《遗传学》课程教学资源（学科前沿）蕴藏在基因组中的生命密码 Nature ENCODE pilot project

Vol 447 14 June 2007 doi: 10. 1038/nature05874 nature ARTICLES Identification and analysis of functional elements in 1% of the human genome by the Encode pilot project The ENCODE Project Consortium* We report the generation and analysis of functional data from multiple, diverse experiments performed on a targeted 1% of the human genome as part of the pilot phase of the ENCODE Project. these data have been further integrated and augmented by a number of evolutionary and computational analyses. Together our results advance the collective knowledge about human genome function in several major areas. First, our studies provide convincing evidence that the genome is pervasively transcribed, such that the majority of its bases can be found in primary transcripts, including non-protein-coding transcripts, and those that extensively overlap one another Second systematic examination of transcriptional regulation has yielded new understanding about transcription start sites, including their relationship to specific regulatory sequences and features of chromatin accessibility and histone modification. third, a more sophisticated view of chromatin structure has emerged including its inter-relationship with DNA replication and transcriptional regulation Finally, integration of these new sources of information, in particular with respect to mammalian evolution based on inter- and intra-species sequence comparisons, has yielded new mechanistic and evolutionary insights concerning the functional landscape of the human genome Together, these studies are defining a path for pursuit of a more comprehensive characterization of human genome function The human genome is an elegant but cryptic store of information. The evolve, our present understanding about the evolution of other func roughly three billion bases encode, either directly or indirectly, the tional genomic regions is poorly developed. Experimental studies instructions for synthesizing nearly all the molecules that form each that augment what we learn from evolutionary analyses are key for human cell, tissue and organ. Sequencing the human genome- pro- solidifying our insights regarding genome function. vided highly accurate DNA sequences for each of the 24 chromosomes. The Encyclopedia of DNA Elements(ENCODE) Project aims to However, at present, we have an incomplete understanding of the provide a more biologically informative representation of the human protein-coding portions of the genome, and markedly less under- genome by using high-throughput methods to identify and catalogue standing of both non-protein-coding transcripts and genomic ele- the functional elements encoded. In its pilot phase, 35 groups pro ments that temporally and spatially regulate gene expression. To vided more than 200 experimental and computational data sets that understand the human genome, and by extension the biological pro- examined in unprecedented detail a targeted 29, 998 kilobases(kb)of cesses it orchestrates and the ways in which its defects can give rise to the human genome. These roughly 30 Mb--equivalent to -1% of disease, we need a more transparent view of the information it encodes. the human genome--are sufficiently large and diverse to allow for The molecular mechanisms by which genomic information directs rigorous pilot testing of multiple experimental and computational the synthesis of different biomolecules has been the focus of much of methods. These 30 Mb are divided among 44 genomic regions; molecular biology research over the last three decades. Previous stud- approximately 15 Mb reside in 14 regions for which there is already ies have typically concentrated on individual genes, with the resulting substantial biological knowledge, whereas the other 15 Mb reside in general principles then providing insights into transcription, chro- 30 regions chosen by a stratified random-sampling method(see matinremodellingmessengerRnasplicingDnareplicationandhttp://www.genome.gov/10506161).Thehighlightsofourfindings numerous other genomic processes. Although many such principles to date include seem valid as additional genes are investigated, they generally have The human is pervasively transcribed, such that the not provided genome-wide insights about biological function. majority of its bases are associated with at least one primary tran E The first genome-wide analyses that shed light on human genome script and many t pts link distal regions to established protei Inction made use of observing the actions of evolution. The ever- coding loci growing set of vertebrate genome sequences- is providing increas-. Many novel non-protein-coding transcripts have been identified, convincingly indicate the presence of numerous genomic regions tionally silent. under strong evolutionary constraint, they have less power in iden- Numerous previously unrecognized transcription start sites tifying the precise bases that are constrained and provide little, if any, have been identified, many of which show chromatin structure insight into why those bases are biologically important. Furthermore, and sequence-specific protein-binding properties similar to well lthough we have good models for how protein-coding regions understood promoters a list of authors and their affiliations appears at the end of the paper. E2007 Nature Publishing Group

ARTICLES Identification and analysis of functional elements in 1% of the human genome by the ENCODE pilot project The ENCODE Project Consortium* We report the generation and analysis of functional data from multiple, diverse experiments performed on a targeted 1% of the human genome as part of the pilot phase of the ENCODE Project. These data have been further integrated and augmented by a number of evolutionary and computational analyses. Together, our results advance the collective knowledge about human genome function in several major areas. First, our studies provide convincing evidence that the genome is pervasively transcribed, such that the majority of its bases can be found in primary transcripts, including non-protein-coding transcripts, and those that extensively overlap one another. Second, systematic examination of transcriptional regulation has yielded new understanding about transcription start sites, including their relationship to specific regulatory sequences and features of chromatin accessibility and histone modification. Third, a more sophisticated view of chromatin structure has emerged, including its inter-relationship with DNA replication and transcriptional regulation. Finally, integration of these new sources of information, in particular with respect to mammalian evolution based on inter- and intra-species sequence comparisons, has yielded new mechanistic and evolutionary insights concerning the functional landscape of the human genome. Together, these studies are defining a path for pursuit of a more comprehensive characterization of human genome function. The human genome is an elegant but cryptic store of information. The roughly three billion bases encode, either directly or indirectly, the instructions for synthesizing nearly all the molecules that form each human cell, tissue and organ. Sequencing the human genome1–3 pro￾vided highly accurate DNA sequences for each of the 24 chromosomes. However, at present, we have an incomplete understanding of the protein-coding portions of the genome, and markedly less under￾standing of both non-protein-coding transcripts and genomic ele￾ments that temporally and spatially regulate gene expression. To understand the human genome, and by extension the biological pro￾cesses it orchestrates and the ways in which its defects can give rise to disease, we need a more transparent view of the information it encodes. The molecular mechanisms by which genomic information directs the synthesis of different biomolecules has been the focus of much of molecular biology research over the last three decades. Previous stud￾ies have typically concentrated on individual genes, with the resulting general principles then providing insights