{
  "id": "next-generation-sequencing",
  "title": "Next-Generation Sequencing",
  "lang": "en",
  "source_lang": "en",
  "summary": "Next-generation sequencing comprises massively parallel technologies for reading nucleic acids, enabling large-scale studies of genomes, gene activity, biological diversity, and disease.",
  "category": "science",
  "subcategory": "biology",
  "type": "concept",
  "aliases": [
    "NGS",
    "high-throughput sequencing",
    "massively parallel sequencing"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/next-generation-sequencing/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/next-generation-sequencing/",
      "json": "https://aiwiki.page/zh/next-generation-sequencing.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Next-generation sequencing (NGS) is a group of technologies that determine the sequences of large numbers of DNA or RNA molecules in parallel. Unlike conventional DNA sequencing based on individually resolved chain-termination reactions, NGS distributes sequencing across millions or billions of spatially separated templates or individual molecules. The resulting sequence fragments, called reads, are analyzed computationally to reconstruct genomes, measure molecular abundance, identify genetic differences, or characterize complex biological samples.\n\nThe term originally referred chiefly to short-read platforms commercialized in the mid-2000s. Usage subsequently broadened, although some classifications distinguish short-read, “second-generation” sequencing from single-molecule, “third-generation” sequencing. These boundaries are not absolute: parallelization, molecular amplification, read length, and detection chemistry vary independently. In contemporary practice, high-throughput sequencing encompasses both short-read and long-read approaches, selected according to the biological question and the properties of the sample.\n\nNGS transformed genetics and molecular biology by making whole-genome and transcriptome-scale measurements routine rather than exceptional. Its applications extend from inherited disease and cancer research to environmental surveys, crop improvement, infectious-disease surveillance, and ancient genomes. A sequencing result nevertheless depends on more than the instrument: sample collection, library preparation, experimental design, statistical analysis, and interpretation all determine which biological conclusions the data can support."
    },
    {
      "heading": "Historical development and technological scope",
      "text": "The immediate predecessor of NGS was Sanger sequencing, introduced in 1977 by Frederick Sanger and colleagues. Its central mechanism is the incorporation of chain-terminating dideoxynucleotides during DNA synthesis. Automated fluorescent detection and capillary electrophoresis later increased throughput, and this technology supplied most of the sequence used in the Human Genome Project, which began in 1990 and announced its completion in 2003. Although many capillaries could operate together, each generally resolved a separate population of fragments derived from one template. Scaling therefore required substantial replication of laboratory operations and instrumentation.\n\nSeveral developments during the 1990s and early 2000s changed that architecture. Researchers devised methods for immobilizing DNA on surfaces, amplifying individual templates into localized colonies, and detecting nucleotide incorporation at many positions simultaneously. These approaches reduced the need to handle each sequencing reaction separately. Important intellectual and technical contributions came from work on pyrosequencing, massively parallel signature sequencing, polony sequencing, and reversible-terminator chemistry. Shankar Balasubramanian and David Klenerman developed the approach underlying Solexa sequencing, while Jonathan Rothberg and colleagues developed the technology commercialized by 454 Life Sciences.\n\nIn 2005, a publication describing the 454 platform demonstrated sequencing in microfabricated reaction wells at a scale substantially beyond conventional capillary methods. Solexa instruments became commercially available in 2006, and Illumina acquired Solexa in 2007. Applied Biosystems’ SOLiD platform used sequencing by ligation, providing a different route to parallel sequence detection. Semiconductor sequencing, commercialized by Ion Torrent in 2010, detected chemical changes associated with DNA synthesis rather than fluorescent light. Platform competition drove rapid improvements in output and cost, but also produced distinct error profiles and analytical requirements.\n\nSingle-molecule approaches developed alongside these amplified-template technologies. Pacific Biosciences commercialized single-molecule real-time sequencing in 2011. Oxford Nanopore Technologies began its MinION early-access program in 2014, introducing a portable instrument that inferred sequence from changes in electrical current as molecules passed through nanopores. These platforms made much longer reads practical and reduced dependence on amplification, although their accuracy and throughput evolved considerably after initial commercialization.\n\nConsequently, NGS is better understood as a family of measurement systems than as one laboratory technique. Short-read instruments usually emphasize large numbers of accurate, relatively short observations; long-read instruments preserve more information about continuity along individual molecules. Neither read length nor throughput alone defines analytical usefulness. The relevant comparison includes molecular input requirements, systematic errors, turnaround time, usable yield, computational demands, and the ability to resolve the particular biological feature under investigation."
