Genome Annotation

Josep F Abril, Genetics, Microbiology and Statistics Dep, University of Barcelona (UB), Barcelona, Catalonia, Spain and Institute of
Biomedicine of the University of Barcelona (IBUB), Barcelona, Catalonia, Spain
Sergi Castellano, Great Ormond Street Institute of Child Health (ICH), University College London (UCL), London, United Kingdom
and UCL Genomics, University College London (UCL), London, United Kingdom
r 2018 Elsevier Inc. All rights reserved.

“The genome sequence of an organism is an information resource unlike any that biologists have previously had access to. But the value of the
genome is only as good as its annotation. It is the annotation that bridges the gap from the sequence to the biology of the organism.” Lincoln Stein,
2001. Nat. Rev. Genet. 2(7), 493–503.

Introduction

The precise ordering of nucleotides and amino acids confer specific functional properties to biological sequences. Functional
and structural features can then be understood as regions contained within those sequences, defined as segments delimited by
an interval of positions, initial and terminal, on the coordinates axis provided by the sequence itself. In a wide sense, sequence
annotation is defined by the procedures leading to determine the location and properties of these sequence features. Genomes
are the organisms’ blueprints, encoding for all the functional elements that lead to a fully developed life form. Therefore,
genomes are complex and contain a mixture of features with different roles, which are interspersed with non-functional
sequences under no evolutionary constraint. Genome annotation is the process of identifying any functional element along the
DNA sequence of a genome, yet at initial stages often focuses on genes. The annotation gives meaning to the genome by
providing the location and function of genes (protein coding or otherwise) and regulatory regions, which underlie the biology
of living organisms. Completing the catalog of functional regions of the genome can facilitate further downstream analyses, as
for instance the assessment of the effect of mutations caused by single nucleotide variants in individuals or populations. As it
has been remarked since the start of the genomics era, raw genomes are thus of limited use to scientists and their annotation
has become central to genome research.
Genome annotation itself has evolved over the years and it is possible to distinguish three, possibly four, major stages based
on the techniques and data used. Each succeeding, yet overlapping, stage has refined previous annotations of the genome,
characterized some of its previously unannotated parts, and provided novel insights into genome biology. The development in
the mid 1990s of computational techniques to predict protein-coding genes, which span thousands of nucleotides in
eukaryotes, marks the start of the efforts to provide genome-wide annotations. This first stage centered on the annotation of
individual reference genomes (one per species), using a combination of the statistical properties of the sequences of the
protein-coding genes themselves (to predict genes ab initio) and the similarity of mRNA and protein sequences to their genome
of origin (to align known transcripts and proteins). Both approaches though rely on the existence of a set of already char-
acterized genes, a priori knowledge based on experimental evidences, which will be used either as training set to build the
models describing the genes species-specific signals and sequence content biases for subsequent computational predictions, or
to be compared against other anonymous sequences to map their location based on similarity, thus annotating the genes in
them (Fig. 1).
When the reference genomes of more species became available in the early and mid 2000s, they ushered in a comparative
and second stage of genome annotation, where mRNA, protein and genome sequences from one species could be homologously
aligned to annotate the genome of another. A third and regulatory stage of genome annotation was initiated in the mid and late
2000s, after genome-wide methods to profile DNA binding, conformation and alteration (e.g., in their copy number) were
introduced. Such genome-wide and multi-omics regulatory annotations have become standard in the last few years and short
(in the tens or hundreds of nucleotides) regulatory elements are commonly annotated. Finally, genome annotation definitely
moved away from one or few genomes per species in the late 2000s and early 2010s. This fourth and population stage of
genome annotation, which uses hundreds if not thousands of genomes from the same species to annotate individual
nucleotides, is where we find ourselves in the late 2010s. We discuss in some detail these different stages of genome annotation
and how their increased resolution and inclusiveness of multi-omics approaches leads to more precise genome biology.

The ab Initio and Similarity Years

Ab initio gene prediction (also known as de novo) refers to the identification of protein-coding genes using the signals that define
and characterize their gene structures along the genome. This is done without the aid of their encoded mRNAs and proteins. When
the sequences of these mRNAs and proteins are sequenced they can be aligned to the genome to locate the genes that encode them.
Thus, allowing the prediction of protein-coding genes by similarity. We discuss both approaches and the increased accuracy of
similarity approaches over the ab initio ones.

Encyclopedia of Bioinformatics and Computational Biology doi:10.1016/B978-0-12-809633-8.20226-4 1

2 Genome Annotation

Fig. 1 In order to annotate an anonymous sequence (query) we have two main approaches. The first approach, depicted on the left, it is to build
models that, based on known sequences, summarize some property or bias characteristic of the functional elements being annotated, and then
apply an algorithm to score these elements along the sequence, taking the optimal prediction as the one to annotate. The second approach uses
the alignment against the query of the set of known sequences, retrieved either from the same (similarity search) or other species (homology
search). Again, the optimal alignment is taken as the one to annotate the functional element. The coordinates – relative to the original sequence –
define the location for each of the functional elements found in the anonymous genome, also known as the annotation set.

Ab Initio Gene Prediction

The structure of prokaryotic and eukaryotic genes is such that start (AUG) and stop (UAA/UAG/UGA) codons signal the initiation
and termination steps of protein translation. These codons, together with some nucleotide context around them, define the coding
DNA sequence (CDS) of the exon or exons in a gene. In eukaryote genomes, in particular, codons in exons along the CDS are not
necessarily adjacent but often interrupted by introns of hundreds or thousands of nucleotides in length (Sharp, 2005). These
breaks are defined, in most cases, by the canonical donor [AG│GTRAGT, where R¼ A or G] and acceptor [YYTTYYYYYYNCAG│G,
where Y¼ C or T and N ¼ A, C, T or G] splice sites in mammalian genomes (Burset et al., 2000; Abril et al., 2005). Albeit GT-AG and
the surrounding context defines a highly conserved motif, splicing is a complex biological process that tolerates small nuclear RNA
variants in the spliceosome (Kyriakopoulou et al., 2006), the RNA-protein complex that mediates the splicing process, and even
non-canonical splice sites, such as those defined by the donor and acceptor pair AT-AC (Burset et al., 2000; Patel and Steitz, 2003).
Together with transcription initiation and termination sites, these signals have allowed the computational definition of prokar-
yotic and eukaryotic genes.
In addition, the genetic code is degenerate and the same amino acid can be encoded by different codons. These codons are
synonymous and the frequency of their occurrence is known as codon usage or bias. Codon usage varies between species and
between genes in a genome due to mutational biases and differences in the strength of natural selection on translational
optimization (Plotkin and Kudla, 2011). Predating the sequencing of genomes there were tools that used codon bias to annotate
expressed sequence tags (ESTs), fragments of mRNAs, which were even capable to correct frameshifts from sequencing errors like
ESTSCAN (Iseli et al., 1999). Thus, codon usage biases within the signals that define prokaryotic and eukaryotic genes can be used
to computationally identify the CDS of genes along a particular species genome. Some of the first ab initio programs to predict
genes this way did so in the short genomes of prokaryotes. For example, the seminal GENEMARK (Lukashin and Borodovsky, 1998)
and GLIMMER (Delcher et al., 1999) programs made use of Markov chains – a stochastic model describing a sequence of
nucleotide-emitting states in which the probability of the next state only depends on the previous one – to capture the usage and
dependence between successive nucleotides and codon positions, while defining the start and end of genes with models of the
nucleotides around these signals (their sequence context). Interestingly, Markov chains of order-five, in which the probability of
the next nucleotide depends on the five previous ones, work best in gene prediction as they also capture dependencies between
consecutive amino acids in proteins (Durbin et al., 1998). These models are used to score potential genes along the genome. The
final prediction is then obtained using algorithms that find the optimal gene structure, that is, the combination of predicted exons
that maximizes the score of the predicted genes. GENEMARK, for example, uses the Viterbi algorithm (Durbin et al., 1998). This
is a dynamic programming algorithm (Eddy, 2004a) – an optimization technique based on breaking down a problem into smaller
ones and using them recursively to find the optimal solution – which, in a Hidden Markov Model (HMM) (Eddy, 2004b) – a
statistical model following a Markov Chain process with hidden states that either emit nucleotides for coding or noncoding
sequences – finds the best scoring gene structures. More recent developments include Prodigal (Hyatt et al., 2010), which
extends dynamic programming with a special set of rules to cope with overlapping genes while looking for the optimal gene
structure prediction along the genome.
Hyatt et al. also established a reference annotation set to evaluate the accuracy of available prokaryotic gene-finders; a difficult
task due to the lack of an experimentally verified translation start in many genes that changes the average length of genes, as well as
the huge variability in GC content across prokaryota. The accuracy of gene prediction in prokaryotic genomes is generally high,
with more than 95% of true protein coding genes being detected in most species (Hyatt et al., 2010). Though the accuracy
predicting the start of the gene is lower, usually around 80%. As a result, the 50 end (upstream) of prokaryotic genes is less likely to
 Genome Annotation 3

