa 561-573, 264-275, and 567-579, respectively. Rabbit anti-QKI (BL1040) was purchased from Bethyl Laboratories.
Recombinant Protein Expression and Purification
pET16 IGF2BP1, -2, and -3 and pDEST17-QKI plasmids, encoding an N-terminal His6-tag, were transformed in E. coli STAR(DE3) (Invitrogen). Cells were grown in LB medium supplemented with 50 μg/ml ampicillin at 37° C. to A600=0.6. The cells were cooled to 25° C., protein synthesis was induced by addition of IPTG to a final concentration of 1 mM, cells were harvested 3 h later. The cell pellet was resuspended in 10 ml lysis buffer (50 mM Tris-HCl pH 8.0, 300 mM KCl, 5 mM MgCl2, 0.1% Triton X-100, and complete EDTA-free protease inhibitor (Roche)) per gram cell pellet. All the following steps were carried out at 4° C. Cells were resuspended in lysis buffer and incubated with 1 mg/ml lysozyme for 30 min and sonicated to reduce viscosity. Insoluble material was removed by centrifugation at 12,000×g for 20 min. For His-tag affinity selection, the supernatant was incubated with 250 μl HIS-Select Cobalt Affinity Gel (Sigma) per 10 ml cell supernatant for 1 h. The gel was washed three times with 10 gel volumes of wash buffer (50 mM Tris-HCl, pH 8.0, 300 mM KCl, 5 mM MgCl2, 1 mM DTT, 0.1% Triton X-100, 25 mM imidazol, and complete EDTA-free protease inhibitor (Roche)). His-tagged proteins were eluted in 3 gel volumes of elution buffer (50 mM Tris-HCl pH 8.0, 300 mM KCl, 5 mM MgCl2, 1 mM DTT, 0.1% Triton X-100, 250 mM imidazol, and complete EDTA-free protease inhibitor (Roche)). The eluted proteins were applied to a Heparin column equilibrated in 20 mM Tris-HCl pH 7.8, 5 mM MgCl2, 100 mM KCl, 1 mM DTT, 0.1% Triton X-100, 10% glycerol. Proteins were eluted with a KCl gradient (0.5-1.5 M) in 20 mM Tris-HCl, pH 7.8, 5 mM MgCl2, 1 mM DTT, 0.1% Triton X-100, 10% glycerol. His6-IGF2BP1, -2, and -3 eluted at 550 to 650 mM KCl and His6-QKI at 1.1 M KCl.
Electrophoretic Mobility-Shift Analysis
Radiolabeled RNA (100 pM) was incubated with recombinant His6-IGF2BP2 protein at indicated concentrations and 100 ng tRNA in 20 μl of 20 mM Tris-HCl, pH 7.8, 140 mM KCl, 2 mM MgCl2 and 0.1% Triton X-100 at 30° C. for 1 h. After addition of 6 μl loading dye (40% glycerol, bromophenol blue in binding buffer), the solution was loaded onto a native 6% acrylamide gel containing 0.5×TBE, running at 200 V for 1 h at room temperature, using 0.5×TBE as running buffer.
Radiolabeled RNA (1 nM) was incubated with recombinant His6-QKI protein at various concentrations and 100 ng tRNA in 20 μl of 20 mM HEPES-KOH, pH 7.4, 330 mM KCl, 10 mM MgCl2, 0.1 mM EDTA and 0.01% IGEPAL CA630 (Sigma). After addition of 6 μl loading dye (40% glycerol, bromophenol blue in binding buffer), the solution was loaded onto a native 10% acrylamide gel containing 0.5×TBE, running at 200 V for 2 h at room temperature, using 0.5×TBE as running buffer. The protein-bound RNA and the free RNA were quantified using a phosphorimager.
Cell Lines and Culture Conditions
HEK293 T-REx Flp-In cells (Invitrogen) were grown in D-MEM high glucose (1×) with 10% (v/v) fetal bovine serum, 1% (v/v) 2 mM L-glutamine, 1% (v/v) 10,000 U/ml penicillin/10,000 μg/ml streptomycin, 100 μg/ml zeocin and 15 μg/ml blasticidin. Cell lines stably expressing FLAG/HA-tagged proteins were generated by co-transfection of pFRT/TO/FLAG/HA or pFRT/FLAG/HA constructs with pOG44 (Invitrogen). Cells were selected by exchanging zeocin with 100 μg/ml hygromycin. Expression of FLAG/HA-IGF2BP1, -2, -3 was induced by addition of 250 ng/ml doxycycline 15 to 20 hrs before crosslinking
SiRNA Transfection and mRNA Affymetrix Array Analysis
siRNA transfections of HEK293 T-REx Flp-In cells were performed in 6-well format using Lipofectamine RNAiMAX (Invitrogen) as described by the manufacturer. Total RNA of transfected cells was extracted using TRIZOL following the instructions of the manufacturer. The RNA was further purified using the RNeasy purification kit (Qiagen). 2 μg of purified total RNA was used in the One-Cycle Eukaryotic Target Labeling Assay (Affymetrix) according to manufacturer's protocol. Biotinylated cRNA targets were cleaned up, fragmented, and hybridized to Human Genome U133 Plus 2.0 Array (Affymetrix). For details of the analysis, see Bioinformatics section.
miRNA Duplex Transfection and mRNA Affymetrix Array Analysis
miRNA duplexes and 2′-O-methyl oligoribonucletide transfections of HEK293 TREx Flp-In cells were carried out using Lipofectamine RNAiMAX (Invitrogen). Total RNA was isolated by TRIZOL RNA extraction. The RNA was further purified and Human Genome U133 Plus 2.0 Array (Affymetrix) analyses were performed as described previously (Landthaler et al., 2008).
miRNA RNA Profiling
miRNAs were extracted from FLAG/HA-AGO immunoprecipitates as described in Meister et al. (Meister et al., 2004). miRNAs from immunoprecipitate and total RNA were cloned and Solexa-sequenced (Hafner et al., 2008) using following bar-coded 5′ adapters:
UV 254 and 365 nm Crosslinking
For UV crosslinking, cells were washed once with ice-cold PBS while still attached to the plates. PBS was removed completely and cells were irradiated on ice with 254 nm UV light (0.15 J/cm2), or 365 nm UV light for cells treated for 14 hrs with 100 μM nucleoside analogues (0.15 J/cm2) in a Stratalinker 2400 (Stratagene), equipped with light bulbs for the appropriate wavelength. Cells were scraped off with a rubber policeman in 1 ml PBS per plate and collected by centrifugation at 500×g for 5 min.