    },
    {
      "heading": "Sample preparation and sequencing libraries",
      "text": "A sequencing library is a collection of nucleic-acid molecules prepared in a form compatible with a sequencing platform. Its construction begins with extraction from material such as blood, cultured cells, tissue, soil, or preserved specimens. Extraction must recover the relevant molecules while limiting inhibitors and degradation. DNA fragment length is especially important for long-read sequencing, whereas RNA workflows must account for RNA’s chemical instability and the large abundance differences among RNA classes. Sample quality cannot be restored simply by generating more reads: material lost or altered before sequencing may leave permanent gaps or biases.\n\nFor many short-read workflows, DNA is fragmented mechanically or enzymatically, repaired at its ends, and joined to synthetic adapter sequences using a DNA ligase. Adapters provide binding sites for sequencing primers, attachment to an instrument surface, or amplification. In tagmentation, a transposase fragments DNA while introducing adapter-associated sequences, combining operations that otherwise require separate steps. Long-read libraries often avoid extensive fragmentation and instead preserve high-molecular-weight DNA. Different platforms require different terminal structures, including circular templates or protein-associated adapters that control passage through a nanopore.\n\nLibraries may be amplified using the polymerase chain reaction (PCR). Amplification increases the amount of material and enriches correctly constructed library molecules, but it can distort relative abundance and preferentially recover fragments with particular lengths or base compositions. Polymerase errors may also become represented in many descendant copies. PCR-free preparation avoids this amplification step but generally requires more input material. In RNA sequencing, reverse transcriptase commonly converts RNA into complementary DNA before library construction; direct RNA sequencing is an alternative available on nanopore platforms.\n\nSample indexes are short, predetermined sequences that identify libraries pooled in the same sequencing run. They should be distinguished from unique molecular identifiers (UMIs), which label original molecules so that amplified descendants can be recognized computationally. Dual indexing uses identifiers associated with both ends of a library molecule and can help identify or reduce some forms of sample misassignment. Neither indexing nor UMIs eliminate every error: index contamination, barcode collisions, damaged templates, and errors introduced before molecular labeling can still affect results.\n\nLibrary preparation may encompass the whole sample or enrich selected regions. Exome sequencing captures principally protein-coding regions, while targeted panels focus on specified genes or loci. Hybrid capture uses complementary probes to recover desired fragments; amplicon sequencing uses targeted PCR. These strategies concentrate sequencing effort but introduce dependence on probe design, primer binding, and capture efficiency. The final library therefore represents a selected and processed version of the original sample, not a perfectly unbiased inventory of all molecules it contained."
    },
    {
      "heading": "Short-read sequencing mechanisms",
      "text": "The most widely used short-read approach is sequencing by synthesis with cyclic nucleotide detection. In an Illumina-type workflow, library fragments bind to a flow-cell surface and undergo localized amplification, producing clusters of related templates. During each sequencing cycle, a DNA polymerase incorporates a labeled, reversibly blocked nucleotide. Imaging identifies the incorporated base, after which the fluorescent label and blocking group are removed so that synthesis can continue. Different instrument generations use different labeling and imaging schemes, but all depend on assigning successive incorporation events to spatially resolved template clusters.\n\nCluster amplification makes each position bright enough for optical detection, but it introduces a synchronization problem. Ideally, all template copies in a cluster incorporate exactly one nucleotide per cycle. In practice, some lag behind or advance relative to others, producing phasing and prephasing. Signals therefore become less clean as sequencing proceeds. Base-calling software models these effects, and read quality commonly changes along the read. High cluster density can increase output until overlapping signals or other instrument constraints begin to reduce usable data.\n\nPaired-end sequencing reads a library fragment from both ends. For example, a \\(2 \\times 150\\)-base run generates two reads of up to 150 bases from each successfully sequenced fragment. The distance between their mapped positions reflects the fragment’s insert length, providing information useful for alignment, small-variant detection, and structural analysis. The pair does not necessarily reveal every base between its ends: if the insert exceeds the combined read lengths, the central interval remains unsequenced. If the reads overlap, their agreement can help validate base calls.\n\nOther short-read platforms detect different consequences of DNA synthesis. Pyrosequencing measures light generated through an enzymatic reaction cascade following release of pyrophosphate during nucleotide incorporation. Semiconductor sequencing detects the associated release of hydrogen ions as a change in local pH. Because a single nucleotide type is presented at a time in these approaches, a run of identical bases must be inferred from signal magnitude. Distinguishing, for example, six consecutive adenines from seven can be more difficult than distinguishing isolated substitutions. Sequencing-by-ligation systems instead use enzymes to join labeled oligonucleotides and infer the underlying sequence from repeated interrogation steps.\n\nShort reads combine high throughput with mature analytical methods and generally strong base-level accuracy, but their limited span creates intrinsic ambiguity. A 150-base sequence originating inside a long repeat may match many genomic locations equally well. Amplification and sequence-dependent chemistry can also make some regions difficult to represent evenly. These limitations explain why excellent average base accuracy does not guarantee complete reconstruction of a genome, accurate measurement of every locus, or reliable detection of every class of genetic change."