be correct. In addition, prokaryotic gene prediction may still slightly overpredict the number of genes in a genome, with the
resulting false positive genes being often short. It is however possible that some of these genes are indeed real. In any case,
prokaryotic gene prediction has become remarkably accurate in the last two decades.
The identification of eukaryotic genes is, compared to their prokaryotic counterparts, more demanding due to their length (often
in the hundreds of thousands of bases) and their split structure. Still, similar approaches to those in prokaryotes can be used. Early
eukaryotic gene predictors included GENEID (Guigo et al., 1992) and GENSCAN (Burge and Karlin, 1997). Both of these programs
use Markov Chains of order five to assess codon usage and score coding exons as a combination of their coding bias and their start,
end and splice sites signals; either in a rule-based (the GENEID Gene Model; (Guigo et al., 1992)) or a generalized HMM (GHMM)
framework for GENSCAN (Burge and Karlin, 1997). To cope with the combinatorial explosion of putative exons and predict the
optimal gene structure along the genome dynamic programming was used (Burge and Karlin, 1997; Guigo, 1998).
The accuracy of such predictions is lower than in prokaryotic genomes (Table 1). An influential work on the assessment of gene
prediction accuracy in eukaryotes (Burset and Guigo, 1996) established the ground metrics at the nucleotide, exon and gene levels, and
provided one of the first reference gene test sets on this field. To begin with, there is in gene prediction a compromise between
sensitivity (Sn in Table 1), the proportion of true coding nucleotides correctly predicted as such, and specificity (Sp in Table 1), the
proportion of predicted coding nucleotides actually coding, so increasing one often decreases the other. At the nucleotide level this is
not much of an issue, as fragments of most genes are predicted; yet missing the right start and end of an exon reduces the sensitivity and
specificity of gene prediction at the exon level (Table 1). The accuracy for the overall gene is even lower due the difficulty in determining
the right combination of exons along the genome. On the other hand, short genes encoding small proteins (Su et al., 2013) are usually
missed (missing exons, ME, in Table 1) as most genome annotation protocols apply a minimum length threshold, i.e., of 100 amino
acids, to reduce the number of falsely predicted genes (wrong exons, WE, in Table 1). These measures show that classic ab initio gene
prediction programs identified no more than 80% of the true coding nucleotides while predicting as coding a large fraction of
noncoding ones (Table 1). Also, a substantial fraction of coding exons had their boundaries mispredicted. Furthermore, around 40% of
the predicted exons are completely wrong, a rather unreasonable figure. Still, ab initio gene prediction provided at the time a first set of
annotations that could be later refined using comparative and/or experimental approaches (see below). To cope with annotation
changes and multiple transcripts in a gene further accuracy measures have been proposed such as the Annotation Edit Distance, which
quantifies changes to a gene annotation, and the Splice Complexity, which quantifies the complexity of alternative splicing in a gene
(Eilbeck et al., 2009). The terms recall and precision were later adopted to measure gene prediction accuracy. Recall is equivalent to the
sensitivity measure described above, whereas precision is equivalent to specificity. Importantly, specificity or precision in the gene
prediction field avoid the estimation of true negatives from incomplete genomes. Nowadays gene predictions are also compared with
respect to the area under the curve in a receiver operating characteristics plot (known as AUC-ROC plot) (Powers, 2011).
The first large-scale assessment on gene-finding accuracy was performed over 2.4 megabases of the Drosophila melanogaster
genome, which contained the alcohol dehydrogenase gene (Adh), in a collaborative experiment known as the Genome Annotation
Assessment Project (GASP) (Reese et al., 2000). This region was first annotated by human curators, with the raw sequence and a
subset of the reference fly annotations later provided to the participants performing computational gene predictions. Similar
assessments have subsequently been done in the worm (nGASP) (Coghlan et al., 2008) and human genomes (EGASP and RGASP,
see Figs. 2 and 3) (ENCODE Consortium, 2012; Guigo et al., 2006).
Human curators have been invaluable in producing reference annotations for the above assessment experiments and in integrating
the experimental evidence that validates (or refutes) individual gene predictions. For example, D. melanogaster was the first eukaryotic
genome assembled from shotgun sequences (where DNA is broken up randomly) (Adams et al., 2000) and Celera, the company
sequencing it, organized an annotation jamboree with bioinformatics experts, protein specialists, and fruit fly biologists to functionally
annotate it (Pennisi, 2000). Similar community annotation projects have since then become a standard for other genomes, among
those it is worth citing the VErtebrate Genome Annotation database (VEGA, [Link] maintained by the human and
vertebrate analysis and annotation (HAVANA) group at the Wellcome Trust Sanger Institute (Harrow et al., 2012), which also has defined
guidelines to integrate gene annotations (Madupu et al., 2010; Loveland et al., 2012). To assist curators in the visualization and editing
of gene annotations different tools have been created. First, to produce static maps of annotations along a genome sequence with
GFF2PS (Abril and Guigo, 2000), GFF2APLOT (Abril et al., 2003), and more recently CIRCOS (Krzywinski et al., 2009). Later, tools to
manually review gene predictions and integrate experimental evidence, such as APOLLO (Lewis et al., 2002), the Integrative Genomics

Table 1 Accuracy of gene prediction programs on human chromosome 22

Program Nucleotide Exon

SnþSp ME
Sn Sp CC Sn Sp 2 WE

“ab initio” gene finding

GENEID 0.73 0.67 0.70 0.65 0.55 0.60 0.21 0.33
GENSCAN 0.79 0.53 0.64 0.68 0.41 0.55 0.15 0.48
Comparative genomics approach
SGP2 0.75 0.73 0.73 0.66 0.58 0.62 0.19 0.28
TWINSCAN 0.72 0.67 0.69 0.69 0.59 0.64 0.18 0.29
4 Genome Annotation

Fig. 2 The assessment of the accuracy of gene predictions (or other functional features) is usually done in three steps. First, a set of well
characterized genes is split into the training and the testing sequence sets. Second, the training set is used to estimate the species-specific
parameters, such as the frequencies of the different codons in coding exons. Finally, the test set is used to predict genes with the computational
approach under assessment. The sequence coordinates of the predicted genes are then compared against those of the known genes. Accuracy
metrics, calculated at the nucleotide, exon and gene levels, are used to assess accuracy of the method and the reliability of the predictions. This
assessment can be used later on either to estimate overall genome predictions reliability or to iterate through the train/test sets again in order to
improve the software parameters (dashed lines).

Fig. 3 Sensitivity versus specificity scatterplots for the prediction of a number of defining signals in coding exons (CDS), on top row, and full-
length exons, on bottom row, for the submitted predictions to RGASP. The accuracy metrics were computed over the genome sequences from
three species: the human chromosomes three and four, the Drosophila melanogaster (fly) genome and the Caenorhabditis elegans (worm)
genome. Each dot corresponds to a set of predictions made by one of each RGASP participant over one of the species sequences. Note the lower
accuracy of predictions in the larger human genome as well as the difficulty to predict the untranslated part of exons due to missing the correct
transcription start (TSS) and end (TES) sites in the three species.

Viewer (IGV) (Thorvaldsdottir et al., 2013) and the GALAXY server (Giardine et al., 2005) grew in importance. Over time, genome
browsers providing access to databases storing the analyses from genome annotation pipelines have become the standard. These include
GBROWSE (Stein et al., 2002), the UCSC Genome Browser (Kent et al., 2002) and ENSEMBL (Hubbard et al., 2002; Aken et al., 2017),
which are widely used today (Fig. 4).
 Genome Annotation 5

Fig. 4 UCSC Genome Browser view of the human chromosome 17, focused on the sequence annotation around the ACADVL gene. Topmost
annotation tracks show several gene/transcript structures obtained by different methods; mid tracks summarize the conservation score and the
human sequence aligned to different vertebrate genomes; bottom tracks include sequence variants (OMIM and common variants), histone marks
from the ENCODE project and transcript gene expression. The gene annotation tracks on top displays predictions from GENEID, GENSCAN,
SGP2, and AUGUSTUS, which can be compared with the current curated annotations from two reference sets, GENCODE and NCBI-RefSeq.
From the aforementioned gene-finding programs, only AUGUSTUS predicts the three alternative spliced transcripts as it integrates experimental
evidence, although it fails to predict some conserved exons detected by the other programs. Note that downstream DVL2 gene was also detected
by the gene-finders, but the upstream DLG4 gene was missed.

