High-throughput sequencing for biology and medicine.

Advances in genome sequencing have progressed at a rapid pace, with increased throughput accompanied by plunging costs. But these advances go far beyond faster and cheaper. High-throughput sequencing technologies are now routinely being applied to a wide range of important topics in biology and medicine, often allowing researchers to address important biological questions that were not possible before. In this review, we discuss these innovative new approaches-including ever finer analyses of transcriptome dynamics, genome structure and genomic variation-and provide an overview of the new insights into complex biological systems catalyzed by these technologies. We also assess the impact of genotyping, genome sequencing and personal omics profiling on medical applications, including diagnosis and disease monitoring. Finally, we review recent developments in single-cell sequencing, and conclude with a discussion of possible future advances and obstacles for sequencing in biology and health.

Introduction

Sequencing has progressed far beyond the analysis of DNA sequences, and is now routinely used to analyze other biological components such as RNA and protein, as well as how they interact in complex networks. In addition, increasing throughput and decreasing costs are making medical applications of sequencing a reality. Below we review various applications of next-generation sequencing as we experience it today and also describe future prospects and challenges, with a particular focus on human biology.

Next-generation sequencing (also ‘Next-gen sequencing' or NGS) refers to DNA sequencing methods that came to existence in the last decade after earlier capillary sequencing methods that relied upon ‘Sanger sequencing' (Sanger et al, 1977). As opposed to the Sanger method of chain-termination sequencing, NGS methods are highly parallelized processes that enable the sequencing of thousands to millions of molecules at once. Popular NGS methods include pyrosequencing developed by 454 Life Sciences (now Roche), which makes use of luciferase to read out signals as individual nucleotides are added to DNA templates, Illumina sequencing that uses reversible dye-terminator techniques that adds a single nucleotide to the DNA template in each cycle and SOLiD sequencing by Life Technologies that sequences by preferential ligation of fixed-length oligonucleotides. A recent review outlines a general timeline of the evolution of sequencing technologies and their features (Pareek et al, 2011). But these advances did not merely make the sequencing of DNA and RNA cheaper and more efficient; they have also helped create innovative new experimental approaches that delve deeper into the molecular mechanisms of genome organization and cellular function.

A prime example of the advances that have been facilitated by new sequencing technologies is the NHGRI-funded ENCODE project, which was launched in late 2003, based largely upon methods first developed in yeast (Iyer et al, 2001; Horak and Snyder, 2002) (Table I). The pilot phase of ENCODE relied heavily on microarray-based assays to analyze 1% of the human genome in unprecedented depth (Birney et al, 2007). With credit to advances in high-throughput sequencing, researchers expanded the scope of this project to include the whole human genome (Bernstein et al, 2012). A total of ∼1650 high-throughput experiments were performed to analyze transcriptomes and map elements, and identify methylation patterns in the human genome. This multi-institution consortia project has assigned biochemical activities to 80% of the genome, particularly annotating the portion of the genome that lies outside the well-studied protein-coding regions, including mapping over four million regulatory regions. This information has also enabled researchers to map genetic variants to gene regulatory regions and assess indirect links to disease (Boyle et al, 2012). Similar projects annotating the genome have also been performed for Drosophila melanogaster (Consortium et al, 2010), Caenorhabditis elegans (Gerstein et al, 2010) and mouse (Stamatoyannopoulos et al, 2012).

Table 1

The various NGS assays employed in the ENCODE project to annotate the human genome

Feature	Method	Description	Reference
	RNA-seq	Isolate RNA followed by HT sequencing	(Waern et al, 2011)
Transcripts, small RNA and transcribed regions	CAGE	HT sequencing of 5'-methylated RNA	(Kodzius et al, 2006)
	RNA-PET	CAGE combined with HT sequencing of poly-A tail	(Fullwood et al, 2009c)
	ChIRP-Seq	Antibody-based pull down of DNA bound to lncRNAs followed by HT sequencing	(Chu et al, 2011)
	GRO-Seq	HT sequencing of bromouridinated RNA to identify transcriptionally engaged PolII and determine direction of transcription	(Core et al, 2008)
	NET-seq	Deep sequencing of 3′ ends of nascent transcripts associated with RNA polymerase, to monitor transcription at nucleotide resolution	(Churchman and Weissman, 2011)
	Ribo-Seq	Quantification of ribosome-bound regions revealed uORFs and non-ATG codons	(Ingolia et al, 2009)
Transcriptional machinery and protein–DNA interactions	ChIP-seq	Antibody-based pull down of DNA bound to protein followed by HT sequencing	(Robertson et al, 2007)
	DNAse footprinting	HT sequencing of regions protected from DNAse1 by presence of proteins on the DNA	(Hesselberth et al, 2009)
	DNAse-seq	HT sequencing of hypersensitive non-methylated regions cut by DNAse1	(Crawford et al, 2006)
	FAIRE	Open regions of chromatin that is sensitive to formaldehyde is isolated and sequenced	(Giresi et al, 2007)
	Histone modification	ChIP-seq to identify various methylation marks	(Wang et al, 2009a)
DNA methylation	RRBS	Bisulfite treatment creates C to U modification that is a marker for methylation	(Smith et al, 2009)
Chromosome-interacting sites	5C	HT sequencing of ligated chromosomal regions	(Dostie et al, 2006)
	ChIA-PET	Chromatin-IP of formaldehyde cross-linked chromosomal regions, followed by HT sequencing	(Fullwood et al, 2009a)