    },
    {
      "heading": "Long-read and single-molecule sequencing",
      "text": "Long-read sequencing preserves sequence continuity over thousands to hundreds of thousands of bases, depending on the platform, preparation, and molecule quality. Its main advantage is not simply producing more sequence per observation. A sufficiently long read can connect unique sequences on opposite sides of a repeat, span a complex rearrangement, or distinguish variants carried together on one physical molecule. These capabilities are especially important for highly repetitive regions, duplicated genes, and the organization of individual chromosomes.\n\nIn Pacific Biosciences single-molecule real-time sequencing, a polymerase synthesizes DNA within a tiny optical observation volume called a zero-mode waveguide. Fluorescent signals identify successive nucleotide incorporations. For circular consensus sequencing, adapters create a circular template containing the DNA insert, and the polymerase can traverse that insert repeatedly. The repeated observations are combined into a consensus sequence. Reads meeting high-accuracy criteria are commonly called HiFi reads. Their usable length depends partly on the trade-off between insert size and the number of passes that can be collected.\n\nNanopore sequencing uses a protein pore embedded in a membrane and an applied electrical potential. A motor protein regulates movement of DNA or RNA through the pore. The measured current depends on several nearby nucleotides rather than on one isolated base, so sequence determination requires a model relating the changing signal to local nucleotide combinations. Modern base callers frequently use artificial neural networks. Read length is strongly influenced by the length and integrity of input molecules; ultralong protocols can produce individual reads extending beyond one million bases, although such reads are not representative of every preparation.\n\nSingle-molecule measurements can retain information that conventional amplification-based workflows erase. Polymerase kinetics and nanopore current patterns can reveal some modified bases, including forms of DNA methylation, when suitable models and controls are available. Nanopore platforms can also sequence native RNA, providing access to transcript sequence and some modification-associated signals without first copying the RNA into DNA. Modification inference remains dependent on the chemical context, training data, coverage, and discrimination from other causes of altered signal.\n\nLong reads facilitate haplotype phasing, the assignment of variants to the same parental chromosome copy, and de novo genome assembly, reconstruction without simply mapping all reads to an existing reference. They contributed substantially to the first essentially gapless sequence of a human genome published by the Telomere-to-Telomere Consortium in 2022, which represented a largely homozygous genome and did not include a Y chromosome. Long reads nevertheless retain platform-specific weaknesses: DNA extraction can be demanding, consensus construction may require substantial coverage, and accuracy differs among sequence contexts and operating modes. Combining short and long reads can exploit complementary strengths."
