Chapter 14Genome Assembly and Annotation Process

Genomic Sequences Used for Assembly

Genomic sequences from the following five data sets are processed for use in the assembly:

High-Throughput Genomic Sequences. Human high-throughput genomic sequences (HTGSs) are retrieved using the Entrez query system (Chapter 15). The query used returns sequences for all entries that contain the HTG keyword, regardless of whether the sequence is finished or is in any of the unfinished draft phases.

Finished Chromosome Sequences. The center that coordinates the sequencing of a finished chromosome submits a specification regarding how to build the sequence of that chromosome from its component clone sequences to a data repository at the European Bioinformatics Institute (EBI). The sequences specified for any finished chromosomes are retrieved from GenBank and included in the set of genomic sequences to be processed for an assembly.

Genomic Sequences from the Tiling Paths of Individual Chromosomes. The human genome sequencing centers use a variety of experimental evidence to compile an ordered list of clones they believe provides the best coverage for each chromosome. At least once every 2 months, the sequencing centers submit an updated minimal tiling path for each chromosome to a data repository at the EBI. These tiling path files (TPFs) include Accession numbers if sequence for a clone is available. The tiling path repository is checked each day for the Accession numbers from any tiling path that has been updated. Any secondary Accession numbers are replaced with the corresponding primary Accession numbers, and any invalid Accession numbers are flagged to prevent those sequences from being used for assembly. The latest version of the sequence for each Accession number in the most recent clone tiling paths is retrieved from GenBank and included in the set of genomic sequences to be processed for assembly.

Assembled Blocks of Contiguous Finished Genomic Sequence. As the sequences for individual clones are finished, they are merged with overlapping finished sequences to form contigs (3). The primary source for identifying neighboring clones is the clone tiling path for each chromosome. Additional information is obtained from some GenBank entries that contain annotation specifying the neighboring clones. BLAST (4) is used to align the sequences of candidate pairs of clones, and a merged sequence is produced automatically if the expected overlap is confirmed by the sequence alignment. When the automatic processes do not find an expected overlap, there is a manual review to find the correct overlap, refining the clone order if necessary. The most recent set of finished contigs are processed for assembly.

Additional Genomic Sequences. A few other specific human genomic sequences are added to the assembly set because they contain genes that may not be represented in the genomic sequences from the other sources.

Genomic Sequences Used for Ordering and Orienting

Sequences known to come from both ends of the same cloned genomic fragment provide valuable linking information that helps to order and orient sequence contigs in the assembly step (5, 6). The SNP Consortium sequenced the ends of the inserts in several million plasmid clones containing small (0.8–6 Kbp) fragments of human genomic DNA. In many cases, both ends of the same insert were sequenced (Table 4 in Ref. 7), thereby providing a set of plasmid paired-end sequences.

Genomic Sequences Used for Annotation

Curated Genomic Regions. The Reference Sequence project (Chapter 18; Refs. 8, 9) provides reviewed annotated sequences for a number of genomic regions that are difficult to annotate correctly by automated processes (e.g., immunoglobulin gene regions). These Reference Sequences (RefSeqs) are aligned to the assembled genome so that the curated annotation can be transferred to the assembled genomic sequence. RefSeqs for known pseudogenes are also aligned, not only to enable transfer of the correct annotation but, more importantly, to prevent prediction of erroneous model transcripts and proteins.

BAC End Sequences. Sequences from the ends of human genomic inserts in Bacterial Artificial Chromosome (BAC) clones are used to help map the location of specific clones onto the assembled genome sequence. The BAC end sequences are obtained from dbGSS (see Chapter 1). The clone names are extracted and converted to a standardized format to facilitate linking of the BAC end sequences with mapping data and additional sequences for the same clone, when these are available.