into transcription, chro￾matin remodelling, messenger RNA splicing, DNA replication and numerous other genomic processes. Although many such principles seem valid as additional genes are investigated, they generally have not provided genome-wide insights about biological function. The first genome-wide analyses that shed light on human genome function made use of observing the actions of evolution. The ever￾growing set of vertebrate genome sequences4–8 is providing increas￾ing power to reveal the genomic regions that have been most and least acted on by the forces of evolution. However, although these studies convincingly indicate the presence of numerous genomic regions under strong evolutionary constraint, they have less power in iden￾tifying the precise bases that are constrained and provide little, if any, insight into why those bases are biologically important. Furthermore, although we have good models for how protein-coding regions evolve, our present understanding about the evolution of other func￾tional genomic regions is poorly developed. Experimental studies that augment what we learn from evolutionary analyses are key for solidifying our insights regarding genome function. The Encyclopedia of DNA Elements (ENCODE) Project9 aims to provide a more biologically informative representation of the human genome by using high-throughput methods to identify and catalogue the functional elements encoded. In its pilot phase, 35 groups pro￾vided more than 200 experimental and computational data sets that examined in unprecedented detail a targeted 29,998 kilobases (kb) of the human genome. These roughly 30 Mb—equivalent to ,1% of the human genome—are sufficiently large and diverse to allow for rigorous pilot testing of multiple experimental and computational methods. These 30 Mb are divided among 44 genomic regions; approximately 15 Mb reside in 14 regions for which there is already substantial biological knowledge, whereas the other 15 Mb reside in 30 regions chosen by a stratified random-sampling method (see http://www.genome.gov/10506161). The highlights of our findings to date include: $ The human genome is pervasively transcribed, such that the majority of its bases are associated with at least one primary tran￾script and many transcripts link distal regions to established protein￾coding loci. $ Many novel non-protein-coding transcripts have been identified, with many of these overlapping protein-coding loci and others located in regions of the genome previously thought to be transcrip￾tionally silent. $ Numerous previously unrecognized transcription start sites have been identified, many of which show chromatin structure and sequence-specific protein-binding properties similar to well￾understood promoters. *A list of authors and their affiliations appears at the end of the paper. Vol 447| 14 June 2007| doi:10.1038/nature05874 799 ©2007 NaturePublishingGroup

ARTICLES NATURE Vol 447 14 June 2007 Regulat surround transcription start sites and what we believe the pi are for a broader with no bias towards upstream investigation of the functional elements in the human id the reader, Box I provides a glossary for many of the e Chromatin accessibility and histone modification patterns are ns used throughout this paper highly predictive of both the presence and activity of transcription start sites Experimental techniques Distal DNasel hypersensitive sites have characteristic histone Table 1(expanded in Supplementary Information section 1.1)lists modification patterns that reliably distinguish them from promo- the major experimental techniques used for the studies reported here, ters; some of these distal sites show marks consistent with insulator relevant acronyms, and references reporting the generated data sets. function These data sets reflect over 400 million experimental data points e DNA replication timing is correlated with chromatin structure. (603 million data points if one includes comparative sequencing e A total of 5% of the bases in the genome can be confidently bases). In describing the major results and initial conclusions, we identified as being under evolutionary constraint in mammals; for seek to distinguish biochemical function'from biological role approximately 60% of these constrained bases, there is evidence of Biochemical function reflects the direct behaviour of a molecule(s) unction on the basis of the results of the experimental assays per- whereas biological role is used to describe the consequence(s)of this formed to date function for the organism. Genome-analysis techniques nearly e Although there is general overlap between genomic regions iden- always focus on biochemical function but not necessarily on bio tified as functional by experimental assays and those under evolu- logical role. This is because the former is more amenable to large tionary constraint, not all bases within these experimentally defined scale data-generation methods, whereas the latter is more difficult to regions show evidence of constraint. assay on a large scale Different functional elements vary greatly in their sequence vari- The ENCODe pilot project aimed to establish redundancy with ability across the human population and in their likelihood of res- respect to the findings represented by different data sets. In some iding within a structurally variable region of the genome instances, this involved the intentional use of different assays that were e Surprisingly, many functional elements are seemingly uncon- based on a similar technique, whereas in other situations, different strained across mammalian evolution. This suggests the possibility techniques assayed the same biochemical function. Such redundancy of a large pool of neutral elements that are biochemically active but has allowed methods to be compared and consensus data sets to be provide no specific benefit to the organism. This pool may serve as a generated, much of which is discussed in warehouse for natural selection, potentially acting as the source as the ChIP-chip platform comparison. L. All ENCODE data have of lineage-specific elements and functionally conserved but non- been released after verification but before this publication, as befits orthologous elements between species. acommunityresource'project(seehttp://www.wellcome.ac.uk/ Below, we first provide an overview of the experimental techniques doc_wtdo03208. html) Verification is defined as when the experiment used for our studies, after which we describe the insights gained from is reproducibly confirmed (see Supplementary Information section halysing and integrating the generated data sets. We conclude with a 1.2). The main portal for ENCoDE data is provided by the UCSC perspectiveofwhatwehavelearnedtodateaboutthis1%oftheGenomebrOwser(http://genome.ucsc.edu/encode/);thisis Box 1 Frequently used abbreviations in this paper at that was inserted into the early ndel An insertion or deletion; two sequences often show a length mammalian lineage and has since become dormant; the majority of difference within alignments, but it is not always clear whether this ancient repeats are thought to be neutrally evolving reflects a previous insertion or a deletion CAGE tag A short sequence from the 5' end of a transcript PET A short sequence that contains both the 5 and 3' ends of CDS Coding sequence: a region of a cDNA or genome that encodes transcri roteins RACE Rapid amplification of cDNA ends: a technique for amplifying ChIP-chip Chromatin immunoprecipitation followed by detection of cDNa sequences between a known internal position in a transcript and the products using a genomic tiling array CNV Copy number variants: regions of the genome that have large factor binding region: a genomic region found by a duplications in some individuals in the human population ChIP-chip assay to be bound by a protein fac CS Constrained sequence: a genomic region associated with evidence RFBR-Seqsp Regulatory factor binding regions that are from of negative selection(that is, rejection of mutations relative to neutral sequence-specific binding factors RT-PCR Reverse transcriptase