Cell Lysis and First Partial RNase T1 Digestion
The pellets of UV365-crosslinked cells were resuspended in 3 cell pellet volumes of NP40 lysis buffer (50 mM HEPES, pH 7.5, 150 mM KCl, 2 mM EDTA, 1 mM NaF, 0.5% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and incubated on ice for 10 min. The typical scale of such an experiment was 3 ml of cell pellet. The cell lysate was cleared by centrifugation at 13,000×g. RNase T1 (Fermentas) was added to the cleared cell lysates to a final concentration of 1 U/μl and the reaction mixture was incubated in a water bath at 22° C. for 15 min and subsequently cooled for 5 min on ice before addition of antibody-conjugated magnetic beads.
Immunoprecipitation and Recovery of Crosslinked Target RNA Fragments
Preparation of Magnetic Beads
10 μl of Dynabeads Protein G magnetic particles (Invitrogen) per ml cell lysate were washed twice with 1 ml of citrate-phosphate buffer (4.7 g/l citric acid, 9.2 g/l Na2HPO4, pH 5.0) and resuspended in twice the volume of citrate-phosphate buffer relative to the original volume of bead suspension. 0.25 μg of anti-FLAG M2 monoclonal antibody (Sigma, F9291) per ml suspension was added and incubated at room temperature for 40 min. Beads were then washed twice with 1 ml of citrate-phosphate buffer to remove unbound antibody and resuspended again in twice the volume of citrate-phosphate buffer relative to the original volume of bead suspension.
Immunoprecipitation (IP), Second RNase T1 Digestion, and Dephosphorylation
10 μl of freshly prepared antibody-conjugated magnetic beads per ml of partial RNase T1 treated cell lysate were added and incubated in 15 ml centrifugation tubes on a rotating wheel for 1 h at 4° C. Magnetic beads were collected on a magnetic particle collector (Invitrogen). Manipulations of the following steps were carried out in 1.5 ml Eppendorf tubes. The supernatant was removed from the bead-bound material. Beads were washed 3 times with 1 ml of IP wash buffer (50 mM HEPES-KOH, pH 7.5, 300 mM KCl, 0.05% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and resuspended in one volume of IP wash buffer. RNase T1 (Fermentas) was added to obtain a final concentration of 100 U/μl, and the bead suspension was incubated in a water bath at 22° C. for 15 min, and subsequently cooled for 5 min on ice. Beads were washed 3 times with 1 ml of high-salt wash buffer (50 mM HEPES-KOH, pH 7.5, 500 mM KCl, 0.05% (v/v) NP40, 0.5 mM DTT, complete EDTA-free protease inhibitor cocktail (Roche)) and resuspended in one volume of dephosphorylation buffer (50 mM Tris-HCl, pH 7.9, 100 mM NaCl, 10 mM MgCl2, 1 mM DTT). Calf intestinal alkaline phosphatase (NEB) was added to obtain a final concentration of 0.5 U/μl, and the suspension was incubated for 10 min at 37° C. Beads were washed twice with 1 ml of phosphatase wash buffer (50 mM Tris-HCl, pH 7.5, 20 mM EGTA, 0.5% (v/v) NP40) and twice with 1 ml of polynucleotide kinase (PNK) Buffer (50 mM Tris-HCl, pH 7.5, 50 mM NaCl, 10 mM MgCl2, 5 mM DTT). Beads were resuspended in one original bead volume of PNK Buffer.
Radiolabeling of RNA Segments Crosslinked to Immunoprecipitated Proteins
To the bead suspension described above, γ-32P-ATP was added to a final concentration of 0.5 μCi/μl and T4 PNK (NEB) to 1 U/μl in one original bead volume. The suspension was incubated for 30 min at 37° C. Thereafter, non-radioactive ATP (Fermentas) was added to obtain a final concentration of 100 μM and the incubation was continued for another 5 min at 37° C. The magnetic beads were then washed 5 times with 800 μl of PNK Buffer and resuspended in 70 μl of SDS-PAGE Loading Buffer (10% glycerol (v/v), 50 mM Tris-HCl, pH 6.8, 2 mM EDTA, 2% SDS (w/v), 100 mM DTT, 0.1% Bromophenol blue).
SDS-PAGE and Electroelution of Crosslinked RNA Protein Compounds from Gel Slices
The radiolabeled bead suspension was incubated for 5 min at 95° C. and vortexed. The magnetic beads were separated on a magnetic separator and 40 μl of supernatant were loaded per well of an SDS-PAGE. The gel was analyzed by phosphorimaging. The radioactive band migrating at the expected molecular weight of the target protein was excised from the gel and electroeluted in a D-Tube Dialyzer Midi (Novagen) in 800 μl SDS running buffer according to the instructions of the manufacturer.