Using mRNAs and Proteins

The sequencing of fragments (ESTs) or full-length mRNAs and proteins from the species whose genome is being annotated has
been the main approach to increase the accuracy of ab initio gene annotation. Instead of relying on the statistical properties of the
CDS and the signals that define them, the direct alignment of transcribed or translated sequences from a genome provides
6 Genome Annotation

experimental evidence to a gene structure. The alignment of such sequences to a naked genome has been computationally
addressed yet again as a dynamic programming algorithm that takes into account the splice sites and exon/intron structure of
eukaryotic genes. In particular, ESTs or full-length transcripts can be properly aligned to the genome using variations of the
Smith-Waterman (Smith and Waterman, 1981; Gotoh, 1982) and Needleman-Wunsch (Needleman and Wunsch, 1970)
algorithms, the classic local and global sequence alignment algorithms, respectively. This is the approach used by the pioneering
EST_GENOME program (Mott, 1997). For the sake of speed, other approaches use the heuristic BLAST algorithm (Altschul et al.,
1990) to produce the starting alignments of coding exons in the genome, which are later recursively chained together to obtain the
best gene structure. SIM4 (Florea et al., 1998) and SPIDEY (Wheelan et al., 2001) were examples of the latter. A more recent tool
to align transcripts to a genomic sequence is EXONERATE (Slater and Birney, 2005), which provides (slower) exhaustive or (faster)
heuristic approaches to align transcripts using different dynamic programming algorithms. On the other hand, spliced-alignment
algorithms with proteins require the conceptual translation of the genomic sequence and finding the optimal alignment of a multi-
exon structure to a related protein. PROCRUSTES (Gelfand et al., 1996) and GENEWISE (Birney and Durbin, 1997, 2000) are well-
known examples, with the HMM-based GENEWISE at the core of the ENSEMBL gene annotation system (Curwen et al., 2004; Aken
et al., 2016). The alignment of proteins to predict genes in large scale genomes can become extremely time- and space-demanding.
This investment, however, is usually at the benefit of prediction accuracy. In this regard, the faster EXONERATE can also align
proteins to DNA sequences and is also heavily used in the ENSEMBL pipeline.
In the last few years, the widespread use of next-generation sequencing technologies for genomes and transcriptomes (RNA
sequencing, RNA-seq) and high-throughput mass spectrometry approaches for proteins have put the approaches above at the forefront
of genome annotation. For example, the alignment of transcripts from RNA-seq experiments has allowed the comprehensive annotation
of alternative splice forms of protein-coding genes. Humans, for example, produce on average four different protein-coding sequences
per gene (ENCODE Consortium, 2012). Pseudogenes, which derived from functional genes through retrotransposition or duplication
but have lost the original functions of their parental genes, are sometimes (about 10%) also transcribed (ENCODE Consortium, 2012).
The high degree of conservation and similarity to functional genes of recent pseudogenes are significant enough to confound con-
ventional gene prediction approaches (Zheng et al., 2007). Other non-standard types of genes that have benefited from experimental
evidence are fused genes, which combine part or all of the exons from transcripts of two collinear genes. In this way, they may
contribute to the increase the protein repertoire in eukaryotes (Parra et al., 2006). Similarly, trans-splicing, where the starting exons from
one transcript are merged with exons from another transcript, far downstream or even located in another chromosome sequence, can be
annotated using sequenced transcripts. Remarkably, up to 70% of the worm Caenorhabditis elegans mRNAs begin with the spliced leader
sequence (SL, 22bp), which is not associated with the gene (Hastings, 2005).

Annotation of Non-Standard Protein Coding Genes

The identification of eukaryotic selenoprotein genes has challenged most computational gene prediction methods. The reason for
this is the alternative use of the UGA codon, typically a termination signal, to code for selenocysteine (Chambers et al., 1986;
Zinoni et al., 1986). The recoding of UGA confounds the computational identification of selenoprotein genes, whose exons
containing a selenocysteine codon are truncated or skipped in their prediction. Selenocysteine, the 21st amino acid in the genetic
code, is incorporated into selenoprotein with the help of an mRNA element, the selenocysteine insertion sequence (SECIS),
located in the 30 untranslated region of selenoprotein genes (48,49). This RNA structure directs the incorporation of selenocysteine
through the interaction with several trans-acting factors. Selenocysteine mediates the biological functions of selenium, an
essential micronutrient whose deficiency is associated to infertility, preterm birth and serious children’s disorders of the bone and
heart (Rayman, 2000).
Efforts to identify selenoproteins were first directed to the computational prediction of the SECIS element in Expressed
Sequence Tags (ESTs, sequenced mRNA fragments). Known SECIS elements were used to derive deterministic models of their
secondary structures (Fig. 5) which, in turn, were used to scan and identify two novel selenoprotein genes (Kryukov et al., 1999;
Lescure et al., 1999). Efforts to identify additional selenoprotein genes in eukaryotic genomes soon followed. The inclusion of
UGA as a selenocysteine codon in the prediction of optimal gene structures along the genome, however, led to the prediction of
many erroneous selenoprotein genes (false positives). The low specificity of selenoprotein gene prediction was due to the difficulty
of using the typical codon usage characteristic of proteins to extend them beyond an in-frame termination codon (Castellano et al.,
2001). A solution to this problem was to use the prediction of SECIS elements along the genome to limit the number of exons
with a TGA in-frame that are incorporated into the dynamic programming recursion used to predict optimal genes. Several new
selenoproteins in different species were identified this way (Castellano et al., 2001; Kryukov et al., 2003; Castellano et al., 2005).
Another example of stop codon recoding is pyrrolysine, which is encoded by UAG but may not require a specific stem loop
structure in the 30 untranslated region for translation (Zhang et al., 2005).

The Comparative and Homology Years

The Zoo Blot is an experimental technique which uses Southern blot analysis to test the ability of a nucleic acid probe from one
species to hybridize with DNA of a range of species. Zoo Blots were already used in the 1980s to characterize the protein coding
 Genome Annotation 7

Fig. 5 (a) Schematic SECIS element (form 1) divided into structural units. Mostly invariant nucleotides are indicated. (b) Alignment of human and
other mammalian glutathione peroxidase 1 SECIS sequences. Note the strong conservation among orthologous sequences. (c) Alignment of human
and other non-mammalian glutathione peroxidase 1 SECIS sequences. Compared to the previous alignment, sequence conservation drops
significantly.

regions of newly discovered genes (Monaco et al., 1986; Katzav et al., 1989). The rationale behind this approach was that protein
coding regions will be conserved across species and would thus retain the capability to hybridize. Testing for hybridization of
genomic probes to genomic libraries from many species will reveal the fragments likely to correspond to coding exons. As often
happens, bioinformatics techniques based on sequence similarity replicate in the computer those experimental techniques based on
the hybridization properties of nucleic acid molecules. In this regard, comparative gene prediction methods that rely on the fact that
protein coding regions are generally more conserved during evolution that non-functional ones (Fig. 6), can be though as a sort of
electronic zoo blots. While phylogenetic footprinting – the phenomenon of the conservation of functional regions across species –
has been used to discover, identify and characterize functional regions other than protein coding genes (Margulies et al., 2003).
Essentially, there are two main classes of comparative approaches for gene identification: the comparison of the DNA query
sequence with a target protein or mRNA sequence, or a database of such sequences (the computational equivalent of a Northern
zoo blot), and the comparison of two or more genomic sequences. In both approaches query and target sequences may be from
the same (as discussed above) or different species. If the latter, the selection of the target organism is not a negligible step; the
organism should be chosen, when possible, at the phylogenetic distance maximizing the correlation between sequence con-
servation and coding function. Thus, allowing the inference of homology.

Using mRNAs and Proteins

The backbone of similarity-aided or similarity-based gene structure determination is constituted by those methods that rely on a
comparison of the query sequence with protein, or mRNA sequences. Although mostly known as a database search program,
BLASTX (Altschul et al., 1990; Gish and States, 1993) illustrates the rationale behind this approach. With BLASTX a genomic
query is translated into a set of amino acid sequences in the six possible frames and compared against a database of known protein
sequences. The assumption is that those segments in the genomic query similar to database proteins are likely to correspond to
homologous coding exons. A similar assumption is behind the comparison of the genomic query against a database of mRNA
sequences (such as ESTs), using BLASTN (Altschul et al., 1990), FASTA (Pearson and Lipman, 1988; Pearson, 2000), or similar
programs. The National Center for Biotechnology Information GENBANK (Benson et al., 2013) and UNIPROT (Bateman et al.,
2017), and their European and Asian counterparts, have been the main sources of nucleotide and protein sequences, respectively,
for homology-based searches.
8 Genome Annotation

Fig. 6 Comparison of the orthologous genomic region of the acyl-Coenzyme A dehydrogenase very long chain (ACADVL) in human, mouse, and
rat. The comparison was done at the amino acid level using the ungapped TBLASTX. Coding exons are shown as red boxes and non-coding ones
(untranslated regions) are depicted in grey. Exons are projected to highlight their conservation (orange/grey ribbons). Note that introns are also
conserved, albeit to a lesser extent than coding exons between human and any of the rodent genomes. The plot was obtained with GFF2APLOT.