polymera n reaction: a Nasel hypersensitive site: a region of the genome showing a echnique for ga spe different sensitivity to DNasel compared with its RxFrag Fragment of a race reaction: a egion found to be ocale present in a RACE product by an unbiased tiling-array assay EST Expressed sequence tag: a short sequence of acDNA indicative of SNP Single nucleotide polymorphism: a single base pair change expression at this point between two individuals in uman population FAIRE Formaldehyde -assisted isolation of regulatory elements: a TAGE Sequence tag analysis of genomic enrichment: a method similar method to open chromatin using formaldehyde crosslinking to ChIP-chip for detecting protein factor binding regions but using ollowed by detection of the products using a genomic tiling array extensive short sequence determination rather than genomic tiling arrays FDR False discovery rate: a statistical method for setting thresholds on SVM Support vector machine: a machine-learning technique that ca statistical tests to correct for multiple testing establish an optimal classifier on the basis of labelled training data GENCODE Integrated annotation of existing cDNA and protein TR50 A measure of replication timing corresponding to the time in the resources to define transcripts with both manual review and GSC Genome structure correction: a method to adapt statistical tests Tss Transcription start site to make fewer assumptions about the distribution of features on the Tx Frag Fragment of a transcript: a genomic region found to be present genome sequence. This provides a conservative correction to standard in a transcript by an unbiased tiling-array assay ests Un. TxFrag A Tx Frag that is not associated with any other functional HMM Hidden Markov model: a machine-learning technique that can establish optimal parameters for a given model to explain the observed ITR Untranslated region: part of a cDNA either at the 5 or 3 end that does not encode a protein sequence E2007 Nature Publishing Group

$ Regulatory sequences that surround transcription start sites are symmetrically distributed, with no bias towards upstream regions. $ Chromatin accessibility and histone modification patterns are highly predictive of both the presence and activity of transcription start sites. $ Distal DNaseI hypersensitive sites have characteristic histone modification patterns that reliably distinguish them from promo￾ters; some of these distal sites show marks consistent with insulator function. $ DNA replication timing is correlated with chromatin structure. $ A total of 5% of the bases in the genome can be confidently identified as being under evolutionary constraint in mammals; for approximately 60% of these constrained bases, there is evidence of function on the basis of the results of the experimental assays per￾formed to date. $ Although there is general overlap between genomic regions iden￾tified as functional by experimental assays and those under evolu￾tionary constraint, not all bases within these experimentally defined regions show evidence of constraint. $ Different functional elements vary greatly in their sequence vari￾ability across the human population and in their likelihood of res￾iding within a structurally variable region of the genome. $ Surprisingly, many functional elements are seemingly uncon￾strained across mammalian evolution. This suggests the possibility of a large pool of neutral elements that are biochemically active but provide no specific benefit to the organism. This pool may serve as a ‘warehouse’ for natural selection, potentially acting as the source of lineage-specific elements and functionally conserved but non￾orthologous elements between species. Below, we first provide an overview of the experimental techniques used for our studies, after which we describe the insights gained from analysing and integrating the generated data sets. We conclude with a perspective of what we have learned to date about this 1% of the human genome and what we believe the prospects are for a broader and deeper investigation of the functional elements in the human genome. To aid the reader, Box 1 provides a glossary for many of the abbreviations used throughout this paper. Experimental techniques Table 1 (expanded in Supplementary Information section 1.1) lists the major experimental techniques used for the studies reported here, relevant acronyms, and references reporting the generated data sets. These data sets reflect over 400 million experimental data points (603 million data points if one includes comparative sequencing bases). In describing the major results and initial conclusions, we seek to distinguish ‘biochemical function’ from ‘biological role’. Biochemical function reflects the direct behaviour of a molecule(s), whereas biological role is used to describe the consequence(s) of this function for the organism. Genome-analysis techniques nearly always focus on biochemical function but not necessarily on bio￾logical role. This is because the former is more amenable to large￾scale data-generation methods, whereas the latter is more difficult to assay on a large scale. The ENCODE pilot project aimed to establish redundancy with respect to the findings represented by different data sets. In some instances, this involved the intentional use of different assays that were based on a similar technique, whereas in other situations, different techniques assayed the same biochemical function. Such redundancy has allowed methods to be compared and consensus data sets to be generated, much of which is discussed in companion papers, such as the ChIP-chip platform comparison10,11. All ENCODE data have been released after verification but before this publication, as befits a ‘community resource’ project (see http://www.wellcome.ac.uk/ doc_wtd003208.html). Verification is defined as when the experiment is reproducibly confirmed (see Supplementary Information section 1.2). The main portal for ENCODE data is provided by the UCSC Genome Browser (http://genome.ucsc.edu/ENCODE/); this is Box 1 | Frequently used abbreviations in this paper AR Ancient repeat: a repeat that was inserted into the early mammalian lineage and has since become dormant; the majority of ancient repeats are thought to be neutrally evolving. CAGE tag A short sequence from the 59 end of a transcript CDS Coding sequence: a region of a cDNA or genome that encodes proteins ChIP-chip Chromatin immunoprecipitation followed by detection of the products using a genomic tiling array CNV Copy number variants: regions of the genome that have large duplications in some individuals in the human population CS Constrained sequence: a genomic region associated with evidence of negative selection (that is, rejection of mutations relative to neutral regions) DHS DNaseI hypersensitive site: a region of the genome showing a sharply different sensitivity to DNaseI compared with its immediate locale EST Expressed sequence tag: a short sequence of a cDNA indicative of expression at this point FAIRE Formaldehyde-assisted isolation of regulatory elements: a method to assay open chromatin using formaldehyde crosslinking followed by detection of the products using a genomic tiling array FDR False discovery rate: a statistical method for setting thresholds on statistical tests to correct for multiple testing GENCODE Integrated annotation of existing cDNA and protein resources to define transcripts with both manual review and experimental testing procedures GSC Genome structure correction: a method to adapt statistical tests to make fewer assumptions about the distribution of features on the genome sequence. This provides a conservative correction to standard tests HMM Hidden Markov model: a machine-learning technique that can establish optimal parameters for a given model to explain the observed data Indel An insertion or deletion; two sequences often show a length difference within alignments, but it is not always clear whether this reflects a previous insertion or a deletion PET A short sequence that contains both the 59 and 39 ends of a transcript RACE Rapid amplification