Genomic Sequences Used for Alignment

Human genomic sequences that are not used for either assembly or annotation are processed so that their relationships to the assembled genome can be displayed in Map Viewer (Chapter 20). Most genomic sequences deposited in GenBank by individual scientists will not be HTG and therefore will not be used for assembly; however, they are used for alignment. The exceptions are a few non-HTG sequences that are used for assembly, because they are included in the clone tiling path of an individual chromosome or in an assembled block of finished sequence. Any sequences intended for assembly but not used, either because they are redundant or are rejected by one of the quality screens in the assembly process, are also used for alignment.

Removing Contaminants

Draft-quality HTGSs sometimes contain segments of sequence derived from foreign sources, most commonly the cloning vector or bacterial host. Finished sequences are usually, but not always, free of such contaminants. Common contaminants introduce artificial blocks of homologous sequence that can give rise to misleading alignments between two unrelated genomic sequences. MegaBLAST (10) is used to compare the raw genomic sequences to a database of contaminant sequences (including the UniVec database of vector sequences, the Escherichia coli genome, bacterial insertion sequences, and bacteriophage). Any foreign segments are removed from draft-quality sequence or masked in finished sequence to prevent them from participating in alignments.

Masking of Repetitive Sequences

Sequences that occur in many copies in the genome will align to many different clones. Such repetitive sequences include interspersed repeats (SINEs, LINEs, LTR elements, and DNA transposons), satellite sequences, and low-complexity sequences (7, 11, 12). Matches between repetitive sequences on unrelated clones make it difficult to identify alignments that indicate a genuine overlap between clones. To eliminate the confounding matches that are based only on repetitive sequences, the genomic sequences are run through RepeatMasker to identify known repeats. Repeats are masked by converting the sequence to lowercase letters so that they do not initiate alignments.

Fragmentation of Draft Sequences

Draft HTGSs consist of a set of sequence contigs derived from a particular clone artificially linked together to form a single sequence. The masked, draft genomic sequences are split at the gaps between their constituent contigs to create separate sequence fragments that can be aligned independently. Vector sequences and other contaminants are also removed at this stage by trimming or further splitting the sequence fragments.

Determination of STS Content

Any STS markers contained within the input genomic sequences are identified by e-PCR (13) using the UniSTS database. The resulting data are used primarily to relate the genomic sequences to independently derived STS maps (genetic, radiation hybrid) but are also used to identify some foreign sequences.

Filtering

Sequences from other clones being sequenced at the same institution can occasionally cross-contaminate draft HTGSs. The contaminating sequences may come from another clone from the same organism or from another organism. The raw genomic sequences are screened in several ways to detect cross-contamination: (a) they are compared with the genome sequences from completely sequenced organisms using MegaBLAST (10); (b) they are screened for the presence of organism-specific interspersed repeats using RepeatMasker; and (3) they are screened for the presence of mapped STS markers from other organisms using e-PCR (13). Any input sequence that contains foreign sequences, repeats, or markers is flagged for removal from the data set used for assembly. Draft sequences longer than the maximum insert length expected for a genomic clone are also rejected because it is likely they are contaminated with sequences from at least one other clone.

At this stage, draft sequences composed of fragments that are too small to contribute significantly to the assembly or that are tagged with the HTGS_CANCELLED keyword are also flagged for removal. Another filter rejects sequences annotated as being from another organism or as being RNA, erroneously included in the input sequences.

Chromosome Assignment

To improve assembly of the genomic sequences, the input genomic sequences are assigned to a specific chromosome before attempting to merge the sequences. Genomic sequences that appear on any of the chromosome tiling paths are automatically assigned to the designated chromosome. Other genomic sequences are assigned to a chromosome based on: (a) annotation on the submitted GenBank record; (b) the presence of multiple STS markers that have been mapped to the same chromosome; (c) fluorescence in situ hybridization (FISH) mapping (14, 15); or (d) personal communication from a scientist with specialized knowledge. If there is no assignment, or the assignments are conflicted, the sequences are treated as unassigned and assembled without constraint by chromosome.