Proteinase K Digestion
An equal volume of 2× Proteinase K Buffer (100 mM Tris-HCl, pH 7.5, 150 mM NaCl, 12.5 mM EDTA, 2% (w/v) SDS) with respect to the electroeluate was added, followed by the addition of Proteinase K (Roche) to a final concentration of 1.2 mg/ml, and incubation for 30 min at 55° C. The RNA was recovered by acidic phenol/chloroform extraction followed by a chloroform extraction and an ethanol precipitation. The pellet was dissolved in 10.5 μl of water.
cDNA Library Preparation and Deep Sequencing
The recovered RNA was carried through a cDNA library preparation protocol originally described for cloning of small regulatory RNAs (Hafner et al., 2008). The first step, 3′ adapter ligation, was carried out as described on a 20 μl scale using 10.5 μl of the recovered RNA. UV 254 nm crosslinked RNAs were processed using standard adapter sets, followed by PCR to introduce primers compatible with 454 sequencing; UV 365 nm crosslinked sample RNAs were processed using Solexa sequencing adapter sets. Depending on the amount of RNA recovered, 5′-adapter-3′-adapter products without inserts may be detected after amplification of the cDNA as additional PCR band. In such case, the longer PCR product of expected size was excised from a 3% NuSieve lowmelting point agarose, eluted from the gel pieces with the Illustra GFX-PCR purification kit (GE Healthcare) and Solexa sequenced.
Generation of Digital Gene Expression (DGEX) Libraries
1 μg each of total RNA from HEK293 cells inducibly expressing tagged IGF2BP1 before and after induction was converted into cDNA libraries for expression profiling by sequencing using the DpnII DGE kit (Illumina) according to instructions of the manufacturer. For details of the analysis, see Bioinformatics section.
Adapter Removal and Sequence Annotation
The basic method for removing adaptors and assigning a functional annotation to the tags was described in (Berninger et al., 2008). Briefly, we used an in-house free ends local alignment algorithm (score parameters: 2 for match, −3 for mismatch, −2 for gap opening, −3 for gap extension) to align the Solexa adapter to the 3′ end of each tag, allowing for the possibility that the adapter was not completely sequenced. We then discarded sequences that were either too short (less than 20 nc) or too repetitive (using a cutoff of 0.7 and 1.5 in the entropy of the mono- and dinucleotide distributions, respectively, of individual tags (Berninger et al., 2008)) and would probably map to multiple genomic locations. We mapped the remaining sequences to the hg18 version of the human genome assembly that we downloaded from the University of California at Santa Cruz and to a database of sequences whose function (rRNA, tRNA, sn/snoRNA, miRNA, mRNA, etc.) is already known. These were obtained from the sources specified in (Berninger et al., 2008). We used the oligomap algorithm (Berninger et al., 2008) for this purpose, and we obtained all the perfect and 1-error (mismatch or indel) mappings. Based on the GMAP (Wu and Watanabe, 2005) genome mapping of human mRNA transcripts from NCBI downloaded on Nov. 4, 2008, we determined whether the tags mapped to intronic or exonic regions of genes. Based on the coding region annotation of transcripts in GenBank, we determined whether the exonic tags originated from the 5′ UTR, CDS or 3′ UTR.
Generation of Clusters of Mapped Sequence Reads
For subsequent analyses we only used sequence tags of at least 20 nt in length and mapping uniquely to the genome, with at most 1 error. We performed single-linkage clustering of the tags, with two tags being placed in the same cluster if they overlapped by at least one nucleotide in their genomic mappings. Each cluster was then annotated based on the functional annotation of tags that covered most of the cluster length. We then considered all the mRNA-annotated clusters containing at least 5 mRNA-annotated tags, and we defined a scoring scheme to identify the clusters that had the highest probability of being real crosslinking sites (see below: Identification of high confidence clusters).
Analysis of Mutational Spectra
From the clusters defined above, we used all tags that mapped uniquely and with one error (mismatch or indel) to the genome to infer the mutational bias of the method. For each library, we calculated the proportion of mutations involving each of the four nucleotides as well as the proportion of each of the four nucleotides in the CLIPed tags (see Supplementary FIG. 3).
Identification of High-Confidence Clusters
We used the CLIPed clusters of PUM2 and QKI, to define criteria for selecting high confidence binding sites. The criteria that we tested reflected the mechanistic aspects of generating the sequence reads. They were the absolute sequence read counts and the presence of a guanosine in the genomic sequence directly upstream of the sequence read cluster, because RNase T1 cleaves after G nucleotides, and criteria reflecting our preliminary analysis of mutation patterns in the tags. For each cut-off on a given property, we calculated an enrichment of binding sites, which is defined as the fraction of clusters with at least one binding site above the given cut-off divided by the fraction of clusters with no crosslinking sites (no T to C mutations) that have at least one binding site. For Pumilio, there are 1020 clusters with no T to C mutation and 27% of these clusters contain the binding motif. For QKI, there are 1482 clusters with no T to C mutation, 35% of which contain the consensus motif. Our preliminary analysis revealed that T to C mutations are by far the most frequently observed mutations in these data sets, and that they are most frequent inside or in the immediate vicinity of the binding motifs as opposed to the rest of the sequence (see FIGS. 2D, 3D, and 4D). This suggested that the observed mutational bias is directly linked to the crosslinking event and should thus be a good criterion for separating true crosslinked sites from background tags. The preliminary analysis also indicates a strong bias for G nucleotides at the last position of a tag and also at the genomic position immediately upstream of a tag. This bias reflects the sequence specificity of the RNase T1, and may again help in the identification of tags that map to multiple sites or for discriminating random RNA turnover products unrelated to RNase T1 treatment. Finally, we observed that many clusters with abundantly sequenced tags contained more than one position with a T to C mutation. The results of testing these criteria for their ability to select clusters that contained the known binding motif are shown in Supplementary FIG. 5. For QKI, binding motifs were defined as occurrences of ACUAA or AUUAA. For Pumilio, in order to account for additional motif variants besides the consensus UGUANAUA, binding motifs were identified as matches to the weight matrix (as inferred by MotEvo) that resulted from the motif search (see below). We found that ranking of the clusters by the number of T to C mutations in all tags in the clusters of sequence reads leads to the strongest enrichment in clusters with a binding site (Supplementary FIG. 5). The figures show the fraction of the CLIPed clusters that contain at least one occurrence of the known binding motif as a function of the number of clusters that passed a given cutoff in the selection criterion (e.g. total number of tags, total number of T to C mutations, total number of tags with a G at position −1 in relative to their genomic locus). It is clear, particularly for PUM2, that the number of T to C mutations strongly correlates with the presence/absence of the motif in the cluster. For comparison, we also show the same plots when using as the ranking criterion not the total number of T to C mutations in the cluster, but just the total number of tags per cluster. For QKI, this leads to a significantly lower enrichment of clusters with recognition elements. We also investigated how the fraction of clusters with the known binding motif depends on the number of distinct crosslinking positions (i.e. positions with at least one T to C mutation) inside the cluster (Supplementary FIG. 5). The fraction of clusters with a binding site increases steadily from 0 to 5 crosslinking positions for both proteins, with the strongest increase from 0 to 1 for PUM2 and between 0 and 2 crosslinking positions for QKI. When requiring that at least two positions with T to C mutations are present in the cluster, the fraction of clusters with a binding site increases roughly by 10% for PUM2, and by more than 40% for QKI, These considerations lead us to the following procedure for defining high confidence clusters for any given RBP. 16 We first selected all the clusters with at least two crosslinking positions and, secondly, within this subset, we ranked all clusters by the total number of T to C mutations in all tags in the cluster.