of cDNA ends: a technique for amplifying cDNA sequences between a known internal position in a transcript and its 59 end RFBR Regulatory factor binding region: a genomic region found by a ChIP-chip assay to be bound by a protein factor RFBR-Seqsp Regulatory factor binding regions that are from sequence-specific binding factors RT–PCR Reverse transcriptase polymerase chain reaction: a technique for amplifying a specific region of a transcript RxFrag Fragment of a RACE reaction: a genomic region found to be present in a RACE product by an unbiased tiling-array assay SNP Single nucleotide polymorphism: a single base pair change between two individuals in the human population STAGE Sequence tag analysis of genomic enrichment: a method similar to ChIP-chip for detecting protein factor binding regions but using extensive short sequence determination rather than genomic tiling arrays SVM Support vector machine: a machine-learning technique that can establish an optimal classifier on the basis of labelled training data TR50 A measure of replication timing corresponding to the time in the cell cycle when 50% of the cells have replicated their DNA at a specific genomic position TSS Transcription start site TxFrag Fragment of a transcript: a genomic region found to be present in a transcript by an unbiased tiling-array assay Un.TxFrag A TxFrag that is not associated with any other functional annotation UTR Untranslated region: part of a cDNA either at the 59 or 39 end that does not encode a protein sequence ARTICLES NATURE|Vol 447| 14 June 2007 800 ©2007 NaturePublishingGroup

NATURE Vol 447 14 June 2007 ARTICLES augmented by multiple other websites(see Supplementary Informa- compared with the total RNA in a cell, suggesting that there are tion section 1.1) numerous RNA species yet to be classified-. In addition, studies A common feature of genomic analyses is the need to assess the of specific loci have indicated the presence of RNA transcripts that ignificance of the co-occurrence of features or of other statistical have a role in chromatin maintenance and other regulatory control. e44 across the genome. We have developed and used a statistical frame- encoded RNA molecule work that mitigates many of these hidden correlations by adjusting Transcript maps. We used three methods to identify transcripts he appropriate null distribution of the test statistics. We term this emanating from the ENCODE regions: hybridization of rNa(either correction procedure genome structure correction(GSC)(see Sup- total or polyA-selected)to unbiased tiling arrays(see Supplementary plementary Information section 1.3) Information section 2.1), tag sequencing of cap-selected RNA at the In the next five sections, we detail the various biological insights of 5 or joint 5 /3 ends(see Supplementary Information sections 2.2 the pilot phase of the ENCODE Project. and S2.3), and integrated annotation of available complementary DNA and EST sequences involving computational, manual, and Transcript experimental approaches(see Supplementary Information section Overview. RNA transcripts are involved in many cellular functions, 2.4). We abbreviate the regions identified by unbiased tiling arrays as either directly as biologically active molecules or indirectly by encod- Tx Frags, the cap-selected RNAs as CAGE or PET tags(see Box 1),and ons other active molecules. In the conventional view of genome the integrated annotation as GENCODE transcripts. When a TxFrag ganization, sets of RNA transcripts(for example, messenger does no lap a GENCODE annotation, we call it an Un. TxFrag RNAs)are encoded by distinct loci, with each usually dedicated to Validation of these various studies is described in papers reporting a single biological role( for example, encoding a specific protein). these data sets(see Supplementary Information sections 2.1.4 and However, this picture has substantially grown in complexity in recent 2.1.5) years 2. Other forms of RNA molecules(such as small nucleolar These methods recapitulate previous findings, but provide RNAs and micro(mi)RNAs)are known to exist, and often these enhanced resolution owing to the larger number of tissues sampled are encoded by regions that intercalate with protein-coding genes. and the integration of results across the three approaches(see Table 2) These observations are consistent with the well-known discrepancy To begin with, our studies show that 14.7% of the bases represented in between the levels of observable mRNAs and large structural RNAs the unbiased tiling arrays are transcribed in at least one tissue sample Consistent with previous work. s, many (63%)Tx Frags reside out- side of GENCODE annotations, both in intronic(40.9%)and inter Table 1 Summary of types of experimental techniques used in ENCODE genic(22.6%)regions. GENCODE annotations are richer than the more-conservative RefSeq or Ensembl annotations, with 2, 608 tran- data points scripts clustered into 487 loci, leading to an average of 5. 4 transcripts 63348656 per locus. Finally, extensive testing of predicted protein-coding sequences outside of GENCODE annotations was positive in only annotation 2% of cases 6, suggesting that GENCODE annotations cover nearly Tag sequencing PET, CAGE 121 864,964 all protein-coding sequences. The GENCODE annotations are cate transcripts gorized both by likely function (mainly, the presence of an open Tiling array Histone 4,401,291 reading frame)and by classification evidence(for example, transcripts based solely on ESTs are distinguished from other scenarios ); this Chromatin QT-PCR, tiling DHS, FAIRE 42 15.318.324 classification is not strongly correlated with expression levels(see upplementary Information sections 2.4.2 and 2.4.3 Analyses of more biological samples have allowed a richer descrip tion of the transcription specificity(see Fig. I and Supplementary Tiling array, tag STAGE, ChIP- 41, 52 324, 846,018 Information section 2.5). We found that 40%of Tx Frags are preser promoter assays Chip, chIP-PET, 11, 1. in only one sample, whereas only 2% are present in all sampl Although exon-containing Tx Frags are more likely(74%)to be expressed in more than one sample, 45% of unannotated TxFrags are also expressed in multiple samples. GENCODE annotations of separate loci often(42%)overlap with respect to their genomic ates, in p plication Tiling array TR50 analysis of GENCODE-annotated sequences with respect to the posi- Computational Computational CC, RFBR cluster tions of open reading frames revealed that some component exons do not have the expected synonymous versus non-synonymous substi- tution patterns of protein-coding sequence(see Supplement Infor mation section 2.6)and some have deletions incompatible with Table 2 Bases detected in processed transcripts either as a GENCODE exon, a TxFrag, or as either a gENCODE exon or a Tx Frag GENCODE exon Either GENCODE exon TxFrag T e1,776,157(59%)1,369611(46%)2519,280(84%) transcripts(bases) copy number Transcripts detected1,447,192(98%)1,369611(93%)2163303(14.7%) ariation Not all da ENCODE Project. t Histone code nomenclature follows the Brno nomenclature as described in ref. 129 Percentages are of total bases in ENCODE in the first row and bases tiled in arrays in the second tAlso contains histone modification. E2007 Nature Publishing Group