Here, we provide an overview of the new fields of biology that were made possible by advancements in DNA and RNA sequencing technologies. We briefly review techniques that were made more efficient, higher-throughput, higher-resolution and genome-wide with the introduction of sequencing, and also discuss fundamentally new types of analyses that rely heavily on the constantly improving sequencing technologies. Their relevance in the clinical context is also highlighted.

Genomes, variation and epigenomics

Genome sequencing with next-generation technologies was first applied to bacterial genomes using 454 technology (Smith et al, 2007). Decreasing costs have made these technologies a sufficiently commonplace that a large number of different organisms have been sequenced. As of June 2012, according to the Genomes Online Database, a total of 3920 bacterial and 854 different eukaryotic genomes have been completely sequenced (Pagani et al, 2012). Although resequencing new lines and closely related organisms is readily achieved, there are still significant challenges (Snyder et al, 2010). Different DNA sequencing platforms have different biases and abilities to call variants (Clark et al, 2011; Lam et al, 2012). Short indels (insertions and deletions) and larger structural variants are also particularly difficult to call (see below). De novo genome assembly can be attempted from short reads, but this remains difficult and leads to short contigs. Increasing read length and accuracy will greatly enhance our abilities to accurately sequence genomes de novo, which will also enable more precise mapping of variants between individuals.

Genome sequence and structural variation

In addition to the sequencing of the genomes of different organisms, projects to characterize the DNA sequence of individuals have gathered pace, and whole-genome sequencing of humans is becoming commonplace (Gonzaga-Jauregui et al, 2012). The reduced costs, increased accuracy and lowered data turn-around time associated with NGS have enabled clinicians and medical researchers to identify susceptibility markers and inherited disease traits (see ‘Medical Genomic Sequencing'). Identifying damaging polymorphisms in coding regions (exonic variants) and those present in other functional regions (discussed below) of the genome are an integral part of clinical genomics. In order to achieve this goal, several groups are studying human genomic variation by sequencing or genotyping large number of individuals, including multi-institute consortia projects such as the 1000 Genomes Project (Consortium, 2010), the Personal Genome Project (Ball et al, 2012), the HapMap project (Consortium, 2003) and the pan-Asian single-nucleotide polymorphism (SNP) project (Abdulla et al, 2009). The different human genome sequencing projects have revealed that individuals have ∼3.1–4 M SNPs between one another and the reference sequence (Consortium, 2003; Frazer et al, 2007), and, thus far, a total of over 30 M SNPs have been discovered from human genome sequencing projects. Studies have been successful in linking variants with a range of conditions, a catalogue of which is available at dbGaP, the database of Genotype and Phenotype (Mailman et al, 2007).

One area that has been particularly challenging in the sequencing of human genomes and other complex genomes are structural variations (SVs): large (>1 kb) segments of the genome that are duplicated, deleted or rearranged relative to reference sequences and among individuals (Figures 1A and B). Early microarray experiments indicated that SVs were abundant in the human genome (Louie et al, 2003; Conrad et al, 2006; Redon et al, 2006), although it was the advent of NGS that revealed that this is much more prevalent than previously appreciated (Ng et al, 2005; Chiu et al, 2006; Dunn et al, 2007; Korbel et al, 2007; Ng et al, 2007). Presently, four different approaches are used to map structural variants in genomes (Snyder et al, 2010). These include paired-end mapping (Korbel et al, 2007), read depth (Abyzov et al, 2011), split reads (Zhang et al, 2011) and mapping sequences to breakpoint junctions (Kidd et al, 2010). Each has its own biases, but typically all four are used to help identify SVs. SVs affect genes as well as transcription factor-binding sites, resulting in altered expression profiles of downstream genes (Snyder et al, 2010). Copy number variation has also been known to be associated with various diseases including glomerulonephritis (Aitman et al, 2006), Crohn's disease (McCarroll et al, 2008), HIV-1/AIDS (Gonzalez et al, 2005) and psoriasis (de Cid et al, 2009). Although much work remains to be done, it is clear that SVs have a significant impact on disease regulation and health, making this an important class of elements to map in eukaryotic genomes.

Figure 1

Dimensionality of the genome. The understanding of the human genome has expanded with advances of sequencing technologies, from (A) 1D sequencing of the human genome to (B) 2D mapping of SVs using methods such as paired-end sequencing, (C) 3D genome-wide chromosomal conformation capture using ChIA-PET and Hi-C, and (D) four dimensions across time.