Extraction of Peaks and T to C-Anchored Regions from Sequence Read Clusters
From each ranked, mRNA-annotated cluster a peak region, defined as a 32-nt long region with the highest average tag density, was extracted. Because the T to C mutation was diagnostic for the site of crosslinking, we focused our motif analysis on regions anchored at the position in a cluster with the most T to C mutations. We then investigated the mutational profile around this position and we found that this profile approaches the background profile after about 20 nt to the left and right of the main site of T to C mutations. Thus, these 41-nt long regions centered on the main site of T to C mutations are most likely to contain the binding sites and we focused our motif search on these regions.
RNA Recognition Site and Motif Search
For the motif search defining the core of a RNA recognition site we selected, for each RBP, the top 100 high confidence clusters, defined as described above. We selected the 41-nt region centered on the main T to C mutation site and searched for over-represented sequence motifs using PhyloGibbs (Siddharthan et al., 2005). We used a first-order Markov model as the background model and searched each set of sequences for three motifs of lengths varying between 4 and 8 nt, demanding an expected total number of 50 motifs. We did not search for motifs shorter than 4 nucleotides as the algorithm is then very likely to get stuck in local maxima. Additionally, significantly overrepresented short motifs are reproducibly found when using longer weight matrices. For each parameter setting, we performed five replicate runs. This generally resulted for each RBP in various shifted versions of the same motif Therefore we hierarchically clustered all the weight matrices that we obtained from these runs, allowing for partial overlap of at least 4 nucleotides between pairs of weight matrices. In the clustering procedure, two weight matrices were fused if the posterior probability of their stemming from the same as opposed to two different probability distribution was larger than 0.2 (for a description of the Bayesian calculation, see (Berninger et al., 2008), section 4.1). For each protein, we selected the largest cluster of weight matrices, i.e. the cluster that contained most of the weight matrices that we obtained in replicate runs, and created the final weight matrix by summing up the counts for each nucleotide of the weight matrices belonging to this cluster. Since the clustering procedure also allows the fusion of only partially overlapping weight matrices, the resulting weight matrices are typically longer (roughly 10 nucleotides) than the motif length that we imposed in individual runs, and can contain stretches of low information content. We therefore selected for each RBP, the window with highest information content. For PUM and QKI, the length of this window was 8 and 6 nt, respectively, in accordance with the known or expected consensus motifs (Galarneau and Richard, 2005; Gerber et al., 2006). For the IGF2BPs, we chose a window length of 4 nt, which is believed to be the size of binding motifs of KHdomains (Valverde et al., 2008). To identify binding sites in Pumilio clusters of aligned sequence reads using the inferred weight matrix, we used the MotEvo algorithm (van
Nimwegen, 2007), which is based on a hidden Markov model that models the input sequences as contiguous stretches of nucleotides drawn from a background or a weight matrix model. We chose for the background a first order Markov model (which makes every nucleotide dependent on the preceding nucleotide in the sequence). The background model parameters (di-nt frequencies) were estimated from the set of input sequences. MotEvo was run in the prior-update mode, meaning that we attempted to find the prior probabilities for sites and background that maximize the likelihood of the sequence data. MotEvo generates as an output a list of sites for the given input weight matrix as well as their corresponding posterior probabilities. Note that not all matches to the weight matrix are reported, but only the subset of matches whose corresponding sequence is more likely under the weight matrix model than the background model. We chose a cut-off of 0.4 on the posterior probability to define the set of binding sites.
Determination of the Location of Clusters of Mapped Sequence Reads within Functional mRNA Regions
For each RBP, we investigated whether clusters of mapped sequence reads preferentially originated in 5′UTR, CDS or 3′UTR. As a result of our annotation pipeline, we can assign probabilities to each cluster to belong to each of the three regions based on the annotation of individual tags within the cluster (see above). Taking together these probabilities for all clusters, we get estimates of the numbers of clusters originating in each of these three regions. We compare these numbers to those that we would expect if clusters were sampled uniformly from anywhere along the transcripts. This would for instance result in many more clusters from 3′ compared to 5′ UTR regions simply because 3′ UTRs tend to be longer than the 5′ UTRs. We determined all the transcripts to which a cluster mapped, and based on the GenBank annotation of the CDS of these transcripts, we calculated the fraction of the cluster nucleotides that fell in the 5′ UTR (f—5), CDS (f_CDS), and 3′ UTR (f—3) cluster. In the cases in which the cluster mapped to several transcripts belonging to the same gene, these fractions were averaged over all transcripts. The expected proportion of nucleotides sequenced from each region can then be calculated by summing these fractions for all clusters. The variance can be determined by noting that the probability that a nucleotide was sampled from a particular region, e.g. 5′ UTR, is Bernoulli distributed with parameter f—5, which has a variance of f—5(1−f—5). The total variance is then given as the sum of all the variances.