Resolution of Conflicts in the Chromosome Tiling Paths

Before the tiling paths are used in the genome assembly, the order of the finished clone sequences included in the tiling paths is compared with the specifications used to assemble the curated contigs of finished sequence. Discrepancies are resolved before proceeding with the assembly. Sequence from any clone should appear at just one place in the assembled genome; therefore, if a clone is listed more than once in the tiling paths, only the location with the best evidence is used in the assembly step.

Genomic Sequences Excluded from the Assembly Step

Clone sequences that consist only of unassembled reads (HTGS_PHASE0) or that were flagged because of suspected cross-contamination or other problem detected in the pre-assembly screens are not used in the assembly step.

Assembly of the Genomic Sequence Contigs

Adjacent, finished clone sequences from the chromosome tiling path that have good sequence overlap are merged. Tiling path draft sequences that are adjacent to and overlap the finished clone sequences or other draft clone sequences are added to extend the initial genomic sequence contigs. After that, genomic sequences from clones not on any chromosome tiling path are added, provided they have good overlaps with the assembled tiling path clones. Genomic sequences from additional clones may be added if they provide the sequence for a known gene that is missing from the existing genomic sequence contigs. Finally, the individual fragments of draft sequences are ordered and oriented.

Assembly of Finished Sequences from Tiling Path Clones. The quality of any overlaps between finished clone sequences that are adjacent in the clone tiling path are assessed using the alignments between pairs of genomic sequences that were produced in advance. Sequences that have high-quality overlaps, or that are known from annotation or other data to abut, are merged to form a genomic sequence contig. Clone sequences that have no good overlaps are retained as separate contigs.

Addition of Draft Sequences from Tiling Path Clones. The procedure used for merging draft sequence from tiling path clones is similar to that described for merging finished sequences, except that the minimum-overlap quality required for merging is different. An overlap involving a draft sequence can contain more mismatches, but must be longer, than an overlap between two finished sequences. Preference is given to finished sequences so that a contig made by merging finished and draft sequences will contain the finished sequence for the overlapping portion. Draft clone sequences that have no good overlaps are retained as separate contigs.

Addition of Sequences from Other Genomic Clones. Genomic sequences from clones that are not on any chromosome tiling path are used to close, or extend into, gaps in the backbone of genomic contigs assembled based on the tiling paths. Genomic sequences that are fully contained within the existing genomic contigs are not used. Any remaining genomic sequences that were either assigned to the relevant chromosome or could not be assigned to any chromosome are evaluated, and sequences that have good-quality overlaps with genomic contigs are merged in to extend a contig or to join two adjacent contigs, if two additional conditions are met: (1) the gap must be sufficiently large to accommodate the additional sequence; and (2) the Sequence Tagged Site (STS) marker content of the additional clone sequence must be compatible with that of the flanking clone sequences when compared with various STS maps.

After all of the chromosomes have been assembled, any remaining genomic clone sequence that contains a known gene not present in the other contigs is added to the assembly as a separate contig.

Ordering and Orienting Draft Sequence Fragments. The order and orientation of the fragments of HTGS_PHASE1 draft sequence need to be defined before sequence made from contigs that include this category of draft sequence can be completed. Some fragments may be ordered and oriented by overlaps with sequences from adjacent clones. Many more can be defined by aligning them with mRNAs, ESTs, or plasmid paired-end sequences. Any fragments whose order and orientation remain undefined are placed in the nearest open gap and given an arbitrary orientation. Fragments of draft sequence are connected to flanking sequences and to each other by runs of 100 unknown bases (Ns), which represent an arbitrarily sized gap in the sequence.

Placement of the Genomic Contigs

After the genomic sequence contigs are assembled, they are oriented and placed in order along each chromosome with appropriately sized gaps inserted between adjacent contigs. The chromosome tiling paths specify the order of the clone sequences and the sizes for some gaps. Therefore, the order and orientation of most of the genomic sequence contigs are derived from the tiling paths. Many of the remaining contigs are placed by comparison of the STS marker content of the contigs, as determined by e-PCR (13) using the UniSTS database, to various STS maps. There are some contigs that can be assigned to a specific chromosome but cannot be placed along that chromosome. Others cannot even be assigned to a specific chromosome and therefore remain unplaced within the genome assembly.