augmented by multiple other websites (see Supplementary Informa￾tion section 1.1). A common feature of genomic analyses is the need to assess the significance of the co-occurrence of features or of other statistical tests. One confounding factor is the heterogeneity of the genome, which can produce uninteresting correlations of variables distributed across the genome. We have developed and used a statistical frame￾work that mitigates many of these hidden correlations by adjusting the appropriate null distribution of the test statistics. We term this correction procedure genome structure correction (GSC) (see Sup￾plementary Information section 1.3). In the next five sections, we detail the various biological insights of the pilot phase of the ENCODE Project. Transcription Overview. RNA transcripts are involved in many cellular functions, either directly as biologically active molecules or indirectly by encod￾ing other active molecules. In the conventional view of genome organization, sets of RNA transcripts (for example, messenger RNAs) are encoded by distinct loci, with each usually dedicated to a single biological role (for example, encoding a specific protein). However, this picture has substantially grown in complexity in recent years12. Other forms of RNA molecules (such as small nucleolar RNAs and micro (mi)RNAs) are known to exist, and often these are encoded by regions that intercalate with protein-coding genes. These observations are consistent with the well-known discrepancy between the levels of observable mRNAs and large structural RNAs compared with the total RNA in a cell, suggesting that there are numerous RNA species yet to be classified13–15. In addition, studies of specific loci have indicated the presence of RNA transcripts that have a role in chromatin maintenance and other regulatory control. We sought to assay and analyse transcription comprehensively across the 44 ENCODE regions in an effort to understand the repertoire of encoded RNA molecules. Transcript maps. We used three methods to identify transcripts emanating from the ENCODE regions: hybridization of RNA (either total or polyA-selected) to unbiased tiling arrays (see Supplementary Information section 2.1), tag sequencing of cap-selected RNA at the 59 or joint 59/39 ends (see Supplementary Information sections 2.2 and S2.3), and integrated annotation of available complementary DNA and EST sequences involving computational, manual, and experimental approaches16 (see Supplementary Information section 2.4). We abbreviate the regions identified by unbiased tiling arrays as TxFrags, the cap-selected RNAs as CAGE or PET tags (see Box 1), and the integrated annotation as GENCODE transcripts. When a TxFrag does not overlap a GENCODE annotation, we call it an Un.TxFrag. Validation of these various studies is described in papers reporting these data sets17 (see Supplementary Information sections 2.1.4 and 2.1.5). These methods recapitulate previous findings, but provide enhanced resolution owing to the larger number of tissues sampled and the integration of results across the three approaches (see Table 2). To begin with, our studies show that 14.7% of the bases represented in the unbiased tiling arrays are transcribed in at least one tissue sample. Consistent with previous work14,15, many (63%) TxFrags reside out￾side of GENCODE annotations, both in intronic (40.9%) and inter￾genic (22.6%) regions. GENCODE annotations are richer than the more-conservative RefSeq or Ensembl annotations, with 2,608 tran￾scripts clustered into 487 loci, leading to an average of 5.4 transcripts per locus. Finally, extensive testing of predicted protein-coding sequences outside of GENCODE annotations was positive in only 2% of cases16, suggesting that GENCODE annotations cover nearly all protein-coding sequences. The GENCODE annotations are cate￾gorized both by likely function (mainly, the presence of an open reading frame) and by classification evidence (for example, transcripts based solely on ESTs are distinguished from other scenarios); this classification is not strongly correlated with expression levels (see Supplementary Information sections 2.4.2 and 2.4.3). Analyses of more biological samples have allowed a richer descrip￾tion of the transcription specificity (see Fig. 1 and Supplementary Information section 2.5). We found that 40% of TxFrags are present in only one sample, whereas only 2% are present in all samples. Although exon-containing TxFrags are more likely (74%) to be expressed in more than one sample, 45% of unannotated TxFrags are also expressed in multiple samples. GENCODE annotations of separate loci often (42%) overlap with respect to their genomic coor￾dinates, in particular on opposite strands (33% of loci). Further analysis of GENCODE-annotated sequences with respect to the posi￾tions of open reading frames revealed that some component exons do not have the expected synonymous versus non-synonymous substi￾tution patterns of protein-coding sequence (see Supplement Infor￾mation section 2.6) and some have deletions incompatible with Table 1 | Summary of types of experimental techniques used in ENCODE Feature class Experimental technique(s) Abbreviations References Number of experimental data points Transcription Tiling array, integrated annotation TxFrag, RxFrag, GENCODE 117 118 19 119 63,348,656 59 ends of transcripts* Tag sequencing PET, CAGE 121 13 864,964 Histone modifications Tiling array Histone nomenclature{, RFBR 46 4,401,291 Chromatin{ structure QT-PCR, tiling array DHS, FAIRE 42 43 44 122 15,318,324 Sequence￾specific factors Tiling array, tag sequencing, promoter assays STAGE, ChIP￾Chip, ChIP-PET, RFBR 41,52 11,120 123 81 34,51 124 49 33 40 324,846,018 Replication Tiling array TR50 59 75 14,735,740 Computational analysis Computational methods CCI, RFBR cluster 80 125 10 16 126 127 NA Comparative sequence analysis* Genomic sequencing, multi￾sequence alignments, computational analyses CS 87 86 26 NA Polymorphisms* Resequencing, copy number variation CNV 103 128 NA * Not all data generated by the ENCODE Project. { Histone code nomenclature follows the Brno nomenclature as described in ref. 129. {Also contains histone modification. Table 2 | Bases detected in processed transcripts either as a GENCODE exon, a TxFrag, or as either a GENCODE exon or a TxFrag GENCODE exon TxFrag Either GENCODE exon or TxFrag Total detectable transcripts (bases) 1,776,157 (5.9%) 1,369,611 (4.6%) 2,519,280 (8.4%) Transcripts detected in tiled regions of arrays (bases) 1,447,192 (9.8%) 1,369,611 (9.3%) 2,163,303 (14.7%) Percentages are of total bases in ENCODE in the first row and bases tiled in arrays in the second row. NATURE| Vol 447|14 June 2007 ARTICLES 801 ©2007 NaturePublishingGroup

ARTICLES NATURE Vol 447 14 June 2007 protein structure. Such exons are on average less expressed (25% detected using RACE followed by hybridization to tiling arrays as versus 87% by RT-PCR; see Supplementary Information section 2.7) Rx Frags. We performed RACE to examine 399 protein-coding loci than exons involved in more than one transcript(see Supple- (those loci found entirely in ENCODE regions)using RNA derived mentary Information section 2.4.3), but when expressed have a tissue from 12 tissues, and were able to unambiguously detect 4,573 distribution comparable to well-established genes. RxFrags for 359 loci(see Supplementary Information section 2.9) Critical questions are raised by the presence of a large amount of Almost half of these RxFrags (2, 324)do not overlap a GENCODE unannotated transcription with respect to how the corresponding exon, and most(90%)loci have at least one novel RxFrag, which sequences are organized in the genome--do these reflect longer tran- often extends a considerable distance beyond the 5 end of the locus. ripts that include known loci, do they link known loci, or are they Figure 2 shows the distribution of distances between these new mpletely separate from known loci? We further investigated these RACE-detected ends and the previously annotated TSS of each locus. issues using both computational and new experimental techniques. The average distance of the extensions is between 50 kb and 100 kb, Unannotated transcription. Consistent with previous findings, the with many extensions(20%)being more than 200 kb. Consistent UnT Exsa information section 2.8). One might expect Un Tx Frags our findings reveal evidence for an overlapping gene at 224 loci, with did not show evidence of encoding proteins(see Sup- with the known presence of overlapping genes in the human genome, ent to be linked within transcripts that exhibit coordinated expression transcripts from 180 of these loci (-50% of the RACE-positive loci) and have similar conservation profiles across species. To test this, we appearing to have incorporated at least one exon from an upstream clustered Un Tx Frags using two methods. The first methodused gene expression