Distance Distribution Between Consecutive CAT-Motifs in the IGF2BP RNA Binding Sites
Since each of the IGF2BPs has 4 KH domains and we found only one clear motif, we hypothesized that all KH domains have the same or a very similar binding specificity. In analogy to what has been observed for Nova (Ule et al., 2006), we propose that the binding specificity of the IGF2BPs arises from the concerted action of several KHdomains that each recognize the same 4 letter sequence (CAUH), which should be apparent by a preferred spacing between subsequent occurrences of the motif as determined by the distance of corresponding KH-domains in the structure of the IGF2BPs. We calculated, for each IGF2BP separately, the distribution of distances between subsequent occurrences of the CAT-motif in clusters unambiguously derived from the 3′UTR of protein coding genes. We restricted ourselves to these clusters since 3′UTR regions are overrepresented in clusters of the IGF2BPs and each region, 5′UTR, CDS and 3′ UTR, has different sequence biases that need to be taken into account when modeling background distributions. In order to reduce boundary effects due to the finite length of the clusters, we extended each cluster region 32 nt to the right and left. We then compared this distance distribution to the distance distribution of consecutive occurrences of the CAT motif in randomly chosen 3′ UTR regions of the same length distribution as the clusters of mapped sequence reads. To estimate the mean and standard deviation of the relative frequency of each inter-motif distance in the background dataset, we repeated the random selection of 3′ UTR regions 1000 times. In Supplementary FIG. 8, we show that for each IGF2BP target set, there is a bias for CAT-motifs to occur at inter-motif distances of 3 to 6 nt.
Analysis of siRNA Knockdown Experiments for QKI and IGF2BPs:
The procedure for the knockdown experiments for QKI and IGF2BP1-3 were described above. We imported the CEL files into the R software using the BioConductor affy package (Gentleman et al., 2004). The transcript probe set intensities were background-corrected, adjusted for non-specific binding and quantile normalized with the GCRMA algorithm (Wu, 2006). Probe sets with more than 6 of the 11 probes mapping ambiguously to the genome were discarded, as were probe sets that mapped to multiple genes. We then collected all probe sets matching a given gene, and we selected for further analysis the RefSeq transcript with median 3′ UTR length corresponding to that gene. In total 16,063 transcripts were identified. The log-intensity of probe sets mapping to the gene were then averaged to obtain the expression level per RefSeq transcript. The level of transcript degradation was computed as the logarithm of the ratio of transcript expression in the cocktails of siRNA treated samples and mocktransfected cells.
To study the effect of individual proteins on the mRNA stability of their targets, we performed the following analysis. We first made the links between clusters of mapped Solexa sequence reads and expression data based on the NCBI Gene ID. That is, both the transcripts that were CLIPed and those whose expression was measured on microarrays have associated Gene IDs in the Gene database of NCBI. We mapped both the mapped sequence read clusters as well as the transcripts on microarrays to their corresponding genes, and thus identified which genes that were represented on microarrays have been CLIPed. From this set of genes we removed those that are likely off-targets of the transfected siRNAs. As previous studies showed, complementarity to the first 8 nucleotides of the miRNA is a good indicator that the transcript will be downregulated by a miRNA or siRNA, so we defined as putative off-targets those genes whose representative RefSeq transcripts carried such complementary sites in their 3′UTR. We divided the list of genes sorted by the maximum score of any cluster associated with a given gene. In order to improve the target identification and the assessment of the target response, we used some specific information that was available for individual data sets. For instance, for the IGF2BPs we only considered clusters with at least 2 positions of T to C changes, because we previously observed that this criterion improves the accuracy of target identification for the positive controls (PUM2 and QKI). Thus, for the IGF2BPs we divided the bound transcripts into the following bins, top 100 genes, 101th-300th genes, 301th-500th genes and 501th-1000th genes, 1001th-2000th, 2001th-3497th, and calculated the log 2 fold change of transcript abundance. To determine whether the siRNA knockdown has an effect on mRNA stability, we compared these distributions with the distribution of log-fold changes of genes that did not have any associated clusters from CLIP analysis. For QKI, whose binding motif is known, we performed the same analysis starting from clusters with a single T to C site, but that additionally contained the known QKI motif.
Generation and Ranking of Clusters of Mapped Sequence Reads for AGO and TNRC6
For subsequent analyses we only used extracted sequence reads of at least 20 nt in length and with unique, perfect or 1-error mapping to the genome. We clustered the reads with single-linkage criterion, meaning that we placed two reads in the same cluster if they overlapped by at least one nucleotide in their genomic mappings. We then selected the clusters that contained at least 5 mRNA-annotated reads and at least 2 positions at which T-to-C mutations occurred in the tags relative to the genomic sequence, and we ranked them by the total number of T-to-C mutations. This is because in our previous analyses of PURE-CLIP data (Landthaler et al., co-submitted) we found that the position of the crosslink is indicated by a T-to-C mutation in the tag relative to the genome sequence, and that clusters with at least two crosslinking positions are enriched in protein-binding sites roughly in proportion to the number of T-to-C mutations in the cluster.
Extraction of T-to-C-Anchored Regions from Sequence Read Clusters for AGO and TNRC6
In each ranked, mRNA-annotated cluster we identified the position with the largest number of T-to-C mutations, and we constructed the mutation frequency profile around this position. We found that this profile approaches the background after about 20 nucleotides to the left and right of the position with the maximum number of T-to-C changes, and we therefore extracted a genomic region of 41 nucleotides centered on this position for further analyses.