Mapping higher-order organization in eukaryotic genomes

New sequencing technologies have also enabled the mapping of three-dimensional (3D) DNA interactions that were previously not possible on a genomic scale and resolution (Figure 1C). DNA analyses first became 3D with the development of chromosome conformation capture techniques such as 3C, 4C and 5C (Dekker et al, 2002; Dekker, 2006; Dostie et al, 2006; Simonis et al, 2006; Zhao et al, 2006; Dostie and Dekker, 2007). However, these techniques offered 3D mapping of DNA interactions only within regions where interactions were already expected (hypothesis-driven). Further, primers had to be designed for each region, which made it very low throughput. With the invention of Hi-C, which utilizes NGS on cross-linked DNA fragments that have been sheared and digested to an optimal size to identify all DNA regions that are physically close together, genome-wide mapping of chromosomal 3D structures became possible, at least at low resolution (20–100 kb) (Lieberman-Aiden et al, 2009; Zhang et al, 2012b).

These newly developed sequencing methods provided important new insights into the global organization of eukaryotic genomes that were previously unattainable. Analyses of individual regions revealed that some distantly located regulatory elements, such as promoters, enhancers and insulators, come into close proximity to better mediate their activities (Branco and Pombo, 2006; Woodcock, 2006; Fraser and Bickmore, 2007; Osborne and Eskiw, 2008). Transcription factor-mediated 3D interactions obtained using immunoprecipitation followed by paired-end sequencing (ChIA-PET) (Fullwood et al, 2009a, Fullwood et al, 2009b, Fullwood et al, 2010) revealed extensive interaction between enhancer and promoter regions, often encoded at long distances from one another on the chromosome (Fullwood et al, 2009a; Handoko et al, 2011; Li et al, 2012). These large-scale analyses also revealed that chromosomal regions are organized together into territories of similar biological activity, such as active and inactive domains. These topological domains seem to be conserved across multiple cell types and mammalian species (Lieberman-Aiden et al, 2009; Cremer and Cremer, 2010; Sung and Hager, 2011; Dixon et al, 2012). Figure 1 summarizes some of the ways that high-throughput sequencing technologies have extended our understanding of the structural organization of genomes.

DNA and histone modification

Besides deciphering the sequence of genomes, NGS has also enabled the mapping of epigenetic marks such as DNA methylation (DNAm) and histone modification patterns in a genome-wide manner (Figure 2).

Figure 2

Sequencing technologies and their uses. Various NGS methods can precisely map and quantify chromatin features, DNA modifications and several specific steps in the cascade of information from transcription to translation. These technologies can be applied in a variety of medically relevant settings, including uncovering regulatory mechanisms and expression profiles that distinguish normal and cancer cells, and identifying disease biomarkers, particularly regulatory variants that fall outside of protein-coding regions. Together, these methods can be used for integrated personal omics profiling to map all regulatory and functional elements in an individual. Using this basal profile, dynamics of the various components can be studied in the context of disease, infection, treatment options, and so on. Such studies will be the cornerstone of personalized and predictive medicine.

Methylation of cytosine residues in DNA is the most studied epigenetic marker and is known to silence parts of the genome by inducing chromatin condensation (Newell-Price et al, 2000). DNAm can be stably inherited in multiple cell divisions, thereby enabling it to regulate biological processes, such as cellular differentiation (Reik, 2007), tissue-specific transcriptional regulation (Lister et al, 2009), cell identity (Feldman et al, 2006; Feng et al, 2006) and genomic imprinting (Li et al, 1993). Hypermethylation of the promoters of tumor-suppressor genes has also been linked to retinoblastoma, colorectal cancer, leukemia, breast and ovarian cancers (Baylin, 2005). Such knowledge of hypermethylation is crucial in treatments, such as in the case of acute myeloid leukemia, where treatment with DNA methyl transferase inhibitor azacytidine has been shown to be successful in clinical trials (Silverman et al, 2002). Precise mapping of these methylation patterns genome wide has only been made possible by various NGS techniques, including methylated DNA immunoprecipitation (Taiwo et al, 2012), MethylC-seq (Lister et al, 2009) and reduced representation bisulfite sequencing (RRBS) (Meissner et al, 2005). The latter two methods make use of sodium bisulfite conversion of unmethylated cytosine to uracil for identification of methylation patterns.

The nucleosomes around which DNA is bound are composed of dimers made up of four basic proteins—H2A–H2B and H3–H4—which are modified post-translationally in a variety of ways, including acetylation, methylation, phosphorylation and sumoylation. Histone modification sites can be identified in a genome-wide manner by the same method used to detect proteins bound to DNA (ChIP-seq), using antibodies that specifically recognize the chemical modifications. Using such a method, 39 different histone modifications were revealed in CD4+ T cells, which were used to delineate between promoters and enhancers (Wang et al, 2008). More recently, the ENCODE project mapped 12 types of histone modifications in 46 cell types (Bernstein et al, 2012), revealing cell type-specific patterns of histone modifications.