    },
    {
      "heading": "Experimental design, coverage, and statistical power",
      "text": "Sequencing depth describes how extensively a target has been sampled. A common approximation for mean genomic coverage is\n\n\\[C = \\frac{NL}{G},\\]\n\nwhere \\(N\\) is the number of reads, \\(L\\) their mean usable length, and \\(G\\) the target size. For paired-end data, \\(N\\) must be defined consistently as the number of individual reads, or the formula adjusted to count both ends. Thirtyfold coverage of a human haploid reference genome requires roughly 90 billion aligned bases before accounting for losses, nonuniformity, and overlap. It does not mean that every position is observed exactly 30 times.\n\nMean depth differs from breadth of coverage, the fraction of the target observed at or above a specified depth. Under an idealized Poisson model of independent and uniform sampling, the probability that a base receives no reads is\n\n\\[P(0) = e^{-C}.\\]\n\nReal libraries violate this model because sequence composition, amplification, capture efficiency, repetitive regions, and fragment structure influence sampling. Consequently, increasing average depth may yield diminishing returns in systematically inaccessible regions. Reporting coverage thresholds across the target is more informative than reporting a single average.\n\nVariant detection depends on both total depth and the proportion of molecules carrying the variant. At a simple heterozygous site in a diploid sample, the expected alternative-allele fraction is approximately one-half. In tumors, mosaic tissues, pooled organisms, or mixed microbial communities, relevant variants may occupy much smaller fractions. Under ideal independent sampling, a variant at frequency \\(f\\) observed with depth \\(d\\) has probability \\((1-f)^d\\) of receiving no supporting reads. Detection requires more than one observation in many workflows, and actual performance is constrained by error rates, mapping ambiguity, contamination, and library complexity.\n\nLibrary complexity is the number and diversity of distinct original molecules represented. Sequencing an exhausted library more deeply mainly generates additional observations of molecules already sampled. PCR duplicates therefore cannot always be treated as independent evidence. UMIs can improve molecule counting, particularly in low-input and rare-variant assays, but the attainable sensitivity still depends on how many original molecules entered the experiment. No amount of computational refinement can reliably detect an allele absent from the sampled material.\n\nExperimental replication addresses a different source of uncertainty. Sequencing the same library twice improves characterization of technical sampling but does not substitute for independent biological samples. In comparisons of gene expression, additional biological replicates may provide more information about population variability than substantially deeper sequencing of a few specimens. Randomization, balanced processing batches, negative controls, and defined reference materials help separate biological effects from laboratory effects. Study design therefore concerns the distribution of sequencing effort across samples and controls, not merely the number of reads purchased."
    },
    {
      "heading": "Bioinformatics and interpretation of sequence data",
      "text": "Analysis begins by translating instrument signals into base calls and associated quality estimates. A Phred quality score expresses an estimated error probability \\(p\\) as\n\n\\[Q = -10\\log_{10}(p).\\]\n\nThus, \\(Q30\\) corresponds to an estimated base-call error probability of \\(0.001\\). Such scores describe a model’s confidence, not an unconditional guarantee of correctness; their calibration can vary by platform, sequence context, and software version. For some technologies, the original signal can be reprocessed with improved base-calling models, making retention of raw measurements scientifically valuable.\n\nReads and base qualities are commonly stored in FASTQ files. Initial quality assessment examines read length, quality profiles, adapter contamination, unusual sequence composition, and possible sample contamination. Adapter trimming and quality filtering may improve downstream performance, but overly aggressive processing can remove useful evidence or distort the distribution of fragments. Multiplexed datasets also require demultiplexing by sample index, and UMI-containing workflows require explicit handling of molecular identifiers rather than treating them as biological sequence.\n\nIn reference-based analysis, sequence alignment places reads against an existing genome or other sequence collection. Short-read aligners use indexing strategies to search large references efficiently, while long-read aligners accommodate longer sequences and different error patterns. Alignment records are often stored as SAM, BAM, or CRAM files. Mapping quality estimates confidence in the assigned location and is distinct from base quality. A read may have highly accurate bases yet map ambiguously because its sequence occurs repeatedly. A reference that omits relevant sequence can cause misalignment or failure to map.\n\nVariant calling infers differences such as single-nucleotide substitutions, short insertions and deletions, copy-number changes, and larger rearrangements. These classes require different evidence and often different algorithms. Small-variant callers may combine base qualities, read orientation, local assembly, and statistical models of genotype likelihood. Structural-variant analysis uses signals such as split alignments, unusual paired-end spacing, depth changes, or reads spanning rearrangement junctions. Results are commonly represented in Variant Call Format, but a uniform file extension does not imply uniform analytical certainty.\n\nAssembly-based and quantitative analyses follow different paths. Genome assembly reconstructs longer sequences from overlaps or graph relationships among reads, whereas transcript quantification estimates the abundance of genes or transcript isoforms. Statistical interpretation must account for issues such as multiple testing, batch effects, and uncertainty in assignment. Reference resources, software versions, parameters, and filtering rules can materially change results. Reproducibility therefore requires recording the complete analytical workflow, including reference builds and annotations, rather than preserving only a final list of variants or differentially expressed genes."