levels in 11 cell lines or conditions, dinucleotide composi- To characterize further the 5 Rx Frag extensions, we performed tion, location relative to annotated genes, and evolutionary conser- RT-PCR followed by cloning and sequencing for 550 of the 5 vation profiles to cluster Tx Frags(both unannotated and annotated ) RxFrags(including the 261 longest extensions identified for each loci,and 21% could be clustered into 200 novel loci (with an average is a combination method previously described and validated in sev- of -7TxFrags per locus). We experimentally examined these novel eral studies 4.170 Hybridization of the RT-PCR products to tiling loci to study the connectivity of transcripts amongst Un Tx Frags and arrays confirmed connectivity in almost 60%of the cases. Sequenced between UnTx Frags and known exons. Overall, about 40% of the clones confirmed transcript extensions. Longer extensions were connections(18 out of 46)were validated by RT-PCR. The second harder to clone and sequence, but 5 out of 18 RT-PCR-positive clustering method involved analysing a time course(0, 2, 8 and 32 h) extensions over 100 kb were verified by sequencing(see Supple- of expression changes in human HL60 cells following retinoic-acid mentary Information section 2.9.7 and ref. 17). The detection of stimulation. There is a coordinated program of expression changes numerous RxFrag extensions coupled with evidence of considerable from annotated loci, which can be shown by plotting Pearson intronic transcription indicates that protein-coding loci are more correlation values of the expression levels of exons inside annotated transcriptionally complex than previously thought. Instead of the loci versus unrelated exons(see Supplementary Information sec- traditional view that many genes have one or more alternative tran Un TxFrags, albeit lower, though still significantly different from gene may both encode multiple protein products and produce other randomized sets. Both clustering methods indicate that there is coor- transcripts that include sequences from both strands and from neigh dinated behaviour of many Un. Tx Frags, consistent with them res- bouring loci(often without encoding a different protein).Figure 3 ding in connected transcripts illustrates such a case, in which a new fusion transcript is expressed in Transcript connectivity. We used a combination of RACe and tiling he small intestine, and consists of at least three coding exons from rrayszo to investigate the diversity of transcripts emanating from the ATP50 gene and at least two coding exons from the DONSON protein-coding loci. Analogous to TxFrags, we refer to transcript 1/112113114115/1 a Intronic proximal hill 宽×889x6 Figure 1 Annotated and unannotated TxFrags detected in different cell lines. The proportion of different types of transcripts detected in the indicated number of cell lines(from 1/ll at the far left to 11/11 at the far t)is shown. The data for annotated and unannotated TxFrags are indicated separately, and also split into different cat based on Extension length(kb) GENCODE classification: exonic, intergenic(proximal being within 5kb of a Figure 2 Length of genomic extensions to GENCODE-annotated gene and distal being otherwise), intronic(proximal being within 5 kb of an the basis of RACE experiments followed by array hybridizations ( intron and distal being otherwise), and matching other ESTs not used in the The indicated bars reflect the frequency of extension lengths amon GENCODE annotation(principally because they were unspliced). The yaxis length classes. The solid line shows the cumulative frequency of indicates the per cent of tiling array nucleotides present in that class for that of that length or greater. Most of the extensions are greater than 50kb from number of samples(combination of cell lines and tissues the annotated gene(see text for details) E2007 Nature Publishing Group

protein structure18. Such exons are on average less expressed (25% versus 87% by RT–PCR; see Supplementary Information section 2.7) than exons involved in more than one transcript (see Supple￾mentary Information section 2.4.3), but when expressed have a tissue distribution comparable to well-established genes. Critical questions are raised by the presence of a large amount of unannotated transcription with respect to how the corresponding sequences are organized in the genome—do these reflect longer tran￾scripts that include known loci, do they link known loci, or are they completely separate from known loci? We further investigated these issues using both computational and new experimental techniques. Unannotated transcription. Consistent with previous findings, the Un.TxFrags did not show evidence of encoding proteins (see Sup￾plementary Information section 2.8). One might expect Un.TxFrags to be linked within transcripts that exhibit coordinated expression and have similar conservation profiles across species. To test this, we clustered Un.TxFrags using two methods. The first method19 used expression levels in 11 cell lines or conditions, dinucleotide composi￾tion, location relative to annotated genes, and evolutionary conser￾vation profiles to cluster TxFrags (both unannotated and annotated). By this method, 14% of Un.TxFrags could be assigned to annotated loci, and 21% could be clustered into 200 novel loci (with an average of ,7 TxFrags per locus). We experimentally examined these novel loci to study the connectivity of transcripts amongst Un.TxFrags and between Un.TxFrags and known exons. Overall, about 40% of the connections (18 out of 46) were validated by RT–PCR. The second clustering method involved analysing a time course (0, 2, 8 and 32 h) of expression changes in human HL60 cells following retinoic-acid stimulation. There is a coordinated program of expression changes from annotated loci, which can be shown by plotting Pearson correlation values of the expression levels of exons inside annotated loci versus unrelated exons (see Supplementary Information sec￾tion 2.8.2). Similarly, there is coordinated expression of nearby Un.TxFrags, albeit lower, though still significantly different from randomized sets. Both clustering methods indicate that there is coor￾dinated behaviour of many Un.TxFrags, consistent with them res￾iding in connected transcripts. Transcript connectivity. We used a combination of RACE and tiling arrays20 to investigate the diversity of transcripts emanating from protein-coding loci. Analogous to TxFrags, we refer to transcripts detected using RACE followed by hybridization to tiling arrays as RxFrags. We performed RACE to examine 399 protein-coding loci (those loci found entirely in ENCODE regions) using RNA derived from 12 tissues, and were able to unambiguously detect 4,573 RxFrags for 359 loci (see Supplementary Information section 2.9). Almost half of these RxFrags (2,324) do not overlap a GENCODE exon, and most (90%) loci have at least one novel RxFrag, which often extends a considerable distance beyond the 59 end of the locus. Figure 2 shows the distribution of distances between these new RACE-detected ends and the previously annotated TSS of each locus. The average distance of the extensions is between 50 kb and 100 kb, with many extensions (.20%) being more than 200 kb. Consistent with the known presence of overlapping genes in the human genome, our findings reveal evidence for an overlapping gene at 224 loci, with transcripts from 180 of these loci (,50% of the RACE-positive loci) appearing to have incorporated at least one exon from an upstream gene. To characterize further the 59 RxFrag extensions, we performed RT–PCR followed by cloning and sequencing for 550 of the 59 RxFrags (including the 261 longest extensions identified for each locus). The approach of mapping RACE products using microarrays is a combination method previously described and validated in sev￾eral studies14,17,20. Hybridization of the RT–PCR products to tiling arrays confirmed connectivity in almost 60% of the cases. Sequenced clones confirmed transcript extensions. Longer extensions were harder to clone and sequence, but 5 out