Depending on the particular modification on nucleosomes, specific regulatory proteins can be recruited to the site, resulting in the activation or repression of nearby genes (Barski et al, 2007). Histone modification is thus a very important epigenetic mark that directly affects gene regulation, and aberrant modifications have been linked to gene dysregulation in disease in multiple studies. Scanning for five histone marks in 183 primary prostate cancer tissues, two subgroups with distinct patterns of histone modifications were obtained that had distinct risks of tumor recurrence, demonstrating the predictive power of histone marks in disease prognosis (Seligson et al, 2005). Aberrant activity of histone-modifying enzymes, such as the histone deacetylases, histone acetyl transferases and histone methyl transferases, or their cofactors, like S-adenosyl methionine and acetyl coenzyme A, results in global changes in histone modification. Apart from using inhibitors to these proteins (Park et al, 2004), site-directed, targeted restoration of the modifications might be a useful and important treatment strategy.

Transcriptomes and other functional elements in genomes

Beyond genome sequencing and interaction analyses, NGS has also enabled the global mapping of the transcriptome using RNA-sequencing (RNA-seq). High-throughput methods have enabled detection and quantification of transcripts, discovery of novel isoforms and linking of their expression to genomic variants (allele-specific variation). Significant interest also lies in uncovering the role of various regulatory factors in controlling the expression of genes, such as transcription factors and non-coding RNAs (Figure 2). We review these aspects in detail in the following sections.

Transcript detection and quantification

Microarray technologies provided the first practical technique for measuring genome-wide transcript levels. However, microarrays were only applicable to studying known genes, had significant problems with cross-hybridization and high noise levels, and had a limited dynamic range of only ∼200 fold (Wang et al, 2009a). Much more accurate measurement of mRNA levels became possible with the introduction of RNA-seq, which was invented in both yeast (Nagalakshmi et al, 2008; Wilhelm et al, 2010) and mammalian cells (Cloonan et al, 2008; Mortazavi et al, 2008). This method employs the high-throughput sequencing of cDNA fragments generated from a library of total RNA or fractionated RNA. It allows unambiguous mapping to unique regions of the genome and hence, essentially, there is little or no background noise. RNA-seq allows the precise quantification of transcripts and exons, and also the analysis of transcript isoforms with at least a 5000-fold dynamic range (Wang et al, 2009b). Not only is RNA-seq able to quantify more accurately the transcriptome consisting of known genes, it is also a great tool for identifying novel genes and RNAs that microarray technologies could not achieve. This includes the identification of novel expressed fusion genes using paired-end RNA-seq (Edgren et al, 2011), as well as the discovery of new non-coding RNAs such as lincRNAs (Prensner et al, 2011).

Mapping transcript isoforms involves precise mapping of reads to known and potential splice junctions or the use of assembly to generate transcript isoforms followed by mapping to genomic regions. Eukaryotic transcriptomes are quite complex, and an average of five or more transcript isoforms have been reported for each gene (Birney et al, 2007). This figure is likely an underestimate as additional novel transcript isoforms may be discovered with increased sequencing depth (Ameur et al, 2010; Wu et al, 2010). Paired-end sequencing allows better mapping of transcript isoforms (Ameur et al, 2010; Wu et al, 2010), although the precise deduction of the ensemble of gene transcripts from multi-exon genes still remains a significant challenge. Increased read length will better enable the complexity of transcripts that are produced. RNA-seq also enables mapping allele-specific expression (ASE) (Zhang et al, 2009) and the identification of editing sites (Li et al, 2009), both of which are extensive in eukaryotic transcriptomes (Chen et al, 2012).

The ability to detect and accurately quantify transcript levels using NGS technologies has significant impacts in the clinic. Altered expression of specific isoforms have been identified to be detrimental in ischemic stroke (Gretarsdottir et al, 2003) and type 2 diabetes (Horikawa et al, 2000) among others; ASE of the TGF beta type 1 receptor confers genetic predisposition to colorectal cancer (Valle et al, 2008); and ASE of proapoptotic gene DAPK1 is associated with chronic lymphocytic leukemia (Lynch et al, 2002). Allelic imbalances that result in altered gene expression profiles were compared across oral squamous cell carcinoma tumors and matched normal tissues (Tuch et al, 2010). These genes were enriched in cancer-related functions and indicate that allelic imbalance is an underlying cause of cancer etiology. Transcriptome profiling using RNA-seq also revealed several novel transcripts and gene fusions in melanoma (Berger et al, 2010) and Alzheimer's disease (Twine et al, 2011), emphasizing the importance of high-throughput sequencing in the understanding of human diseases.

Profiling transcript production and ribosome-bound mRNAs

Transcript abundance is only one measure for analyzing the expression of gene products. Recently, it has become possible to measure the production of nascent RNAs by bromo-uridinating nuclear run-on RNA molecules and sequencing them (GRO-Seq, for Global Run-On Sequencing) (Core et al, 2008) or by immunoprecipitation of RNA polymerase followed by sequencing the bound RNA fragments, a process called NET-seq (Churchman and Weissman, 2001; Churchman and Weissman, 2011). The dynamics of transcript synthesis and decay can also be tracked using dynamic transcriptome analysis (DTA) (Miller et al, 2011). These methods not only identify RNA polymerase II-bound transcripts but also the direction of transcription and its rate of decay. These efforts have revealed promoter-proximal pausing and active genes. More than twice the number of active genes has also been discovered in the lung fibroblast, as compared with the number of active genes obtained from a microarray of the same cell line (Core et al, 2008).