Database search programs, however, are not dedicated gene prediction tools; they are not capable of automatically identifying
start and stop codons or splice sites, the signals defining the exonic structure of the genes. Thus, after a database search and the
identification of potential targets of homology, additional tools are required to define these structures. One approach is to use the
top database match as target sequence and obtain a so-called spliced alignment between this and the genomic query. In such an
alignment, large gaps – likely to correspond to introns – are only allowed at legal splice junctions. To calculate such spliced
alignments between transcript and genomic sequences several programs we have already mentioned have been used: SIM4 (Florea
et al., 1998), EST_GENOME (Mott, 1997), SPIDEY (Wheelan et al., 2001), and EXONERATE (Slater and Birney, 2005). Regarding
the alignment of proteins of one species to another species genome, GENEWISE (Birney and Durbin, 1997, 2000) and
EXONERATE (Slater and Birney, 2005) are used by ENSEMBL. In this direction, Meyer and Durbin (Meyer and Durbin, 2004) have
developed the program PROJECTOR, which makes explicit use of the conservation of the exon-intron structure between two
related genes. Because human and mouse orthologous genes show a remarkable conservation of their exonic structure, (Mouse
Genome Sequencing Consortium, 2002), the PROJECTOR program outperforms GENEWISE predictions when human targets are
used in mouse genomic sequences, and vice versa, in particular, when the conservation at the amino level is weak.
In an alternative approach, the results of a database search can be integrated ad-hoc into the framework of a typical ab initio gene
prediction program. In essence, these methods promote candidate exons in the query sequence for which similar known coding
sequences exist. Indeed, the score of the candidate exon – initially a function of the score of the splice (start, stop) sites and of the
coding potential of the exon sequence – is increased as a function of the similarity between the candidate exon and the known
coding sequences. In this way, candidate exons showing similarity to known coding sequences and thus likely orthologous, are
promoted into the final gene prediction. In theory, this approach should produce predictions as accurate as pure ab initio programs
when no similar target sequences exist, but more accurate ones (ideally, as accurate as those from splicing alignment tools) when
such target sequences do exist. One example of this approach is the program GENOMESCAN (Yeh et al., 2001), an extension of
GENSCAN (Burge and Karlin, 1997) over which it reports increased accuracy. GRAILEXP (Xu et al., 1997), CRASA (Chuang et al.,
2003) and AUGUSTUS (Stanke et al., 2006a) are examples of methods using ESTs or mRNAs instead.

Using Whole-Genomes
With the increasing availability of the genome sequences for a wide range of eukaryotic organisms, whole genome sequence
comparisons gained popularity as a mean of identifying protein coding genes. Under the assumption that regions conserved in
genome sequence alignments will tend to correspond to coding exons from homologous genes, a number of programs were
developed. The program EXOFISH (Crollius et al., 2000) was one of the first such programs. Basically, it predicted human exons
aided by the comparison of the human genome with sequences from Tetraodon nigroviridis, a puffer fish species (within Acti-
nopterygii) that diverged about 400 Myr ago from the lineage later leading to humans. Latter developments followed notably
different approaches.
In one such approach (Pedersen and Scharling, 2002; Blayo et al., 2003), the problem was stated as a generalization of pairwise
sequence alignment: given two genomic sequences coding for homologous genes, the goal was to obtain the predicted exonic
structure in each sequence maximizing the score of the alignment of the resulting amino acid sequences. Both, (Blayo et al., 2003)
 Genome Annotation 9

and Pedersen and Scharling (Pedersen and Scharling, 2002) solved this problem through a complex extension of the classical
dynamic programming algorithm for sequence alignment.
In a different approach, the programs SLAM (Alexandersson et al., 2003) and DOUBLESCAN (Meyer and Durbin, 2002)
combined sequence alignment pair HMMs (Durbin et al., 1998), with gene prediction generalized HMMs (Burge and Karlin, 1997)
into the so-called generalized pair HMMs. In these, gene prediction is not the result of the sequence alignment, as in the programs
above, but both gene prediction and sequence alignment are obtained simultaneously.
A third class of programs adopted a more heuristic approach, and clearly separated gene prediction from sequence alignment.
The programs ROSETTA (Batzoglou et al., 2000), SGP1 (from Syntenic Gene Prediction, (Wiehe et al., 2001)), and CEM (from the
Conserved Exon Method, (Bafna and Huson, 2000)) are representative of this approach. All these programs start by aligning two
syntenic sequences – understood here as sequences from different species, but containing homologous genes in a similar order–,
and then predict gene structures in which the exons are compatible with the alignment.
Although similarity-based gene prediction with homologous genomic sequences produced high quality results (Miller, 2001),
an obvious shortcoming was the need for two homologous sequences. Also, genes without a homologue in the partner sequence
will escape detection. This is particularly problematic if species are compared at genomic regions where synteny is not preserved.
Given only a single query sequence, it is therefore desirable to automatically search for homologous sequences or syntenic
chromosome stretches in other species that are suited to similarity-based approaches. The programs TWINSCAN (Korf et al., 2001)
and SGP2 (Parra et al., 2003) attempted to address this limitation. The approach in these programs was reminiscent of that used in
GENOMESCAN (Yeh et al., 2001). Essentially, the sequence from the query genome is compared against a collection of sequences
from the target or informant genome – which can be a single homologous sequence to the query sequence, a whole assembled
genome, or a collection of shotgun reads –, and the results of the comparison are used to modify accordingly the scores of the
exons produced by ab initio gene predictors. In TWINSCAN, the query and target genome sequences are compared at the nucleotide
level with BLASTN and the results serve to modify the underlying probability of the potential exons predicted by GENSCAN. In
SGP2, the genome sequences are compared using TBLASTX (Altschul et al., 1990), and the results used to modify the scores of
the potential exons predicted by GENEID.
TWINSCAN, SGP2, and SLAM were successfully applied to the annotation of the mouse genome (Mouse Genome Sequencing,
C, 2002), and helped to identify previously unconfirmed genes (Guigo et al., 2003). Importantly, a number of studies consistently
indicate that comparative gene finders improve over their ab initio counterparts (Parra et al., 2003; Wu et al., 2004). Indeed, in
Table 1, we show that, when evaluated in human chromosome 22, TWINSCAN and SGP2 outperformed GENSCAN and GENEID,
respectively. The exhaustive scrutiny to which the sequence of human chromosome 22 (Dunham et al., 1999) was subjected
through the Vertebrate Genome Annotation database project at the Wellcome Sanger Institute offers an excellent platform to
obtain estimates of the accuracy of current gene finders.
Annotating a genome has required a complex craftsmanship from its inception, given the many different layers of annotations
that have to be considered and therefore put together. Common to the tools and approaches described above is the first task of
masking the repetitive sequences, often taxa-specific, that populate the genome; this step facilitates and speeds up the downstream
analyses just by replacing the segments where the repeats are found by N’s, a common nucleotide wildcard – meaning A or C or G
or T–. RepeatMasker (Smit et al., 2013-2015), has been the classical tool for this purpose, yet it has been described as sensitive
but computationally costly when applied to large eukaryotic genomes (Bedell et al., 2000). After repeat-masking, genome
sequences are scanned for protein-coding and non-coding genes using different prediction tools while, simultaneously, homology-
based searches and whole-genome alignments to other species are performed. The result are sets of annotations with varying
degrees of supporting evidence that need to be integrated into one set of annotated genes, also known as the gene-build of a
genome. Additional genome annotation layers result from high-throughput experimental approaches, such as RNA-seq and the
transcriptional evidence it produces (which may be used on the gene-build itself), or the typing and resequencing experiments that
sample sequence variation across populations (human or otherwise), and so on. Genes from the gene-build, and the proteins they
encode, require these other annotation layers to assign them a molecular function, a process known as functional annotation.
These functions rely on controlled vocabularies to describe them, for instance the Gene Ontology (GO) (Ashburner et al., 2000;
The Gene Ontology Consortium, 2017).
As the complexity of the annotation process increased due to the need to integrate new layers of information, guidelines and/
or automatic or semiautomatic workflows to annotate novel genomes or to re-annotate existing ones (every time the
assemblies are improved) were developed (Mudge and Harrow, 2016; Elsik et al., 2014). For example, the NCBI prokaryotic
(Tatusova et al., 2016) and eukaryotic (Thibaud-Nissen et al., 2016) pipelines, or EnsEMBL (Potter et al., 2004), which
rely on in-house protocols for large-scale genome annotation. In addition, gene-finding tools were wrapped into specia-
lized genome annotation workflows. For instance, the Fgenesh (Salamov and Solovyev, 2000), SNAP þ Exonerate, and
GeneMark þ AUGUSTUS were integrated into the Fgenesh þ þ (Solovyev et al., 2006), MAKER (Cantarel et al., 2008), and
BRAKER (Hoff et al., 2016) annotation pipelines, respectively.
Comparative genomics can be also useful to assess the completeness of genome annotations. REFSEQ (Pruitt et al., 2012) has
been used as a reference annotation set for the annotation of novel gene-builds, but also to validate the completeness of genome
assemblies. A set of 458 highly reliable core proteins, conserved across plants, fungi and mammals, was collected and integrated
into a gene-finding validation pipeline known as CEGMA (Parra et al., 2007). More recently, it has been superseded by a set of
single-copy orthologs in the BUSCO pipeline (Simao et al., 2015), which can be used to benchmark the completeness of both the
annotation of genes and the genome assembly.
10 Genome Annotation

Annotation on Non-Standard Protein-Coding Genes

Comparative methods for gene prediction have been also useful to annotate selenoprotein genes. The fact that downstream of a
recoded stop codon lies a real protein implies this region to have the typical pattern of sequence conservation among its protein
homologues. At the least, to the extent of the conservation in the upstream protein sequence. Therefore, sequence conservation
beyond a TGA codon in the genome strongly suggests selenocysteine coding function. It is then feasible to predict selenoprotein
genes in two different species genomes, without information from their SECIS elements, and compare them to identify a pattern of
symmetrical protein sequence conservation around a predicted selenocysteine codon. Application of this approach to the human
and fugu genomes identified a new selenoprotein in this fish (Castellano et al., 2004). More recently, selenoproteins have been
annotated across eukaryotes, in a SECIS-independent manner, using alignment profiles derived from the known selenoprotein
families (Mariotti and Guigo, 2010).