of 18 RT–PCR-positive extensions over 100 kb were verified by sequencing (see Supple￾mentary Information section 2.9.7 and ref. 17). The detection of numerous RxFrag extensions coupled with evidence of considerable intronic transcription indicates that protein-coding loci are more transcriptionally complex than previously thought. Instead of the traditional view that many genes have one or more alternative tran￾scripts that code for alternative proteins, our data suggest that a given gene may both encode multiple protein products and produce other transcripts that include sequences from both strands and from neigh￾bouring loci (often without encoding a different protein). Figure 3 illustrates such a case, in which a new fusion transcript is expressed in the small intestine, and consists of at least three coding exons from the ATP5O gene and at least two coding exons from the DONSON 1/11 2/11 3/11 4/11 5/11 6/11 7/11 8/11 9/11 10/11 11/11 cell lines Intronic proximal Intronic distal Intergenic proximal Intergenic distal Other ESTs GENCODE exonic 12 Annotated transcripts Novel transcripts 10 8 6 4 2 0 2 Tiling array nucleotides (%) 4 6 8 10 12 Figure 1 | Annotated and unannotated TxFrags detected in different cell lines. The proportion of different types of transcripts detected in the indicated number of cell lines (from 1/11 at the far left to 11/11 at the far right) is shown. The data for annotated and unannotated TxFrags are indicated separately, and also split into different categories based on GENCODE classification: exonic, intergenic (proximal being within 5 kb of a gene and distal being otherwise), intronic (proximal being within 5 kb of an intron and distal being otherwise), and matching other ESTs not used in the GENCODE annotation (principally because they were unspliced). The y axis indicates the per cent of tiling array nucleotides present in that class for that number of samples (combination of cell lines and tissues). Per cent of RxFrag extensions (shaded boxes) 0 5 10 15 Extension length (kb) Cumulative per cent of extensions this length or greater (line) < 0.5 0.5–1 5–10 10–25 25–50 50–100 100–200 200–300 300–400 400–500 ≥ 1–5 500 0 10 20 30 40 50 60 70 80 90 100 Figure 2 | Length of genomic extensions to GENCODE-annotated genes on the basis of RACE experiments followed by array hybridizations (RxFrags). The indicated bars reflect the frequency of extension lengths among different length classes. The solid line shows the cumulative frequency of extensions of that length or greater. Most of the extensions are greater than 50 kb from the annotated gene (see text for details). ARTICLES NATURE|Vol 447| 14 June 2007 802 ©2007 NaturePublishingGroup

NATURE Vol 447 14 June 2007 ARTICLES ch.2133.900000,33950000134000.000 34,150000 ATP50 H+ ↓4‖↓ Figure 3 Overview of RACE experiments showing a gene fusion. ray analyses(RxFrags) are shown along the top. Along th Transcripts emanating from the region between the doNSON and ATP5O genes. A 330-kbinterval ofhuman chromosome 21(within ENm005)is shot om the DONSON gene f and sequenced RT-PCR productt ollowed by three exons from the APso genes which contains four annotated genes: DONSON, CRYZLI, ITSNI and ATP50 ences are separated by a 300 kb intron in the genome. A PET tag The 5" RACE products generated from small intestine RNA and detected by termini of a transcript consistent with this RT-PCR product. gene, with no evidence of sequences from two intervening protein- Information sections 2 11 and 2.9.3); the predictions were validated PseudogenesPseudogenes,reviewed in refs 21 and 22, are generally respec%, and 63% rate for Evofold, RNAz and dual predictions, coding genes(ITSNI and CRYZLI) ata56%,65% Dat of genes, are sometimes tran- Primary transcripts. The detection of numerous unannotated scribed and often complicate analysis of transcription owing to close transcripts coupled with increasing knowledge of the general com- ei quence similarity to functional genes. We used various computa- plexity of transcription prompted us to examine the extent of prim onal methods to identify 201 pseudogenes(124 processed and 77 ary(that is, unspliced) transcripts across the ENCODE regions. non-processed)in the ENCODE regions(see Supplementary Infor- Three data sources provide insight about these primary transcripts mation section 2.10 and ref 23). Tiling-array analysis of 189 of these the GENCODE annotation, PETs, and RxFrag extensions. Figure 4 revealed that 56% overlapped at least one TxFrag. However, possible summarizes the fraction of bases in the ENCODE regions that over- cross-hybridization between the pseudogenes and their correspond- lap transcripts identified by these technologies. Remarkably, 93% of ing parent genes may have confounded such analyses. To assess better bases are represented in a primary transcript identified by at least two the extent of pseudogene transcription, 160 pseudogenes(lll pro- independent observations(but potentially using the same techno- cessed and 49 non-processed)were examined for expression using logy ) this figure is reduced to 74% in the case of primary transcripts RACE/tiling-array analysis(see Supplementary Information section detected by at least two different technologies. These increased spans 2.9.2). Transcripts were detected for 14 pseudogenes( 8 processed are not mainly due to cell line rearrangements because they were and 6 non-processed)in at least one of the 12 tested RNA sources, present in multiple tissue experiments that confirmed the spans the majority(9)being in testis(see ref. 23). Additionally, there was (see Supplementary Information section 2.12). These estimates evidence for the transcription of 25 pseudogenes on the basis of their assume that the presence of PETs or RxFrags defining the terminal proximity(within 100 bp of a pseudogene end)to CAGE tags(8), ends of a transcript imply that the entire intervening DNA is tran- PETs(2), or cDNAS/ESTs(21). Overall, we estimate that at least 19% scribed and then processed. Other mechanisms, thought to be of the pseudogenes in the ENCODE regions are transcribed, which is unlikely in the human genome, such as trans-splicing or polymerase consistent with previous estimates umping would also produce these long termini and potentially Non-protein-coding RNA Non-protein-coding RNAs(ncRNAs) should be reconsidered in more detail. clude structural RNAs(for example, transfer RNAs, ribosomal Previous studies have suggested a similar broad amount of tran RNAS, and small nuclear RNAs) and more recently discovered scription across the human 4 and mouse2genomes. Our studies regulatory RNAs(for example, miRNAs). There are only 8 well- confirm these results, and have investigated the genesis of these characterized ncRNA genes within the ENCODE regions (U70, transcripts in greater detail, confirming the presence of substantial ACA36, ACA56, mir-192, mir-194-2, mir-196, mir-483 and H19), intragenic and intergenic transcription. At the same time, many of whereas representatives of other classes, (for example, box C/D the resulting transcripts are neither traditional protein-coding snoRNAs, tRNAs, and functional snRNAs)seem to be completel absent in the ENCODE regions. Tiling-array data provided evidence for transcription in at least one of the assayed rna samples for all of one observation One techn hese ncRNAs, with the exception of mir-483(expression of mir-483 might be specific to fetal liver, which was not tested). There is also two observations evidence for the transcription of 6 out of 8 pseudogenes of ncRNA: (mainly snoRNA-derived ). Similar to the analysis of protein pseudogenes, the hybridization results could also originate from All three the known snoRNa gene elsewhere in the genome Many known nCRNAs are characterized by a well-defined RNA secondary structure. We applied two de novo ncRNA prediction algorithms--EvoFold and RNAz--to predict structured ncRNAs (as well as functional structures in mRNAs)using the multi-species sequence alignments(see below, Supplementary Information section 2. 11 and ref. 26). Using a sensitivity threshold capable of detecting all Figure 4 Coverage of primary transcripts across ENCODE region known miRNAs and snoRNAs, we identified 4986 and 3.707 can- different technologies(integrated annotation from GENCODE, R didate ncRNA loci with Evo Fold and RNAZ, respectively. Only 268 experiments (RxFrags)and PET tags)were used to assess the pr loci(5% and 7%, respectively) were found with both program representing a 1. 6-fold enrichment over that expected by chance; opportunity to have multiple observations of each finding. The proportion the lack of more extensive overlap is due to the two programs having the following scenarios is depicted: detected by all three technologies, by two e experimentally exami50 hese targets using RACE/ and by one technologies, by one technology but wi四m山k optimal sensitivity at different levels of GC content and conservation. of th iling-array analysis for brain and testis tissues(see Supplementary genomic bases without any detectable coverage of primary transcripts. E2007 Nature Publishing Group