In addition to transcriptional control, protein expression is controlled at the level of translation. Ingolia et al (2009) developed Ribo-Seq to measure the quantities of ribosome-bound fragments by first freezing ribosomes and using the translation inhibitor cycloheximide. The mRNA is then digested and the resulting fragments sequenced to reveal mRNA regions occupied by ribosomes. The quantification of ribosome-bound regions is used as a proxy for translation efficiency. These studies have revealed that many upstream ORFs in mRNA are bound to ribosomes, that many non-ATG codons are used, and that ribosome occupancy and mRNA show a partial correlation. Thus, high-throughput sequencing has provided considerable insight into many levels of gene expression.

Genome-wide identification of protein–DNA interactions

Much of gene regulation is thought to occur at the level of transcriptional control, and the binding sites of transcription factors are associated with regulation of gene expression. Experimental identification of these sites has been an area of high interest and constant improvement. The first experiments to map transcription factor-binding sites genome wide used chromatin immunoprecipitation (ChIP) of a transcription factor of interest followed by recovery of the associated DNA and probing on DNA microarrays (ChIP–chip) (Iyer et al, 2001; Horak and Snyder, 2002). This method, however, was noisy and expensive to apply to large genomes. Sequencing technologies made widespread application of genomic ChIP profiling to the human genome practical. Protein–DNA interactions based on NGS (ChIP-seq) not only provided clear indications of transcription factor-binding sites at high resolution, but also enabled genome-wide mapping of histone marks (Figure 2). ChIP-seq (Johnson et al, 2007; Robertson et al, 2007) was similar to ChIP–chip in that DNA associated with a transcription factor or histone modification of interest was enriched by immunoprecipitation, but was followed by NGS of the DNA and mapping the sequence reads back to the genome (Robertson et al, 2007) rather than hybridization to a microarray. ChIP-seq has been applied to many studies such as global analyses of several DNA-binding regions (as in the ENCODE project), as well as mapping regulatory differences between individuals and in disease settings. Genome-wide binding profiles across 10 individuals (lymphoblastoid cell lines) for two transcription factors, NFκB and PolII, revealed significant binding differences between any two individuals (7.5% for NFkB and 25% for PolII-binding sites). These also correlate to the expression of the downstream target genes (Kasowski et al, 2010). In another example, polymorphisms in a gene desert associated to coronary artery disease were found to affect STAT1 binding, resulting in altered expression of neighboring genes. These long-range enhancer interactions support the importance of regulatory polymorphisms as disease biomarkers (Harismendy et al, 2011).

Other complementary techniques to globally identify potential regulatory regions include the identification of DNAse1 hypersensitive sites, using formaldehyde-assisted isolation of regulatory elements (FAIRE) (Nammo et al, 2011) and Sono-Seq (Auerbach et al, 2009) (Figure 2). These methods globally map large numbers of potential regulatory sites across the human genome, although in most cases what these elements bind is not known.

Besides proteins that map to chromosomes, RNA species such as long non-coding RNAs (lncRNA) are also important regulators of the chromatin structure and are involved in several biological processes (Wang and Chang, 2011). An effective method, ChIRP (chromatin isolation by RNA purification), has been developed (Chu et al, 2011), which can effectively detect the interaction of lncRNAs and chromatin in a genome-wide scale (Figure 2). LncRNA is crosslinked with glutaraldehyde and hybridized to oligonucleotide tiles. The sequence bound to the complex is then determined using NGS.

Medical genomic sequencing

Genomic sequencing will have an enormous impact on the field of medicine. Until recently, cost and throughput limitations have made general clinical applications infeasible. Currently, though, the price of about 5000USD for a normal human genome sequence (not counting analysis) and fast throughput (several days to a few weeks) is rapidly making medical sequencing practical. Indeed, high-throughput sequencing has already been used to help diagnose highly genetically heterogeneous disorders, such as X-linked intellectual disability, congenital disorders of glycosylation and congenital muscular dystrophies (Zhang et al, 2012a); to detect carrier status for rare genetic disorders (Tester and Ackerman, 2011; Zhang et al, 2012a); and to provide less-invasive detection of fetal aneuploidy through the sequencing of free fetal DNA (Fan et al, 2008, 2012).

While this is a promising start for high-throughput sequencing in the clinic, these technologies must be used with caution as they have non-negligible false-positive and false-negative rates owing to sequencing errors and amplification biases, which need to be improved upon with optimized library construction methods, improved sequencing technologies or filtering algorithms. Nonetheless, medical sequencing could potentially be applied in a wide range of settings in the future. Here, we highlight three main areas: cancer, hard-to-diagnose diseases and personalized medicine.