Filtering Out “Background” Clusters for AGO and TNRC6
It is still possible that a fraction of the clusters that we obtained simply contain degradation products of abundantly expressed mRNAs. Moreover, because a number of proteins that associate with the RISC complex have a molecular weight that is similar to that of AGO proteins, some of the tags/clusters that we obtained in the experiment with FLAG-tagged AGO may actually have been bound by other proteins. We have collected PURE-CLIP data for a number of proteins and we used to identify the AGO-specific clusters as follows. We built similar clusters for all the proteins that we investigated previously (IGF2BP1-3) (Landthaler, co-submitted). Then we compared the clusters that we obtained for different proteins, and when two clusters bound by two different proteins overlapped by more than 75% of their total length we considered that the two proteins shared a cluster. Finally, we discarded the following AGO clusters: clusters in which no position had a T-to-C mutation rate greater than 0.2, the experimentally determined T-to-C mutation rate at non-crosslinked sites; clusters that were shared between AGO libraries and libraries of other RBPs, with the number of tags in the AGO libraries being less than 1/10 of the number of tags in the other library. After applying these filters we obtained 17,319 AGO1-4 binding regions. We applied the same procedure to the clusters that we obtained from miR-124 and miR-7 transfection experiments.
Location of Clusters of Mapped Sequence Reads within Functional mRNA Regions
To investigate whether the clusters of mapped sequence reads preferentially originated in the 5′UTR, CDS or 3′UTR of protein-coding transcripts, we identified all the transcripts to which a cluster mapped, and based on the GenBank annotation of the CDS of these transcripts, we calculated the fraction of the cluster nucleotides that fell in the 5′ UTR, CDS, and 3′ UTR. When a cluster mapped to several transcripts belonging to the same gene, these fractions were averaged over all transcripts. The expected proportion of nucleotides sequenced from each functional region can then be calculated by summing these fractions for all clusters. The variance can be similarly determined by summing the variance over all clusters. This can be determined by noting that if the probability to sample a nucleotide from a particular region is Bernoulli distributed with parameter f its variance is f(1−f). We compared these numbers to those that we would expect if clusters were sampled uniformly from anywhere along the transcripts. This would for instance result in many more clusters from 3′ compared to 5′ UTR regions simply because 3′ UTRs tend to be longer than the 5′ UTRs.
Position of the Crosslink Relative to miRNA Seed-Complementary Sequence
We identified all the target regions (T-to-C anchored regions of 41 nucleotides) that have an 8-mer (A opposite miRNA position 1 and perfect match at miRNA positions 2-8) seed match and we extended symmetrically the seed-complementary region by 20 nts to the left and right. We then computed the positional T-to-C mutation frequency in these regions and normalized it over the length of the target region.
Regions of the miRNAs that Pair Contiguously with the Target Sites
To determine whether positions other than the seed region may be involved in base-pairing interaction with targets, we first took the T-to-C anchored target regions and identified those that had at least a 6mer (2-6 and A opposite miRNA position 1, 2-7 or 3-8) seed complementarity to at least one of the top 100 most expressed miRNAs in HEK293 cells. For each of these T-to-C anchored regions and each miRNA that matched to it, we identified all the occurrences of complementarities of at least 4 nucleotides between the miRNA and the putative target region. Each of these was counted with a weight 1/n towards the positional profile of miRNA-target site matches, with n being the number of miRNAs that matched the putative target region.
Analysis of miRNA Knockdown and Overexpression Experiments
CEL files of mRNA profiles from miRNA/2′O-methyl oligonucleotide and mocktransfected cell were loaded into the R software using the BioConductor affy package (Gentleman et al., 2004). The transcript probe set intensities were background-corrected, adjusted for non-specific binding and quantile normalized with the GCRMA algorithm (Wu, 2006). Probe sets with more than 6 of the 11 probes mapping ambiguously to the genome were discarded, as were probe sets that mapped to multiple genes. We then collected all probe sets matching a given gene, and we computed the logarithm of gene-level expression as the average of the log-intensity of probe sets mapping to that gene. For sequence analyses we also selected a representative transcript for each gene. This was the Genbank transcript with median 3′ UTR length corresponding to that gene. In total this procedure gave us expression measurements for 16,063 genes (and representative transcripts). The difference between the log-expression levels of a gene in samples treated with a cocktail of 2′-O-methyl oligoribonucleotides and mock transfected cells was taken as a measure of transcript stabilization upon miRNA inhibition.
We made the link between microarray and PURE-CLIP data through the NCBI Gene ID. That is, both the transcripts that were CLIPed and those whose expression was measured on microarrays have associated Gene IDs in the Gene database of NCBI, and we used this to identify the genes that were represented on microarrays and also yielded PURE-CLIP clusters. Of the 16063 transcripts measured by the Affymetrix array, 6,166 did not have any associated PURE-CLIP clusters.
Analysis of Transcript Stabilization as a Function of the Type of miRNA Target Sites
We constructed the distribution of log-fold-changes of transcripts with various types of PURE-CLIP clusters, and we compared them with the distribution of log-fold27 changes of transcripts that did not yield PURE-CLIP clusters, although they were expressed, as determined by the microarray measurements. The categories of transcripts were the following:
The sequence reads from the Digital Gene Expression experiment have been analyzed in a manner similar to that described above in the section “Adapter removal and sequence annotation”. We only considered genomic and transcript matches containing the GATC recognition sequence of the DpnII restriction enzyme directly upstream of the mapped sequence tag. For our analyses we further used sequence reads that had a perfect match in the genome. The probability that a tag originates in a given locus was then computed as 1/n of loci to which the tag can be mapped. The sequence reads were also mapped to the mRNA sequences and then we computed an expression level per gene. This was defined as the sum of the weighted copies of all tags that can be mapped to transcripts that originate in that gene. Finally, to assess the accuracy of the expression level measurements, we correlated the logarithm of the expression level measured on the lysate Affymetrix GeneChip® microarray with the logarithm expression level measured using the Digital Gene Expression technology. The Spearman correlation coefficient was higher than 0.68. We found a considerable number of transcripts that could be detected by sequencing (20,993) and that were undetectable on the microarrays (altogether 16,063 transcripts). Correlation between biological replicates of HEK293 cells was higher than 0.99. We correlated the logarithm of the expression level measured on the lysate Affymetrix GeneChip microarray with the logarithm expression level measured using the Digital Gene Expression technology. The Spearman correlation coefficient was 0.68. We could detect 22′466 genes by sequencing and 16′063 by microarray measurements. The correlation between biological replicates of the DGE experiment for HEK293 RNA was higher than 0.99.