    },
    {
      "heading": "Transcriptomics, single-cell analysis, and regulatory genomics",
      "text": "RNA sequencing measures the RNA molecules represented in a sample. In a typical experiment, RNA is converted into complementary DNA and sequenced, with counts assigned to genes or transcripts. Because ribosomal RNA often dominates total RNA, library preparation may enrich polyadenylated RNA or remove ribosomal molecules. These choices define the observable transcript population: polyadenylated-RNA enrichment excludes many noncoding and non-polyadenylated transcripts, while depletion-based methods can capture a broader range of RNA species. Strand-specific protocols preserve information about the strand from which an RNA originated.\n\nRead counts are influenced by transcript abundance, sequencing depth, transcript length, and protocol-specific efficiencies. Comparisons between samples therefore require appropriate normalization and statistical models, not a direct comparison of unadjusted counts. Many differential-expression methods use count models that allow variability beyond simple Poisson sampling. Short reads can identify splice junctions and reveal alternative splicing, but assigning reads among transcripts that share exons is difficult. Long-read RNA approaches can observe larger portions of individual transcript structures, although incomplete molecules, coverage, and library artifacts remain important limitations.\n\nSingle-cell sequencing separates molecular measurements by individual cell or nucleus. Some workflows isolate cells in wells; others encapsulate cells and barcoded beads in droplets. Cellular barcodes identify the compartment of origin, while UMIs assist in counting captured RNA molecules. Large single-cell RNA-sequencing datasets have enabled detailed descriptions of cell populations and developmental transitions. Most high-throughput protocols capture only a fraction of each cell’s RNA, and many emphasize one end of a transcript rather than its complete structure.\n\nSingle-cell datasets require special attention to artifacts. Two cells may enter one compartment, producing a doublet; RNA released from damaged cells may contaminate other compartments; and differences in capture efficiency can resemble biological variation. A zero count may reflect true absence, low abundance, or unsuccessful sampling. Computational clustering groups cells with similar expression patterns, but the resulting clusters are analytical constructs that require biological interpretation. Spatial transcriptomic methods retain information about tissue position, with resolution ranging from multicellular capture areas to subcellular measurements depending on the method.\n\nSequencing also measures regulatory features beyond RNA abundance. Chromatin immunoprecipitation followed by sequencing identifies DNA fragments enriched with a chosen protein or histone modification. Assays of accessible chromatin, including ATAC-seq, use transposase insertion to identify regions more accessible under the assay conditions. Bisulfite-based sequencing converts many unmethylated cytosines while leaving 5-methylcytosine relatively protected, allowing modification inference with important caveats concerning DNA damage and related cytosine modifications. Together, these approaches support epigenetics and regulatory genomics, but enrichment, accessibility, and molecular abundance remain measurements of distinct properties rather than interchangeable indicators of gene activity."