The RNA and Regulatory Years

The functional annotation of the regulatory and non-coding part of the genome has lagged behind the annotation of protein-
coding genes. Systematic analysis of RNA genes and regulatory regions was made possible in the mid and late 2000s by a number
of experimental approaches for the characterization of transcribed regions, transcription factor binding sites, DNA methylation
sites and chromatin structure. The ENCODE project was an early adopter of these methods to systematically characterize the
human genome, (2012ENCODE Project Consortium). Among these, tiling arrays and RNA-seq have been important to under-
stand the transcriptional landscape of the human genome, which is pervasively transcribed (Djebali et al., 2012; Bertone et al.,
2004). This was perhaps not unexpected if one acknowledged the high initiation promiscuity of Pol II, the eukaryotic RNA
polymerase complex (Struhl, 2007). In any case, transcription from regions where protein-coding genes or known RNA genes –
such as ribosomal and transfer RNAs, small nuclear RNAs involved in splicing, microRNAs, short interfering RNAs, and a few
others – have so far not been annotated, produced thousands of long noncoding RNAs with no obvious function. These unknown
RNAs (Derrien et al., 2012), which standard gene prediction programs cannot detect using codon bias measures, constitute a vast
array of potential genes that do not encode for proteins (Kowalczyk et al., 2012), albeit some may do (Ji et al., 2015). While a few
of these new RNAs have been shown to have functions (Nesterova et al., 2001; Sleutels et al., 2002), there is little functional
evidence for the majority of them. Their transcription is often extremely low and they tend to have limited sequence or RNA
structure conservation (Rivas et al., 2017), raising the possibility that they are transcriptional noise. Indeed, between and within
species analyses indicate that most of these RNA genes are not under strong evolutionary constraint (purifying selection) (Wiberg
et al., 2015). Alternatively, it could be their transcription itself rather than their sequence that is under selection. The functional
annotation of RNA genes thus remains open and additional work is needed before this controversy is solved.
The high-throughput sequencing of transcriptomes and other experimental approaches has required the development of fast
aligners like BWA (Li and Durbin, 2009) and Bowtie (Langmead and Salzberg, 2012), which can align millions of short sequences
to a reference genome sequence. Specialized tools can integrate those alignments into predicted gene structures, such as
AUGUSTUS (Stanke et al., 2006b), mGENE (Schweikert et al., 2009), or the Transomics pipeline (Solovyev et al., 2006). Other
tools, like SpliceGrapher (Rogers et al., 2012), have been developed to refine the alignments of mapped sequences to predict
“splicing graphs”, a representation of the splicing variants of a gene in which exons are shown as nodes that are connected by the
intervening introns represented as edges. As discussed, obtaining simultaneously the genome and transcriptome of a species
facilitates the task of gene annotation, by aligning the transcriptome sequences along the newly assembled genome, and provides
an estimation of gene expression levels when translating the alignment coverage of read sequences into normalized counts
(Mortazavi et al., 2008). An extensive assessment of the sequence mappers (Engstrom et al., 2013) and the gene-finders (Steijger
et al., 2013) was undertaken by the GENCODE consortium (Harrow et al., 2012). Among the results obtained it is worth
mentioning that 50 - and 30 -UTR exons were still difficult to predict (see Fig. 3) as well as the different exon combinations for
splicing transcript isoforms, and that prediction accuracy correlated with the gene expression level, as highly expressed genes are
often more abundant in the training sets. However, this is changing as single molecule sequencing methods, such as PacBio or
ONT nanopore, produce long sequences recovering the full length splicing isoforms of a gene (Sharon et al., 2013). Thus easing the
task of protein-coding and non-coding mRNA genes annotation, even at single cell resolution (Byrne et al., 2017).
Regulatory regions where proteins (e.g., transcription factors) bind have been studied using chromatin immunoprecipitation
followed by sequencing (ChIP-seq). Around 8% of the genome across different cell types is typically protein-bound with sequence-
specific signals known to bind proteins being common in them. In addition, DNase I hypersensitivity has been used to identify
those regions of the genome open to transcription. These regions can be bound by transcription factors and were found to be
upstream of the predicted transcription start sites, coinciding with the regions defined by the ChIP-seq experiments. Methylation of
cytosine, usually at CpG dinucleotides, is involved in the epigenetic regulation of gene expression, with promoter and gene
methylation leading to repression and promotion of transcription, respectively. Levels of DNA methylation correlate with chro-
matin accessibility as defined by the DNase I hypersensitivity experiments. Interestingly, most variable regions in their methylation
correspond to genes rather than their promoters.
One surprising result from the ENCODE project was the finding that the majority (about 80%) of the human genome
participated in one or more of the experiments above (ENCODE Project Consortium, 2012), mostly in the RNA transcription
 Genome Annotation 11

experiments discussed above. This was initially interpreted as providing evidence for a functional role to most of the human DNA
sequence but low specificity of the assays used is a simpler explanation. These assays often do not distinguish between meaningful
function – some molecular function that impacts fitness in the individual – and one that does not. As a result, having the majority
of the genome performing important roles contrasts with the small fraction of it – probably 5%–10% – that has been shown to be
under evolutionary constraint (Asthana et al., 2007; Davydov et al., 2010). It follows that some fraction of non-coding DNA is
functional in the typical meaning of this word but that the majority of it – mostly transposable elements that make up about half
of the human genome (Eddy, 2012) – is not (albeit being biochemically active in different experiments). Indeed, one of the
common tasks in gene prediction is masking about half of the human genome for repeats and transposons using the venerable
REPEATMASKER program.

The Population Years

The functional annotation of a species genome provides a reference to which additional genomes from the same species can be
compared. Thousands of human exomes (protein-coding fraction of the genome) or genomes have now been sequenced (1k
Genomes Project Consortium, 2012, 2015) and, in doing so, many differences have been found between them and their reference
– e.g., between four and five million nucleotide differences in each human genome –. These differences, which include single
nucleotide variants, insertions and deletions, copy number variants and others, are often novel and unannotated, which each
ensuing genome having more of them. These rare variants in humans, which are found in specific populations and at low
frequency, are to a large the result of the recent growth of human populations (Keinan and Clark, 2012). Indeed, humans have
accumulated in the last 5000–10,000 years an enormous number of rare variants, particularly in non-African populations. Protein-
coding variants at low frequency are often slightly deleterious as measured by approaches that infer the probability of amino acid
changes impacting the structure or function of proteins (Fu et al., 2013). Programs like POLYPHEN (Adzhubei et al., 2010) and
SIFT (Kumar et al., 2009), using amino acid conservation (within and between species) and different amino acid properties,
annotate their deleteriousness across the genome. In addition, a few hundred variants disrupt the translation of proteins in a
typical human genome. Non-coding variants in regulatory regions have also been annotated with regard to their deleteriousness
using, for example, their conservation among species with PHASTCONS (Siepel et al., 2005). It follows that the functional
annotation of single nucleotide variants, particularly those that change (amino acid substitutions) or disrupt proteins (e.g.,
changing splice sites or creating stop codons), are important to understand interindividual differences linked to diseases that are
both rare and severe. Indeed, rare diseases, understood as those that occur in one out of 2000 individuals, have often a genetic
basis and the annotation of the deleteriousness of the causal mutation(s) in single-gene disorders is the first step in their diagnosis
and treatment.
Still, most of the genetic variation between individuals genomes of a species is shared among some (or most) of them. This
common variation was described in a landmark study that provided not only a catalog of single nucleotide variants in humans but
the genetic association among them (International HapMap Consortium, 2005). This association, known as linkage dis-
equilibrium, reflects the coinheritance of sets of single nucleotide variants (haplotypes) and their functional annotation is needed
to investigate, using genome-wide association studies, the hereditary factors involved in disease. This is important as each human
genome has around 10,000 rare and common amino acid changes, a few hundred protein-disrupting ones and about 500,000
non-coding changes in regulatory regions (1k Genomes Project Consortium, 2015). Of these, about 2000 are linked through
genome-wide studies with common disease whereas a few dozen are implicated with rare disease per genome. Today, thousands of
mutations have been associated to common or rare disease in humans and their genome and functional annotation can be
obtained from different databases, for example OMIM ([Link] CLINVAR (Landrum et al., 2016) and the NHGRI-EBI
GWAS catalog (MacArthur et al., 2017).