Gaps between the clone contigs laid out in the chromosome tiling paths are arbitrarily set at 50 Kbp, and 3 Mbp for the centromere, unless another gap size is specified in the tiling path. Any remaining gaps between genomic sequence contigs are arbitrarily set to 10 Kbp.

Alignment of Transcripts to the Assembled Genome

The alignments between RefSeq RNA sequences, mRNA and EST sequences from GenBank and the component genomic sequences are remapped to produce alignments of these transcripts to the assembled genomic contigs.

Production of Candidate Gene Models

A candidate gene model is produced from each set of alignments between a particular transcript and one strand of a particular genomic contig as follows: (1) putative exons are identified by looking for mRNA splice sites near the ends of those alignments that satisfy minimum length and percentage identity criteria; (2) a mutually compatible set of exons for the model is selected by applying rules, such as restrictions on the size of an intron, that define plausible exon–intron structures; and (3) BLAST (4) may be used to produce additional alignments to try to identify exons that were missed because they were too short to be represented in the initial set of transcript alignments. Candidate gene models are only retained if good-quality alignments between their exons and the defining transcript cover either more than half the length of the transcript or more than 1 Kbp.

Selection of the Best RefSeq RNA-based Gene Models

Each RefSeq RNA represents a distinct transcript produced from a particular gene (Chapter 18; Refs. 8, 9). Hence, there should not be more than one gene model corresponding to any given RefSeq RNA. Therefore, all gene models based on a particular RefSeq RNA are compared, and the best one is selected. Because the RefSeq RNA is taken to be the best representation of a particular transcript, this gene model is preserved without any further modification. Any extra models may represent paralogs; therefore, they are included with the mRNA- and EST-based models for further processing. Between builds, RefSeq RNAs are refined based on a review of related gene models and transcript alignments produced during the genome annotation process.

Exon Refinement

Many gene models may be produced for the same gene because the input data set frequently contains multiple EST or mRNA sequences representing the same transcript. This redundancy is used to refine the splice sites defining a particular exon. Similar exons are clustered, and splice sites may be adjusted in some models to match those used by the majority of models containing the same exon. Inconsistent models may be discarded at this stage, unless they have sufficient support to be retained as likely splice variants.

Chaining of Transcript-based Gene Models

Many of the mRNAs and most of the ESTs used to generate the initial gene models provide sequence for only part of the native transcript. Overlapping gene models that are compatible with each other are combined into an extended model. This chaining step produces models more likely to represent the full gene.

Dividing the Genomic Sequences into Segments

GenomeScan produces better results when long genomic sequences are broken into shorter segments at putative gene boundaries. The locations of gene models based on RefSeq RNA alignments are, therefore, used to divide the assembled genomic contigs into segments. Repetitive sequences are masked by remapping the repeats found in the component genomic sequences.

Producing Protein Hints for GenomeScan

GenomeScan can use data on protein homologies to improve its gene predictions (17). The locations of genomic sequences that potentially code for polypeptides with homology to other proteins are obtained from three sources. Significant alignments between translated genomic segments and vertebrate proteins are obtained by filtering and remapping the precomputed alignments. Significant alignments between translated genomic sequences and conserved protein domains are obtained in the same manner. A third set of alignments comes from running GenomeScan without any hints. The proteins predicted by this initial run are aligned to proteins from SWISS-PROT (18) and NCBI RefSeq proteins (8, 9) using blastp (4). The eukaryotic protein with the best match is then aligned to the genomic sequence segments using tblastn (4). These three sets of data are converted into the format required by GenomeScan and merged to produce a single set of protein hints.