Analysis of miRNA-Induced Destabilization of CLIPed and UNCLIPed miR-124 and miR-7 Targets
We intersected the transcripts with the background-noise-filtered PURE-CLIP clusters obtained after miR-124 and miR-7 transfection (see “filtering the background” section above) with those for which we had destabilization and AGO-IP Affymetrix microarray measurements. We then constructed, for each miRNA, three non-overlapping sets of transcripts: those with PURE-CLIP clusters exclusively in the 3′UTR, with PURECLIP clusters exclusively in the CDS only, and transcripts that did not yield any PURECLIP clusters. For each set, we computed the average log 2 fold change upon miRNA transfection, and the average log 2 fold enrichment in the AGO-IP. We compared these values between transcripts with and transcripts without PURE-CLIP clusters (FIG. S9A). The error bars on the bar plot represent 95% confidence intervals on the mean log 2 fold changes. Finally, we performed Wilcoxon's rank sum test to assess the significance of the difference in the log 2 fold changes of pairs of transcript sets. We also looked at various combinations of CLIP cluster locations (Supplementary FIG. 9)) that occurred more than 25 times in a given data set. Finally, we also compared the destabilization and AGObinding of PURE-CLIPed and UNCLIPed single miR-124 and miR-7 seed matches (Supplementary FIG. 9c). A seed match was defined as a match to nucleotides 1-7, 2-8 or 1-8 of the miRNA (both miRNAs start with U, so a 1-7 or 1-8 seed match also means having an A opposite nucleotide 1 of the miRNA). A seed match was considered “CLIPed” if it overlapped with a CLIP cluster from the corresponding transfection library.
Estimation of miRNA Expression Based on Deep Sequencing
The miRNA profile was generated from a Solexa deep sequencing run containing small RNAs from the following libraries: AGO1-IP and lysates of AGO1-4 IP, which were combined and denoted lysate in FIG. 1c. The miRNA annotation was preformed as described in (Berninger et al., 2008; Landgraf et al., 2007).
Plots of Motif Frequency-Vs-Enrichment
We performed a 7mer word enrichment analysis based on the T-to-C anchored target regions from the miRNA transfection experiments. We enumerated all words of length 7 and we determined their frequency in the real set as well as in a background set of shuffled sequences with the same dinucleotide content. For each 7-mer, we then calculated its enrichment as the ratio of the two frequencies. Additionally, we calculated for each 7mer the posterior probability that the frequency of the 7-mer is different in foreground and background allowing for sampling noise (Berninger et al., 2008). To determine whether the enriched motifs may correspond to miRNAs, all significantly enriched motifs (with a posterior>=0.99) were aligned with Needleman-Wunsch algorithm (penalties: gapopening −4, gapextension −4) to the reverse complemented of the transfected and to the top 20 most expressed in HEK293 miRNAs. We only reported cases in which the enriched word mapped with 0 or 1 errors to the first 9 positions of one of these miRNAs.
Identification of Significantly Enriched miRNA Binding Site Types
In order to identify individual miRNA binding sites in the sequence data we first defined a set of putative “binding models”. These were either contiguous matches to at least 6 nucleotides of a miRNA, or matches that had a single structural defect. This was defined as either an internal loop or a bulge either in the miRNA or in the mRNA. For each of the 553 miRNAs we enumerated all these binding models, and we determined the enrichment of the T-to-C anchored regions in each of these models, relative to 10 dinucleotide randomized sequence sets. Using a cutoff of 1.0e-20 in the probability that the real set had a lower frequency of occurrence compared to the randomized sets, which we used as a measure of the significance of the enrichment, we found all the T-to-C anchored regions that contained at least one significantly enriched binding model from one of the top 100 most expressed miRNAs within 10 nucleotides of the T-to-C mutation site. To obtain a comprehensive list of target sites we added to these the 7mer nucleotide matches (within the same 10 nucleotides of the T-to-C mutation) to positions 1-7 or 2-8 of one of the top 100 most expressed miRNAs, irrespective of whether the T-to-C anchored regions were enriched in these 7mers.
Correlation of miRNA Seed Expression with Seed-Complementary Motif
From all samples of smirnadb (Landgraf et al., 2007), all miRNAs that had at least 50 counts in total from all samples were used to build seed groups (defined by the motif found at positions 2-8). We added an additional sample which was generated by pooling together the miRNA reads from deep sequencing of HEK293 lysate as well as AGO1-4 IPs without crosslinking. For each sample, we computed the expression of a seed group as the sum of the tags of all miRNAs that were part of the seed group. We correlated the seed expression with the frequency of the seed-complementary motif in the T-to-C anchored regions.