Conclusion

Genome annotation has come a long way since the first protein-coding genes were annotated using ab initio and later comparative
approaches. Experimental evidence is now routinely integrated in the annotation of these genes and their multiple alternative
splice forms. At the same time, promoters and smaller regulatory elements, as well as RNA genes (including controversial ones),
are being annotated with a variety of experimental approaches genome-wide. Knowing the location of these coding and noncoding
functional features has provided the framework to understand the role of sequence variants in health and disease, which remains
largely undetermined. Still, since the early nineties, when the Human Genome Project was launched, our knowledge of genome
biology has greatly improved. This is not only due to the ever increasing computational capabilities but also to novel discoveries in
molecular biology from experimental methods with increased resolution and throughput, which has been harnessed in annotation
consortiums like the ENCODE project. As a result, well annotated genomes like that of human, fly or mice, the result of years of
intense human curation, as well as gold standards such REFSEQ or VEGA genes, provide reference annotations to other species.
This is important as the pace of sequencing new species genomes, transcriptomes, and environmental (metagenomics) samples
is only accelerating given the ever decreasing cost of sequencing technologies. Thus, a shift towards more automated annotation
pipelines is warranted to cope with the scale of the analyses in genome annotation. Automated annotation pipelines have room to
12 Genome Annotation

improve though, with gene-builds still having a significant impact in the downstream analyses. For example, on the variants found
in a genome and their predicted functional consequences (Frankish et al., 2015). This is due to the fact that most gene-finding tools
return a single and often different set of optimal gene structures. This also poses a problem for overlapping genes, including those
on different strands, nested genes, genes located within introns of other genes, non-standard gene structures, and genes with
translation recoding like selenoprotein genes, as they are often not included in the optimal prediction of genes. Furthermore, most
gene prediction tools do not consider frame shifts, RNA editing, and the impact of sequencing errors in the identification of gene
structures. Still, some tools are capable to infer more than one alternative splicing isoform provided that there is some sort of
experimental evidence, for example, from RNA-seq experiments. However, problems may appear when transcripts contain start or
stop codons split by introns. Together with non-cannonical splice-sites, these are some of the special cases that human curators
solve in the final refinement steps of genome annotation when all sort of evidences are integrated. Many of aforementioned issues
will be overcome by the use of single molecule sequencing technologies, which make possible to map the full length of individual
protein-coding and non-coding transcript isoforms. In combination with single cell gene expression experiments, it will be finally
feasible to produce gene-builds and their corresponding functional annotation at the single molecule and cell resolution, an
astounding promise just two decades ago.

Acknowledgements

This study was partially funded (Sergi Castellano) by the NIHR HS&DR Programme (14/21/45) and supported by the NIHR
GOSH BRC. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of
Health.

References

Abril, J.F., Castelo, R., Guigo, R., 2005. Comparison of splice sites in mammals and chicken. Genome Res. 15 (1), 111–119.
Abril, J.F., Guigo, R., 2000. gff2ps: Visualizing genomic annotations. Bioinformatics 16 (8), 743–744.
Abril, J.F., Guigo, R., Wiehe, T., 2003. gff2aplot: Plotting sequence comparisons. Bioinformatics 19 (18), 2477–2479.
Adams, M.D., et al., 2000. The genome sequence of Drosophila melanogaster. Science 287 (5461), 2185–2195.
Adzhubei, I.A., et al., 2010. A method and server for predicting damaging missense mutations. Nat. Methods 7 (4), 248–249.
Aken, B.L., et al., 2016. The Ensembl gene annotation system. Database (Oxford) 2016.
Aken, B.L., et al., 2017. Ensembl 2017. Nucleic Acids Res. 45 (D1), D635–D642.
Alexandersson, M., Cawley, S., Pachter, L., 2003. SLAM: Cross-species gene finding and alignment with a generalized pair hidden Markov model. Genome Res. 13 (3), 496–502.
Altschul, S.F., et al., 1990. Basic local alignment search tool. J. Mol. Biol. 215 (3), 403–410.
Ashburner, M., et al., 2000. Gene ontology: Tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet. 25 (1), 25–29.
Asthana, S., et al., 2007. Widely distributed noncoding purifying selection in the human genome. Proc. Natl. Acad. Sci. USA 104 (30), 12410–12415.
Bafna, V., Huson, D.H., 2000. The conserved exon method for gene finding. Proc. Int. Conf. Intell. Syst. Mol. Biol. 8, 3–12.
Bateman, A., et al., 2017. UniProt: The universal protein knowledgebase. Nucleic Acids Res. 45 (D1), D158–D169.
Batzoglou, S., et al., 2000. Human and mouse gene structure: Comparative analysis and application to exon prediction. Genome Res. 10 (7), 950–958.
Bedell, J.A., Korf, I., Gish, W., 2000. MaskerAid: A performance enhancement to RepeatMasker. Bioinformatics 16 (11), 1040–1041.
Benson, D.A., et al., 2013. GenBank. Nucleic Acids Res. 41 (D1), D36–D42.
Bertone, P., et al., 2004. Global identification of human transcribed sequences with genome tiling arrays. Science 306 (5705), 2242–2246.
Birney, E., Durbin, R., 1997. Dynamite: A flexible code generating language for dynamic programming methods used in sequence comparison. In: Proceedings of Ismb-97 –
Fifth International Conference on Intelligent Systems for Molecular Biology, pp. 56–64.
Birney, E., Durbin, R., 2000. Using GeneWise in the Drosophila annotation experiment. Genome Res. 10 (4), 547–548.
Blayo, P., Rouze, P., Sagot, M.F., 2003. Orphan gene finding – An exon assembly approach. Theor. Comput. Sci. 290 (3), 1407–1431.
Burge, C., Karlin, S., 1997. Prediction of complete gene structures in human genomic DNA. J. Mol. Biol. 268 (1), 78–94.
Burset, M., Guigo, R., 1996. Evaluation of gene structure prediction programs. Genomics 34 (3), 353–367.
Burset, M., Seledtsov, I.A., Solovyev, V.V., 2000. Analysis of canonical and non-canonical splice sites in mammalian genomes. Nucleic Acids Res. 28 (21), 4364–4375.
Byrne, A., et al., 2017. Nanopore long-read RNAseq reveals widespread transcriptional variation among the surface receptors of individual B cells. Nat. Commun. 8, 16027.
Cantarel, B.L., et al., 2008. MAKER: An easy-to-use annotation pipeline designed for emerging model organism genomes. Genome Res. 18 (1), 188–196.
Castellano, S., et al., 2001. In silico identification of novel selenoproteins in the Drosophila melanogaster genome. EMBO Rep. 2 (8), 697–702.
Castellano, S., et al., 2004. Reconsidering the evolution of eukaryotic selenoproteins: A novel nonmammalian family with scattered phylogenetic distribution. EMBO Rep. 5 (1), 71–77.
Castellano, S., et al., 2005. Diversity and functional plasticity of eukaryotic selenoproteins: Identification and characterization of the SelJ family. Proc. Natl. Acad. Sci. USA 102
(45), 16188–16193.
Chambers, I., et al., 1986. The structure of the mouse glutathione peroxidase gene: The selenocysteine in the active site is encoded by the ‘termination’ codon, TGA. EMBO J.
5 (6), 1221–1227.
Chuang, T.J., et al., 2003. A complexity reduction algorithm for analysis and annotation of large genomic sequences. Genome Res. 13 (2), 313–322.
Coghlan, A., et al., 2008. nGASP – The nematode genome annotation assessment project. BMC Bioinform. 9, 549.
[Link], Project Consortium, 2012. An integrated encyclopedia of DNA elements in the human genome. Nature 489 (7414), 57–74.
Crollius, H.R., et al., 2000. Estimate of human gene number provided by genome-wide analysis using Tetraodon nigroviridis DNA sequence. Nat. Genet. 25 (2), 235–238.
Curwen, V., et al., 2004. The Ensembl automatic gene annotation system. Genome Res. 14 (5), 942–950.
Davydov, E.V., et al., 2010. Identifying a high fraction of the human genome to be under selective constraint using GERP plus. PLOS Comput. Biol. 6 (12),
Delcher, A.L., et al., 1999. Improved microbial gene identification with GLIMMER. Nucleic Acids Res. 27 (23), 4636–4641.
Derrien, T., et al., 2012. The GENCODE v7 catalog of human long noncoding RNAs: Analysis of their gene structure, evolution, and expression. Genome Res. 22 (9), 1775–1789.
Djebali, S., et al., 2012. Landscape of transcription in human cells. Nature 489 (7414), 101–108.
Dunham, I., et al., 1999. The DNA sequence of human chromosome 22. Nature 402 (6761), 489–495.
 Genome Annotation 13