Predicting Genes Using GenomeScan

Each segment of genomic sequence is processed by GenomeScan using the combined set of protein homology-based hints as an additional input. This produces one model containing all of the predicted exons for each putative gene. Models with coding sequences shorter than 90 amino acids are discarded. Each remaining model is aligned to proteins from SWISS-PROT and NCBI RefSeq proteins using blastp. The eukaryotic protein with the best match to any model is used as evidence for that model and to provide a clue as to the possible function of that model.

Gene Models Superceded by Curated Genomic Regions

The manually reviewed annotations from curated genomic region RefSeqs are used in preference to any corresponding gene models generated by automated processing. The curated genomic regions are aligned to the assembled genomic contigs by remapping the alignments between these RefSeqs and the component genomic sequences. Any gene model that significantly overlaps a segment of the assembled sequence that corresponds to a curated genomic region is discarded.

Gene Models Likely to Be Pseudogenes

When transcripts from a particular gene are aligned to the genomic sequences, they will align not only to the active copy of the gene but also to any segment of the genome containing a pseudogene derived from the active gene. Because model transcripts or model proteins that represent nontranscribed pseudogenes are undesirable, an attempt is made to identify and remove such models.

Whenever possible, alignments of RefSeqs for pseudogenes, either curated genomic regions or RNAs, are used to annotate pseudogenes. Some additional models derived from pseudogenes that are not yet represented by RefSeqs are eliminated by the following mechanism. All models based on the same supporting mRNA are compared with respect to the percent identity of the alignments and the number of exons. Only the model with the strongest evidence is retained.

Conflicting Gene Models

When two gene models are found to have an extensive overlap, then in general only the model with the stronger evidence is retained. However, models based on RefSeqs are always retained. Whereas any model not based on a RefSeq is discarded if it overlaps a model that is RefSeq based, two RefSeq-based models that overlap are both retained.

Relating the Model Transcripts to Known Transcripts

To provide continuity from build to build and to identify genes based on their predicted transcripts, MegaBLAST (10) is used to compare model RNAs to: (a) RefSeq RNAs; (b) mRNAs from GenBank; and (c) model RNAs from the previous build. These comparisons are reported as reciprocal best hits if: (a) they produce a significant hit; (b) no other model has a better hit to that particular RNA; and (c) no other RNA has a better hit to that particular model.

Relating the Model Proteins to Known Proteins

The eukaryotic proteins with the best match to each protein predicted by the annotation process are used to identify the best model for a possible gene and to assign a name to gene models that are novel. The proteins encoded by the models are aligned to proteins from SWISS-PROT (18), NCBI RefSeq proteins (8, 9), and the NCBI non-redundant protein database using blastp (4). The name of the eukaryotic protein with the best match, its sequence identifier, and match score are recorded for each predicted protein with a significant hit.

Maps Based on Sequence Coordinates

Maps that display those features annotated on the genomic sequence contigs (genes, STSs, clones, and SNPs) are generated by translating the positions of the features on the contigs into chromosome coordinates. Contig coordinates are translated into chromosome coordinates using the positions of the contigs along each chromosome, as determined in the genome assembly step. Using this same method, alignments between various sequences and the genomic contigs are translated into chromosome coordinates to produce additional maps that show the locations of the aligned sequences on the chromosomes. Maps generated from sequence alignments include maps that show the genomic positions of mRNA plus EST sequences, or genomic sequences from GenBank. The specifications used to build each genomic contig are also translated into chromosome coordinates to produce one map that shows the component sequences used to assemble each contig and another that simply shows the finished and draft sections of the contigs.

Maps Based on Other Coordinate Systems

Cytogenetic maps, genetic linkage maps, and radiation hybrid maps use different coordinate systems that are not based on sequence. To generate data for these types of maps, the locations of the map elements are listed in the coordinate system appropriate to each map. Map Viewer can scale maps defined in different coordinate systems so that they can de displayed side by side.