Genome sequencing in cancer

Cancer is a genetic disease, both in predisposition and somatic growth. High-throughput sequencing of cancer genomes has been a major factor in the understanding of the genetics of this complex disease. Exome sequencing, RNA sequencing, paired-end sequencing and whole-genome sequencing of cancer genomes have led to a dramatic increase in the number of known recurrent somatic alterations, such as mutations, amplifications, deletions and translocations (Bass et al, 2011; Salzman et al, 2011; Fujimoto et al, 2012).

These studies have revealed many interesting findings. As a recent example, using paired-end sequencing, Inaki et al (2011) discovered that approximately half of all structural rearrangements in breast cancer genomes result in fusion transcripts, where single segmental tandem duplication spanning multiple genes is a major source. They estimated that 44% of these fusion transcripts are potentially translated, and found a novel RPS6KB1–VMP1 fusion gene that is recurrent in a third of breast cancer samples analyzed, with potential association with prognosis. Simultaneously, Hillmer et al (2011) applied paired-end sequencing on cancer and non-cancer human genomes, and found that non-cancer genomes contain more inversion, deletions and insertions, whereas cancer genomes are dominated by duplications, translocations and complex rearrangements. Recent works from Korbel et al and others have found that cancer genomes lacking p53 often contain genomic regions that undergo extensive rearrangements called ‘chromothripsis', suggestive of complex chromosome shattering and rejoining in a single event (Nowell, 1976; Korbel et al, 2007; Stratton et al, 2009; Kloosterman et al, 2011; Stephens et al, 2011; Tubio and Estivill, 2011; Rausch et al, 2012). Much work has also been done on matched tumor–normal pairs and revealed that extensive somatic SNVs and SVs occur in cancer genomes (Kumar et al, 2011; Wei et al, 2011; Banerji et al, 2012; Wang et al, 2012; Zang et al, 2012).

One important medical conclusion that has emerged from this work is that every tumor is genetically different but that common pathways are often activated. Thus, the sequencing of cancer genomes can help reveal the activated pathways and the information used to suggest therapeutic treatments. As an example, the detection of novel fusion transcripts in a difficult diagnostic case of acute promyelocytic leukemia that were previously missed in a regular diagnosis was used to influence the medical care of the patient (Welch et al, 2011). In addition, sequencing of carefully selected samples could lead to interesting discoveries of cancer evolution and mutational processes (Nik-Zainal et al, 2012a, 2012b).

Genome sequencing for clinical assessment of ‘mysterious' diseases

Whole-genome and -exome sequencing is likely to prove useful in the diagnosis of rare diseases and in selecting the optimal individualized treatment option for patients. This approach typically involves the use of families; sequencing of affected individuals and relatives along with inheritance patterns is used to deduce variants that are associated with a disease. Whole-exome sequencing performed on a four-member family led to the discovery of the causative gene for Miller's syndrome, an extremely rare condition that gives rise to micrognathia and cleft lips among other features (Ng et al, 2010). Nicholas Volker received a bone marrow transplant after his genome sequence indicated he had a mutation on the X chromosome that led to an inherited immune disorder that was giving him multiple problems. With the new diagnosis at hand, Volker was successfully treated and his severe inflammatory bowel disease alleviated (Worthey et al, 2011). Richard Gibbs describes using complete genome sequences of twins diagnosed with dopa-responsive dystonia to identify the appropriate treatment option, which eventually resulted in significant clinical improvements of the twins (Bainbridge et al, 2011). With multiple examples of whole-genome sequencing aiding the diagnosis and treatment of tough medical cases, sequencing in medical care is promising. However, it should be noted that in many cases, whole-genome sequencing of families does not always reveal the causative mutation. In some cases, it may suggest a list of possible candidates and in others, no obvious gene candidate is revealed. Clearly, a major bottleneck is the interpretation of gene variants and their effect on human health.

Personal genome sequencing for detecting medically actionable risks

Whole-genome sequencing and transcriptome analyses have shed light on mutations and expression alterations in individuals and in disease states. However, until recently, the power of genome sequencing for otherwise healthy individuals was unknown. Moreover, the integration of multiple different sequencing technologies amplifies the amount of information one can derive from medical examples by many fold. A recent example by Chen et al examined the power of personal genome sequencing of a healthy person to access disease risk, using integrated multiple ‘omics' data sets of a single individual in what they termed integrated personal omics profiling (Chen et al, 2012) (Figures 1D and 2). This study sequenced the genome of an individual at high accuracy and followed the transcriptomic, metabolomic and proteomic profiles of the single individual over a 14-month period. The integrated analysis not only allowed more complete understanding of the individual's genetic make-up and disease risks, but also tracked the emergence of type 2 diabetes. The extensive study revealed how various biological systems function and change together over the course of time as well as during the transition from a healthy to diseased state. The dynamic and complex nature of the human biological system emphasizes that such longitudinal monitoring of trends and changes may be the future of disease monitoring and even diagnosis.

However, many obstacles still lie between current medical practice and this kind of in-depth longitudinal patient monitoring. For one, the amount of time, money and effort needed to process such massive amounts of data for each patient is not practical at present. Further, the cost benefits of longitudinal patient monitoring in tracking disease onset and progression need to be more comprehensively assessed. Despite these formidable challenges, one cannot deny the promise such information holds for improving medical treatment and health management.