gene, with no evidence of sequences from two intervening protein￾coding genes (ITSN1 and CRYZL1). Pseudogenes. Pseudogenes, reviewed in refs 21 and 22, are generally considered non-functional copies of genes, are sometimes tran￾scribed and often complicate analysis of transcription owing to close sequence similarity to functional genes. We used various computa￾tional methods to identify 201 pseudogenes (124 processed and 77 non-processed) in the ENCODE regions (see Supplementary Infor￾mation section 2.10 and ref. 23). Tiling-array analysis of 189 of these revealed that 56% overlapped at least one TxFrag. However, possible cross-hybridization between the pseudogenes and their correspond￾ing parent genes may have confounded such analyses. To assess better the extent of pseudogene transcription, 160 pseudogenes (111 pro￾cessed and 49 non-processed) were examined for expression using RACE/tiling-array analysis (see Supplementary Information section 2.9.2). Transcripts were detected for 14 pseudogenes (8 processed and 6 non-processed) in at least one of the 12 tested RNA sources, the majority (9) being in testis (see ref. 23). Additionally, there was evidence for the transcription of 25 pseudogenes on the basis of their proximity (within 100 bp of a pseudogene end) to CAGE tags (8), PETs (2), or cDNAs/ESTs (21). Overall, we estimate that at least 19% of the pseudogenes in the ENCODE regions are transcribed, which is consistent with previous estimates24,25. Non-protein-coding RNA. Non-protein-coding RNAs (ncRNAs) include structural RNAs (for example, transfer RNAs, ribosomal RNAs, and small nuclear RNAs) and more recently discovered regulatory RNAs (for example, miRNAs). There are only 8 well￾characterized ncRNA genes within the ENCODE regions (U70, ACA36, ACA56, mir-192, mir-194-2, mir-196, mir-483 and H19), whereas representatives of other classes, (for example, box C/D snoRNAs, tRNAs, and functional snRNAs) seem to be completely absent in the ENCODE regions. Tiling-array data provided evidence for transcription in at least one of the assayed RNA samples for all of these ncRNAs, with the exception of mir-483 (expression of mir-483 might be specific to fetal liver, which was not tested). There is also evidence for the transcription of 6 out of 8 pseudogenes of ncRNAs (mainly snoRNA-derived). Similar to the analysis of protein￾pseudogenes, the hybridization results could also originate from the known snoRNA gene elsewhere in the genome. Many known ncRNAs are characterized by a well-defined RNA secondary structure. We applied two de novo ncRNA prediction algorithms—EvoFold and RNAz—to predict structured ncRNAs (as well as functional structures in mRNAs) using the multi-species sequence alignments (see below, Supplementary Information section 2.11 and ref. 26). Using a sensitivity threshold capable of detecting all known miRNAs and snoRNAs, we identified 4,986 and 3,707 can￾didate ncRNA loci with EvoFold and RNAz, respectively. Only 268 loci (5% and 7%, respectively) were found with both programs, representing a 1.6-fold enrichment over that expected by chance; the lack of more extensive overlap is due to the two programs having optimal sensitivity at different levels of GC content and conservation. We experimentally examined 50 of these targets using RACE/ tiling-array analysis for brain and testis tissues (see Supplementary Information sections 2.11 and 2.9.3); the predictions were validated at a 56%, 65%, and 63% rate for Evofold, RNAz and dual predictions, respectively. Primary transcripts. The detection of numerous unannotated transcripts coupled with increasing knowledge of the general com￾plexity of transcription prompted us to examine the extent of prim￾ary (that is, unspliced) transcripts across the ENCODE regions. Three data sources provide insight about these primary transcripts: the GENCODE annotation, PETs, and RxFrag extensions. Figure 4 summarizes the fraction of bases in the ENCODE regions that over￾lap transcripts identified by these technologies. Remarkably, 93% of bases are represented in a primary transcript identified by at least two independent observations (but potentially using the same techno￾logy); this figure is reduced to 74% in the case of primary transcripts detected by at least two different technologies. These increased spans are not mainly due to cell line rearrangements because they were present in multiple tissue experiments that confirmed the spans (see Supplementary Information section 2.12). These estimates assume that the presence of PETs or RxFrags defining the terminal ends of a transcript imply that the entire intervening DNA is tran￾scribed and then processed. Other mechanisms, thought to be unlikely in the human genome, such as trans-splicing or polymerase jumping would also produce these long termini and potentially should be reconsidered in more detail. Previous studies have suggested a similar broad amount of tran￾scription across the human14,15 and mouse27 genomes. Our studies confirm these results, and have investigated the genesis of these transcripts in greater detail, confirming the presence of substantial intragenic and intergenic transcription. At the same time, many of the resulting transcripts are neither traditional protein-coding No coverage One technology, one observation One technology, two observations Two technologies All three technologies Figure 4 | Coverage of primary transcripts across ENCODE regions. Three different technologies (integrated annotation from GENCODE, RACE-array experiments (RxFrags) and PET tags) were used to assess the presence of a nucleotide in a primary transcript. Use of these technologies provided the opportunity to have multiple observations of each finding. The proportion of genomic bases detected in the ENCODE regions associated with each of the following scenarios is depicted: detected by all three technologies, by two of the three technologies, by one technology but with multiple observations, and by one technology with only one observation. Also indicated are genomic bases without any detectable coverage of primary transcripts. 33,900,000 33,950,000 34,000,000 34,050,000 34,100,000 34,150,000 34,200,000 RxFrag DONSON CRYZL1 ATP5O PETs (–) strand (–) strand (+) strand ITSN1 DONSON Cloned RT-PCR product ATP5O Chr. 21 GENCODE reference genes Figure 3 | Overview of RACE experiments showing a gene fusion. Transcripts emanating from the region between the DONSON and ATP5O genes. A 330-kbinterval of human chromosome 21 (within ENm005) is shown, which contains four annotated genes:DONSON,CRYZL1,ITSN1 andATP5O. The 59 RACE products generated from small intestine RNA and detected by tiling-array analyses (RxFrags) are shown along the top. Along the bottom is shown the placement of a cloned and sequenced RT–PCR product that has two exons from the DONSON gene followed by three exons from the ATP5O gene; these sequences are separated by a 300 kb intron in the genome. A PET tag shows the termini of a transcript consistent with this RT–PCR product. NATURE| Vol 447|14 June 2007 ARTICLES 803 ©2007 NaturePublishingGroup