Durbin, R., et al., 1998. Biological Sequence Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge, United Kingdom: Cambridge University Press, (xi, 356
pages: Illustrations; 26 cm).
Eddy, S.R., 2004a. What is dynamic programming? Nat. Biotechnol. 22 (7), 909–910.
Eddy, S.R., 2004b. What is a hidden Markov model? Nat. Biotechnol. 22 (10), 1315–1316.
Eddy, S.R., 2012. The C-value paradox, junk DNA and ENCODE. Curr. Biol. 22 (21), R898–R899.
Eilbeck, K., et al., 2009. Quantitative measures for the management and comparison of annotated genomes. BMC Bioinform. 10, 67.
Elsik, C.G., et al., 2014. Finding the missing honey bee genes: Lessons learned from a genome upgrade. BMC Genomics 15, 86.
Engstrom, P.G., et al., 2013. Systematic evaluation of spliced alignment programs for RNA-seq data. Nat. Methods 10 (12), 1185–1191.
Florea, L., et al., 1998. A computer program for aligning a cDNA sequence with a genomic DNA sequence. Genome Res. 8 (9), 967–974.
Frankish, A., et al., 2015. Comparison of GENCODE and RefSeq gene annotation and the impact of reference geneset on variant effect prediction. BMC Genom. 16 (Suppl. 8), S2.
Fu, W.Q., et al., 2013. Analysis of 6515 exomes reveals the recent origin of most human protein-coding variants. Nature 493 (7431), 216–220.
Gelfand, M.S., Mironov, A.A., Pevzner, P.A., 1996. Gene recognition via spliced sequence alignment. Proc. Natl. Acad. Sci. USA 93 (17), 9061–9066.
1k Genomes Project Consortium, et al., 2012. An integrated map of genetic variation from 1092 human genomes. Nature 491 (7422), 56–65.
1k Genomes Project Consortium, et al., 2015. A global reference for human genetic variation. Nature 526 (7571), 68–74.
Giardine, B., et al., 2005. Galaxy: A platform for interactive large-scale genome analysis. Genome Res. 15 (10), 1451–1455.
Gish, W., States, D.J., 1993. Identification of protein coding regions by database similarity search. Nat. Genet. 3 (3), 266–272.
Gotoh, O., 1982. An improved algorithm for matching biological sequences. J. Mol. Biol. 162 (3), 705–708.
Guigo, R., et al., 1992. Prediction of gene structure. J. Mol. Biol. 226 (1), 141–157.
Guigo, R., 1998. Assembling genes from predicted exons in linear time with dynamic programming. J. Comput. Biol. 5 (4), 681–702.
Guigo, R., et al., 2003. Comparison of mouse and human genomes followed by experimental verification yields an estimated 1019 additional genes. Proc. Natl. Acad. Sci. USA
100 (3), 1140–1145.
Guigo, R., et al., 2006. EGASP: The human ENCODE genome annotation assessment project. Genome Biol. 7 (Suppl. 1), S2 1–31.
Harrow, J., et al., 2012. GENCODE: The reference human genome annotation for The ENCODE project. Genome Res. 22 (9), 1760–1774.
Hastings, K.E.M., 2005. SL trans-splicing: Easy come or easy go? Trends Genet. 21 (4), 240–247.
Hoff, K.J., et al., 2016. BRAKER1: Unsupervised RNA-seq-based genome annotation with GeneMark-ET and AUGUSTUS. Bioinformatics 32 (5), 767–769.
Hubbard, T., et al., 2002. The Ensembl genome database project. Nucleic Acids Res. 30 (1), 38–41.
Hyatt, D., et al., 2010. Prodigal: Prokaryotic gene recognition and translation initiation site identification. BMC Bioinform. 11, 119.
International HapMap Consortium, C., 2005. A haplotype map of the human genome. Nature 437 (7063), 1299–1320.
Iseli, C., Jongeneel, C.V., Bucher, P., 1999. ESTScan: A program for detecting, evaluating, and reconstructing potential coding regions in EST sequences. Proc. Int. Conf. Intell.
Syst. Mol. Biol. 138–148.
Ji, Z., et al., 2015. Many lncRNAs, 50 UTRs, and pseudogenes are translated and some are likely to express functional proteins. eLife 4.
Katzav, S., Martin-Zanca, D., Barbacid, M., 1989. vav, a novel human oncogene derived from a locus ubiquitously expressed in hematopoietic cells. EMBO J. 8 (8), 2283–2290.
Keinan, A., Clark, A.G., 2012. Recent explosive human population growth has resulted in an excess of rare genetic variants. Science 336 (6082), 740–743.
Kent, W.J., et al., 2002. The human genome browser at UCSC. Genome Res. 12 (6), 996–1006.
Korf, I., et al., 2001. Integrating genomic homology into gene structure prediction. Bioinformatics 17 (Suppl. 1), S140–S148.
Kowalczyk, M.S., Higgs, D.R., Gingeras, T.R., 2012. Molecular biology: RNA discrimination. Nature 482 (7385), 310–311.
Kryukov, G.V., et al., 2003. Characterization of mammalian selenoproteomes. Science 300 (5624), 1439–1443.
Kryukov, G.V., Kryukov, V.M., Gladyshev, V.N., 1999. New mammalian selenocysteine-containing proteins identified with an algorithm that searches for selenocysteine insertion
sequence elements. J. Biol. Chem. 274 (48), 33888–33897.
Kumar, P., Henikoff, S., Ng, P.C., 2009. Predicting the effects of coding non-synonymous variants on protein function using the SIFT algorithm. Nat. Protoc. 4 (7), 1073–1082.
Kyriakopoulou, C., et al., 2006. U1-like snRNAs lacking complementarity to canonical 50 splice sites. RNA 12 (9), 1603–1611.
Krzywinski, M., et al., 2009. Circos: an Information Aesthetic for Comparative Genomics. Genome Res. 19, 1639–1645.
Landrum, M.J., et al., 2016. ClinVar: Public archive of interpretations of clinically relevant variants. Nucleic Acids Res. 44 (D1), D862–D868.
Langmead, B., Salzberg, S.L., 2012. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9 (4), 357–359.
Lescure, A., et al., 1999. Novel selenoproteins identified in silico and in vivo by using a conserved RNA structural motif. J. Biol. Chem. 274 (53), 38147–38154.
Lewis, S.E., et al., 2002. Apollo: A sequence annotation editor. Genome Biol. 3 (12), (RESEARCH0082).
Li, H., Durbin, R., 2009. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25 (14), 1754–1760.
Loveland, J.E., et al., 2012. Community gene annotation in practice. Database – J. Biol. Databases Curation.
Lukashin, A.V., Borodovsky, M., 1998. [Link]: New solutions for gene finding. Nucleic Acids Res. 26 (4), 1107–1115.
MacArthur, J., et al., 2017. The new NHGRI-EBI Catalog of published genome-wide association studies (GWAS Catalog). Nucleic Acids Res. 45 (D1), D896–D901.
Madupu, R., et al., 2010. Meeting report: A workshop on best practices in genome annotation. Database (Oxford) 2010, baq001.
Margulies, E.H., et al., 2003. Identification and characterization of multi-species conserved sequences. Genome Res. 13 (12), 2507–2518.
Mariotti, M., Guigo, R., 2010. Selenoprofiles: Profile-based scanning of eukaryotic genome sequences for selenoprotein genes. Bioinformatics 26 (21), 2656–2663.
Meyer, I.M., Durbin, R., 2002. Comparative ab initio prediction of gene structures using pair HMMs. Bioinformatics 18 (10), 1309–1318.
Meyer, I.M., Durbin, R., 2004. Gene structure conservation aids similarity based gene prediction. Nucleic Acids Res. 32 (2), 776–783.
Miller, W., 2001. Comparison of genomic DNA sequences: Solved and unsolved problems. Bioinformatics 17 (5), 391–397.
Monaco, A.P., et al., 1986. Isolation of candidate cDNAs for portions of the Duchenne muscular dystrophy gene. Nature 323 (6089), 646–650.
Mortazavi, A., et al., 2008. Mapping and quantifying mammalian transcriptomes by RNA-Seq. Nat. Methods 5 (7), 621–628.
Mott, R., 1997. EST_GENOME: A program to align spliced DNA sequences to unspliced genomic DNA. Comput. Appl. Biosci. 13 (4), 477–478.
Mouse Genome Sequencing Consortium, 2002. Initial sequencing and comparative analysis of the mouse genome. Nature 420 (6915), 520–562.
Mudge, J.M., Harrow, J., 2016. The state of play in higher eukaryote gene annotation. Nat. Rev. Genet. 17 (12), 758–772.
Needleman, S.B., Wunsch, C.D., 1970. A general method applicable to the search for similarities in the amino acid sequence of two proteins. J. Mol. Biol. 48 (3), 443–453.
Nesterova, T.B., et al., 2001. Characterization of the genomic Xist locus in rodents reveals conservation of overall gene structure and tandem repeats but rapid evolution of
unique sequence. Genome Res. 11 (5), 833–849.
Parra, G., et al., 2003. Comparative gene prediction in human and mouse. Genome Res. 13 (1), 108–117.
Parra, G., et al., 2006. Tandem chimerism as a means to increase protein complexity in the human genome. Genome Res. 16 (1), 37–44.
Parra, G., Bradnam, K., Korf, I., 2007. CEGMA: A pipeline to accurately annotate core genes in eukaryotic genomes. Bioinformatics 23 (9), 1061–1067.
Patel, A.A., Steitz, J.A., 2003. Splicing double: Insights from the second spliceosome. Nat. Rev. Mol. Cell Biol. 4 (12), 960–970.
Pearson, W.R., 2000. Flexible sequence similarity searching with the FASTA3 program package. Methods Mol. Biol. 132, 185–219.
Pearson, W.R., Lipman, D.J., 1988. Improved tools for biological sequence comparison. Proc. Natl. Acad. Sci. USA 85 (8), 2444–2448.
Pedersen, C.N.S., Scharling, T., 2002. Comparative methods for gene structure prediction in homologous sequences. Proc. Algorithms Bioinform. 2452, 220–234.
Pennisi, E., 2000. Ideas fly at gene-finding jamboree. Science 287 (5461), 2182. þ .
Plotkin, J.B., Kudla, G., 2011. Synonymous but not the same: The causes and consequences of codon bias. Nat. Rev. Genet. 12 (1), 32–42.
14 Genome Annotation