Single-cell sequencing

Biological research often involves the analysis of tissues, cell populations and whole organisms. However, much variation occurs at the single-cell level where understanding of each individual cell is crucial for the analysis of the entire system. Cancer cells, for example, are heterogeneous populations of multiple clonal expansions, and analyzing a tumor as one entity could mask many important characteristics of the tumor. The ideal approach to such systems biology thus requires analyzing ‘parts' of these systems individually, using methods and technologies that can extract data at few or single-cell levels (Schubert, 2011).

Single-cell sequencing in cancer

Most sequencing techniques that have been developed to date require DNA or RNA from over 105 cells (Metzker, 2008; Schuster, 2008; Metzker, 2010). This is a significant problem in solid tumors because of the heterogeneous nature of the tumors. In addition to multiclonal populations of cancer cells within each solid tumor, non-cancerous cells, such as blood cells and fibroblasts, are also present (Heppner, 1984; Marusyk and Polyak, 2010). This complex mixture of cells complicates analyses of data obtained from tumor sequencing, and signals from cancer cells tend to be masked by that from other cells. Determining gene expression and copy number by ‘averaging' across these complex cell populations is also far from ideal, and can give a measurement that is vastly different from the truth at the level of the individual cell (Wang and Bodovitz, 2010). Thus, separating these distinct cell populations and analyzing them individually is critical to a more thorough and accurate understanding of cancer. Laser capture microdissection is a method used to isolate tumor cells from their neighboring normal cell counterparts, in an attempt to get ‘pure' tumor cells for sequencing (Espina et al, 2006a, Espina et al, 2006b, , Espina et al 2007). Flow cytometry can also do the same, for tumor cells that are known to have a specific protein that is differentially expressed as compared with normal cells (Glogovac et al, 1996; van Beijnum et al, 2008). However, the heterogeneity of tumors still serves as a major problem, masking signals and making it difficult to differentiate signal from noise in bulk tissue analyses. Single-cell analysis using cytological methods and aCGH is possible, but only at limited resolution and coverage (Mark et al, 1998; Le Caignec et al, 2006; Fiegler et al, 2007; Fuhrmann et al, 2008; Hannemann et al, 2011).

Recent advances in single-cell sequencing enable significantly higher resolution than has been previously achieved. Navin et al was the first group to analyze tumors in such a manner. Using breast cancer as a model, they sequenced 100 single nuclei from distinct sections of a polygenomic breast tumor to obtain 50-kb copy number profiles, and showed that the tumor originated from three clonal subpopulations. They then sequenced another 100 single nuclei from a monogenomic primary tumor with matched liver metastasis, demonstrating that the primary tumor was from a single clonal expansion, and that the metastasis had arose from one of the cells in the primary tumor (Navin et al, 2011).

The Beijing Genomics Institute team extended this further by developing a high-throughput single-cell sequencing method that could reach single-nucleotide resolution. This technique was applied to conduct single-cell exome analysis of the JAK-2 negative neoplasm (Hou et al, 2012). Results demonstrated that this type of neoplasm arise from a single clonal expansion, and many novel mutated genes were identified (at >96% accuracy) that could be further explored for therapeutic purposes. The same technique was applied to a solid tumor of clear cell renal cell carcinoma, which revealed greater genetic complexity of the cancer than previously expected (Hou et al, 2012; Xu et al, 2012).

Taken together, it has been demonstrated that single-cell analyses of highly heterogeneous tissues provide much clearer intratumoral genetic pictures and developmental histories than previous bulk tissue sequencing. These developments finally allow tumor populations to be probed at an extremely high resolution with significantly lower noise signals from any non-cancerous cells and different subclones. This platform will serve to improve our understanding of how tumors develop, expand and progress.

Potential clinical applications of single-cell sequencing include detection of rare circulating tumor cells. These circulating tumor cells that are found in bodily fluids, such as the blood or urine, can now be isolated by microfluidic methods (Lien et al, 2010; Dickson et al, 2011; Xia et al, 2011; O'Flaherty et al, 2012). Genomics analyses can then be applied to the patients' DNA and RNA without the need of even a biopsy, which could be useful for both diagnosis and prognosis of the cancer in a non-invasive way. Single-cell sequencing of the biopsied tumor could also reveal if there is a multiclonal subpopulation of the cells as shown by Navin et al, and better personalized treatment options targeting the different mutations and aberrations in the subpopulations can be offered (Navin et al, 2011).

To make this a reality in clinics, work has to be done to compare single-cell diagnosis and prognosis of cancer to the current gold standards of clinical diagnosis and prognosis. Given the many different types of cancer, single-cell sequencing may only be useful for certain cancer types, depending on the amount of circulating tumor cells and the impact of clonality on the prognosis of each cancer subtype.