Co-Occurrence of miRNA Seed Pairs
We set to determine if the CLIPed regions are enriched in pairs of binding sites for highly expressed miRNAs. Assuming that not all of these sites may have been captured in our experiment, we used for this purpose the 17,319 cluster regions which we extended by 32 nucleotides on either side. We scanned these regions for non-overlapping 7mers corresponding to the positions 2-8 of the top 20 most expressed miRNAs in HEK293 cells. We performed a similar procedure using 100 randomized variants of the extended clusters that preserved the di-nucleotide composition. The results are shown in Supplementary FIG. 6a. As additional controls we performed 1. The same procedure using 20 randomly selected miRNAs (Supplementary FIG. 6b): 2. Counting of the number of seed match pair occurrence in the extended clusters for 100 sets of 20 randomly selected miRNAs (Supplementary FIG. 6c). A visualization of seed match pair occurrence is shown in Supplementary FIG. 6d.
Properties of PURE-CLIPed and UNCLIPed miRNA Seed Matches
For the analyses whose results are presented in Supplementary FIG. 9 we needed to intersect the CLIP transcript sets with the transcript set measured by the Affymetrix microrray. In order to study the properties of CLIPed and UNCLIPe seed matches we do not need to make this intersection, and we therefore considered the entire set of miRNA seed matches that are present in the representative RefSeq transcripts. We chose as the representative RefSeq transcript for a given gene that transcript that had the median 3′UTR length from all RefSeq transcripts corresponding to a gene. RefSeq transcripts that could not be detected in the DGE transcriptome profile were discarded. For the analysis of the miR-124 and miR-7 transfection libraries, we scanned the 5′UTR, CDS and 3′UTRs of representative expressed RefSeq transcripts for 7mer or 8mer seed matches to miR-124 or miR-7, and intersected these with the background-noise-filtered miR-124 and miR-7 PURE-CLIP clusters to CLIPed and UNCLIPed seed matches. In parallel, we scanned the 5′UTR, CDS and 3′UTRs of representative expressed RefSeq transcripts for 7mer and 8mer seed matches to miR-15, miR-20, miR-103, miR-19, let-7 representing the top expressed miRNA families in HEK293 cells. These seed matches were then intersected with the background-noise-filtered AGO1-4, PURE-CLIP clusters.
Furthermore, because we wanted to analyze properties of the environment of the putative miRNA target sites, we only considered seed matches located at least 100 nucleotides away from either of the boundaries of the transcript. For each individual seed match, we computed the following quantities: Selection pressure: is the posterior probability that a seed complementary region is under evolutionary selection pressure, as computed by the ElMMo algorithm described in (Gaidatzis et al., 2007).
Predicted destabilization score: is a score that characterizes the extent to which the environment of a seed match is favorable for its functionality in mRNA destabilization, as computed by the TargetScanS algorithm (Grimson et al., 2007). For the analysis, we downloaded the TargetScan 4.2 from the TargetScan website. local AU content: was defined as the proportion of A+U nucleotides within 50 nucleotides upstream and 50 nucleotides downstream of the miRNA binding site, which was a 20 nt-long region, anchored at the 3′end by the seed-matching region.
Target site Eopen: was similarly defined in terms of the energy required to open the secondary structure of the target in a region of 20 nucleotides, anchored at the 3′end by the seed-complementary region (opposite positions 1-8 of the miRNA). This was computed using the program RNAup of the Vienna package (Hofacker, 2003) with the following parameters: u=20 (length of the window required to be single-stranded), w=50 (maximal distance between 2 nucleotides allowed to interact). The rest of the parameters were left with their default values. The negative value of this energy can be viewed as a measure of accessibility. We tested whether the four properties introduced above took significantly different values when comparing CLIPed to unCLIPed seed matches using Wilcoxon's rank sum test.
Codon Adaptation Index Around PURE-CLIPed and UNCLIPed Seed Matches
We compared the Codon Adaptation Index (CAI) (Sharp and Li, 1987) around CLIPed and unCLIP seed matches as follows. We obtained a reference codon usage by analyzing all the CDS from representative RefSeq transcripts from all expressed genes. We then anchored all sequences at the codon covering the 5′ end of seed match (1-7, 2-8, or 1-8 of miR-15, miR-20, miR-103, miR-19, let-7 miRNAs) and computed the CAI for the 70 codons upstream and downstream of the anchor, i.e. a total of 141 codons. The 7mer or 8mer seed match is entirely covered by codons 0, 1 and 2, which highly constrains the codon usage at these positions, making it uninformative. The figure therefore does not show the CAI at these positions. For CLIPed seed matches, we smoothed the profile using a moving average of 5.
Analysis of Positional Bias of CLIPed and UNCLIPed Regions
We set to determine whether CLIPed seed matches (1-7, 2-8, or 1-8 of miR-15, miR-20, miR-103, miR-19, let-7 miRNAs) have a positional bias relative to the STOP codon. Noting that at least in the 4 Ago libraries, CLIPed seed matches tended to be located in CDS of shorter lengths than their unCLIPed counterparts, we performed local polynomial regression (loess, see W. S. Cleveland, E. Grosse and W. M. Shyu (1992) Local regression models. Chapter 8 of Statistical Models in S, eds J. M. Chambers and T. J. Hastie, Wadsworth & Brooks/Cole), fitting the distance between the seed matches and the STOP codon to the CDS length. The loess fit and standard errors on the distance to the STOP codon given the CDS length were obtained using R's loess and predict.loess functions with default parameters. The miRNA transfection (Supplementary FIG. 11a) and Ago (Supplementary FIG. 11b) CLIP libraries were analyzed separately, and loess fits were computed separately for crosslinked and uncrosslinked seed matches (shown in red and black, respectively). Finally, we represented the expected distance to the STOP codon as a function of the CDS length assuming that seed matches are distributed uniformly over the CDS (dashed blue curve). We used the same methodology to determine whether CLIPed sites are located preferentially towards a 3′UTR boundary (STOP codon or polyA tail) instead of the STOP codon.
Incorporated herein by reference in its entirety is the Sequence Listing for the application. The Sequence Listing is disclosed on a computer-readable ASCII text file titled, “sequence_listing—1119-29PCTUS.txt”, created on Jan. 26, 2011. The sequence_listing.txt file is 39.9 kb in size.