Potter, S.C., et al., 2004. The Ensembl analysis pipeline. Genome Res. 14 (5), 934–941.
Powers, D., 2011. Evaluation: From precision, recall and F-measure to ROC, informedness, markedness and correlation. Int. J. Mach. Learn. Technol. 2, 37–63.
Pruitt, K.D., et al., 2012. NCBI Reference Sequences (RefSeq): Current status, new features and genome annotation policy. Nucleic Acids Res. 40 (Database issue), D130–D135.
Rayman, M.P., 2000. The importance of selenium to human health. Lancet 356 (9225), 233–241.
Reese, M.G., et al., 2000. Genome annotation assessment in Drosophila melanogaster. Genome Res. 10 (4), 483–501.
Rivas, E., Clements, J., Eddy, S.R., 2017. A statistical test for conserved RNA structure shows lack of evidence for structure in lncRNAs. Nat. Methods 14 (1), 45–48.
Rogers, M.F., et al., 2012. SpliceGrapher: Detecting patterns of alternative splicing from RNA-Seq data in the context of gene models and EST data. Genome Biol. 13 (1), R4.
Salamov, A.A., Solovyev, V.V., 2000. Ab initio gene finding in Drosophila genomic DNA. Genome Res. 10 (4), 516–522.
Schweikert, G., et al., 2009. mGene: Accurate SVM-based gene finding with an application to nematode genomes. Genome Res. 19 (11), 2133–2143.
Sharon, D., et al., 2013. A single-molecule long-read survey of the human transcriptome. Nat. Biotechnol. 31 (11), 1009–1014.
Sharp, P.A., 2005. The discovery of split genes and RNA splicing. Trends Biochem. Sci. 30 (6), 279–281.
Siepel, A., et al., 2005. Evolutionarily conserved elements in vertebrate, insect, worm, and yeast genomes. Genome Res. 15 (8), 1034–1050.
Simao, F.A., et al., 2015. BUSCO: Assessing genome assembly and annotation completeness with single-copy orthologs. Bioinformatics 31 (19), 3210–3212.
Slater, G.S., Birney, E., 2005. Automated generation of heuristics for biological sequence comparison. BMC Bioinform. 6, 31.
Sleutels, F., Zwart, R., Barlow, D.P., 2002. The non-coding Air RNA is required for silencing autosomal imprinted genes. Nature 415 (6873), 810–813.
Smith, T.F., Waterman, M.S., 1981. Identification of common molecular subsequences. J. Mol. Biol. 147 (1), 195–197.
Smit, A.F.A., Hubley, R., Green, P., RepeatMasker Open-4.0, 2013-2015, [Link]
Solovyev, V., et al., 2006. Automatic annotation of eukaryotic genes, pseudogenes and promoters. Genome Biol. 7 (Suppl. 1), S10 1–12.
Stanke, M., et al., 2006a. Gene prediction in eukaryotes with a generalized hidden Markov model that uses hints from external sources. BMC Bioinform. 7.
Stanke, M., Tzvetkova, A., Morgenstern, B., 2006b. AUGUSTUS at EGASP: Using EST, protein and genomic alignments for improved gene prediction in the human genome.
Genome Biol. 7.
Steijger, T., et al., 2013. Assessment of transcript reconstruction methods for RNA-seq. Nat. Methods 10 (12), 1177–1184.
Stein, L.D., et al., 2002. The generic genome browser: A building block for a model organism system database. Genome Res. 12 (10), 1599–1610.
Struhl, K., 2007. Transcriptional noise and the fidelity of initiation by RNA polymerase II. Nat. Struct. Mol. Biol. 14 (2), 103–105.
Su, M., et al., 2013. Small proteins: Untapped area of potential biological importance. Front. Genet. 4, 286.
Tatusova, T., et al., 2016. NCBI prokaryotic genome annotation pipeline. Nucleic Acids Res. 44 (14), 6614–6624.
The Gene Ontology Consortium,, 2017. Expansion of the Gene Ontology knowledgebase and resources. Nucleic Acids Res. 45 (D1), D331–D338.
Thibaud-Nissen, F., et al., 2016. The NCBI eukaryotic genome annotation pipeline. J. Anim. Sci. 94, 184.
Thorvaldsdottir, H., Robinson, J.T., Mesirov, J.P., 2013. Integrative Genomics Viewer (IGV): High-performance genomics data visualization and exploration. Brief. Bioinform. 14
(2), 178–192.
Wheelan, S.J., Church, D.M., Ostell, J.M., 2001. Spidey: A tool for mRNA-to-genomic alignments. Genome Res. 11 (11), 1952–1957.
Wiberg, R.A.W., et al., 2015. Assessing recent selection and functionality at long noncoding RNA loci in the mouse genome. Genome Biol. Evol. 7 (8), 2432–2444.
Wiehe, T., et al., 2001. SGP-1: Prediction and validation of homologous genes based on sequence alignments. Genome Res. 11 (9), 1574–1583.
Wu, J.Q., et al., 2004. Identification of rat genes by TWINSCAN gene prediction, RT-PCR, and direct sequencing. Genome Res. 14 (4), 665–671.
Xu, Y., Mural, R.J., Uberbacher, E.C., 1997. Inferring gene structures in genomic sequences using pattern recognition and expressed sequence tags. In: Proceedings Ismb-97 –
Fifth International Conference on Intelligent Systems for Molecular Biology, pp. 344–353.
Yeh, R.F., Lim, L.P., Burge, C.B., 2001. Computational inference of homologous gene structures in the human genome. Genome Res. 11 (5), 803–816.
Zhang, Y., et al., 2005. Pyrrolysine and selenocysteine use dissimilar decoding strategies. J. Biol. Chem. 280 (21), 20740–20751.
Zheng, D., et al., 2007. Pseudogenes in the ENCODE regions: Consensus annotation, analysis of transcription, and evolution. Genome Res. 17 (6), 839–851.
Zinoni, F., et al., 1986. Nucleotide sequence and expression of the selenocysteine-containing polypeptide of formate dehydrogenase (formate-hydrogen-lyase-linked) from
Escherichia coli. Proc. Natl. Acad. Sci. USA 83 (13), 4650–4654.

Biographical Sketch

Josep F. Abril, PhD Associate Professor at the Department of Genetics, Microbiology and Statistics of the Uni-
versitat de Barcelona. He earned his Bachelor’s degree in Biology from the Universitat de Barcelona, and his PhD
in Bioinformatics from the Universitat Pompeu Fabra in Barcelona. His research focuses on the computational
analysis of sequences and their annotated features. He has worked on a variety of genomic and transcriptomic
projects, but also on the functional characterization of proteins, and on different areas in computational biology.
These include the development of protocols for the assembly and annotation of genomes, the use of comparative
genomics methods, the analysis of gene expression, the modeling of splice sites and exonic structure of eukaryotic
genes, the visualization of whole-genome annotations – this includes the human genome map published in
Science in 2001 – the integration of expression and variation data into interaction networks, as well as the
characterization of viral samples from metagenomic experiments. His organisms of interest are planarians, flies,
and humans, not necessarily in that order, although he contributed to the annotation of other species too. Finally,
he has been involved in the organization and analysis of three of the most relevant gene-prediction accuracy
assessments in computational gene prediction, namely GASP in the Drosophila genome and EGASP and RGASP in
the human one.
 Genome Annotation 15

Sergi Castellano, PhD Associate Professor at the Genetics and Genomics Medicine Programme at University
College London. He received his Bachelor’s degree in Biology from the Universitat de Barcelona, and his PhD in
Bioinformatics from the Universitat Pompeu Fabra, also in Barcelona. His group works on understanding the role
of essential micronutrients, with particular emphasis on selenium, in the adaptation of human metabolism to the
different environments encountered by archaic and modern humans. Much of his early work focused on the
identification and characterization of functional elements in eukaryotic genomes, primarily selenium-containing
genes, which are missannotated in standard databases. At the University of Hawaii, he developed the first data-
base, SELENODB, with correct annotations for selenium-containing genes across animal genomes. Later, while at
Cornell University and Howard Hughes Medical Institute, he used population genetics and molecular evolution
approaches to study the evolution of the use of selenocysteine, the selenium-containing amino acid, compared to
cysteine in proteins. More recently, at Max Planck in Germany, his group contributed to the population genetics
analysis of Neandertals as it relates to their coding variation and their first encounters – as early as 100,000 years
ago – with modern humans. His group is currently working on the role of micronutrient deficiencies in common
and rare disease.