Single-cell sequencing in embryonic stem cell developmental biology

Previously, transcriptome analyses and whole-genome sequencing required a large number of cells, which made it inherently difficult to study gene expression or genomic variation within rare totipotent and pluripotent cell populations or within early embryos consisting of only a limited number of cells. How so many cell types can be derived from each pluripotent stem cell, and how each stem cell ‘knows' how to behave differently, has been an area of intense research. Indeed, within early animal embryos, each cell is likely to express specific transcriptional programs that define its eventual developmental fate (Gage and Verma, 2003; Sylvester and Longaker, 2004).

With the emergence of technologies that allow single-cell expression analysis, expression programs in 64-cell human blastocysts were determined resulting in the identification of distinct markers uniquely expressed in the different cell types of the blastocyst (Guo et al, 2010). Single-cell RNA sequencing enabled an even more in-depth and comprehensive analysis of the stem cell transcriptome at a genome-wide manner. With such capabilities, Tang et al (2009) identified over 1500 previously unknown splice junctions that could be critical for oogenesis.

Further comprehensive analysis of complex biological systems using single-cell approaches will undoubtedly provide new insights. It will likely reveal the myriad of underlying biological states that exist (e.g., cell cycle) as well as the role that stochastic events have in the formulation of complex cellular and developmental processes. For instance, although genetically identical cells in the same tissue type are usually analyzed as a homogenous population, they really are not (Elowitz et al, 2002). It has been suggested in multiple occasions that crosstalk happens between genetically identical cells (Elowitz et al, 2002; Sachs et al, 2005; Maheshri and O'Shea, 2007; Raj and van Oudenaarden, 2008; Snijder et al, 2009). Much is known about what happens within a cell, but knowledge of how information is transmitted from one cell to another, how cells communicate to accommodate variability, is very much lacking. There remains much to be discovered about how normal cells interact with one another, and how they function to maintain a homeostatic status despite high cell-to-cell variability. Resolution and depth have served as one of the most major obstacles in achieving this (Elowitz et al, 2002). With the new ability to analyze large populations of single-cell transcriptomes by high-throughput single-cell RNA sequencing, a largely unchartered realm of molecular biology will finally be accessible (Pelkmans, 2012).

Future developments

High-throughput sequencing, with its rapidly decreasing costs and increasing applications, is replacing many other research technologies. For example, gene expression studies are slowly moving from expression array technologies to RNA sequencing for the higher resolution, lower biases and ability to discover novel transcripts and mutations. With the availability of deeply sequenced RNA-seq data sets and high-resolution variation information, it has been possible to delineate allele-specific binding of transcription factors and allele-specific binding based on the maternal- versus paternal-derived alleles (Rozowsky et al, 2011). As more personal genomes become available, the functional elements can be mapped specifically to the individual's own genetic information. Cytogenetics is being replaced by paired-end sequencing to identify genomic rearrangements and copy number variants at a much higher resolution and throughput.

Nonetheless, significant challenges remain with NGS. These include data processing and storage. In 5 years' time, we are likely to have sequenced more than a million human genomes. Where and how these data will be stored will be a big problem. Another significant challenge is genome interpretation. This includes not only the analysis of genomes for functional elements but the understanding of the significance of variants in individual genomes on human phenotypes and disease. All these add to the still-impractical costs of vast sequencing applications in the clinic. Although sequencing costs have dipped tremendously in recent years, further decrease in costs have to occur before more ambitious applications, such as whole-genome sequencing and longitudinal monitoring, can have a chance in the clinic.

Cost–benefit analyses of sequencing applications in the clinic have to be conducted before actual medical application. Comparison with currently available techniques needs to be done, and a decision made to whether such screens should be made routine or only under exceptional cases. The benefits of sequencing applications in the medical clinic definitely look promising, but much remains to be done in ironing out minute details to make it practical and applicable.

With many people's genomes sequenced, security also becomes an important factor. How will these information be stored, and who will have access to them? Will the individuals know every detail of their genome, or only those pertinent to disease diagnosis or treatment? How can we prevent the possible emergence of ‘genetic discrimination'? Ethical issues will definitely emerge with the commonalization of personal genomes, and these issues need to be resolved before we arrive there.

Our current knowledge and understanding of the human genome still lies largely in the coding regions of the genome. SNPs and SVs that are discovered in non-coding regions are generally dismissed as ‘less important' and ‘not causal'. Although this approach allows us to prioritize and focus resources on the more probable damaging mutations, the effects of non-coding regions in regulation and human disease are becoming more evident. Based on the wealth of non-genic functional regulatory regions obtained from the ENCODE project, RegulomeDB has been developed as a resource for integrating and cross-validating polymorphisms to the regulatory regions (Boyle et al, 2012). Disease-associated SNPs obtained from GWAS studies might point to gene deserts, but could essentially lie in regulatory sites of downstream genes. Often, the SNP that is in linkage disequilibrium with the reported SNP might be more informative (Boyle et al, 2012). It is necessary, in the future, to develop ways to map sequencing data onto currently difficult-to-map regions, such as highly repetitive and low-expressed regions. Sequencing technology is rapidly improving, but the analytical capabilities to understand everything that is being generated by the sequencers is lagging far behind. We need to advance the computational technologies as we progress towards the systemic use of high-throughput sequencing in research and medicine.