    },
    {
      "heading": "Population genomics, microbial communities, and evolutionary research",
      "text": "Large-scale sequencing reveals genetic variation across populations and species. Whole-genome datasets support estimates of genetic diversity, demographic history, relatedness, recombination, and the distribution of alleles. The 1000 Genomes Project, launched in 2008, produced a widely used reference for human genetic variation; its 2015 main analysis included 2,504 individuals from 26 populations. Such resources improved variant interpretation and statistical imputation, while also demonstrating that the usefulness of a reference panel depends on which populations and forms of variation it represents.\n\nPopulation studies connect sequencing with population genetics and genome-wide association studies. Sequencing can discover rare variants and variants poorly represented on genotyping arrays, but association does not by itself establish causation. Population structure, relatedness, environmental differences, and technical batches can create misleading correlations. Read-based measurements also interact with reference bias: molecules more similar to the reference may align more readily than divergent molecules, distorting apparent allele frequencies. Graph-based references and pangenomes aim to represent a broader range of sequences and genomic arrangements than a single linear reference.\n\nIn microbial research, targeted marker sequencing and metagenomics answer different questions. Sequencing the 16S ribosomal RNA gene can profile many bacteria and archaea, while shotgun metagenomics samples DNA across all recoverable genomes in a specimen. Shotgun data can reveal genes and support reconstruction of microbial genomes without cultivation. However, the recovered DNA may originate from living organisms, dead organisms, or extracellular material, so its presence alone does not establish biological activity. RNA-based metatranscriptomics provides complementary information about expressed genes but introduces additional sampling and stability constraints.\n\nCommunity composition is usually measured relatively: increasing the abundance of one organism can reduce the apparent proportion of others even when their absolute abundance remains unchanged. Extraction efficiency, cell-wall structure, marker-gene copy number, and database completeness further influence estimates. Closely related strains may share most of their sequence, complicating assembly and abundance estimation. Negative controls are particularly important in low-biomass samples because reagent contaminants can become a substantial fraction of the reads. Environmental DNA surveys similarly require careful interpretation of transport, persistence, and incomplete reference collections.\n\nSequencing ancient DNA has expanded research on human evolution, extinct organisms, and past ecosystems. Ancient molecules are often short, chemically damaged, and mixed with abundant microbial DNA. Cytosine deamination can produce characteristic apparent substitutions, especially near fragment ends, while modern contamination can overwhelm endogenous material. Authentication therefore uses multiple lines of evidence, including molecular damage patterns and laboratory controls. Comparisons among ancient and present-day genomes have revealed population movements and interbreeding, but genetic patterns do not directly identify language, cultural identity, or archaeological affiliation."
    },
    {
      "heading": "Clinical and public-health applications",
      "text": "Clinical sequencing applies high-throughput methods to defined medical questions under validated laboratory workflows. Targeted panels, exome sequencing, and whole-genome sequencing differ in the range of variants and regions they examine. For suspected inherited disorders, sequencing can identify variants that explain a phenotype, especially when patient data are compared with parental sequences. However, sequencing a gene is not equivalent to establishing that a detected change causes disease. Interpretation combines molecular findings with inheritance, population frequency, functional evidence, and the patient’s documented features.\n\nThe distinction between analytical and clinical performance is fundamental. Analytical validity concerns whether an assay accurately detects a specified variant under specified conditions; clinical validity concerns the relationship between that finding and a condition. Clinical utility concerns consequences for care and outcomes. A variant may be technically real but clinically uninterpretable. The 2015 joint guidelines from the American College of Medical Genetics and Genomics and the Association for Molecular Pathology established widely used categories for interpreting sequence variants in Mendelian disease, including “pathogenic,” “likely pathogenic,” and “variant of uncertain significance.” Classification can change as evidence accumulates.\n\nCancer sequencing examines inherited and acquired alterations, often using tumor tissue together with a matched non-tumor sample. Tumor specimens may contain mixtures of malignant and nonmalignant cells and multiple tumor subclones. Their observed allele fractions depend on tumor purity, local copy number, and clonal composition, not merely the presence or absence of a mutation. Some assays examine circulating cell-free DNA in plasma. Because tumor-derived molecules may constitute only a small fraction of that material, low-frequency error suppression and careful interpretation are central. Variants associated with age-related blood-cell clones can also complicate attribution of plasma findings to a solid tumor.\n\nSequencing-based prenatal screening illustrates the difference between screening and diagnosis. Such tests analyze cell-free DNA in maternal blood, including a placental component, to estimate the likelihood of selected fetal chromosomal conditions. The result is not a direct measurement of every fetal cell or a definitive genome-wide diagnosis. Other clinical applications include identifying infectious agents and investigating drug-resistance determinants. The reliability of pathogen detection depends on specimen type, sampling time, contamination control, genome coverage, and the biological relevance of the detected sequence.\n\nIn public health, pathogen whole-genome sequencing supports outbreak investigation and surveillance of transmission-associated variation. It was used extensively during the COVID-19 pandemic to track SARS-CoV-2 lineages. Closely related genomes can support an epidemiological connection, but they do not by themselves prove direct transmission or its direction. Sampling gaps, mutation rates, within-host diversity, and epidemiological context matter. Across clinical uses, laboratories validate their assays for defined specimen types and variant classes; performance demonstrated for one application does not automatically transfer to another."
    },
    {
      "heading": "Technical limitations, standards, and data governance",
      "text": "NGS errors arise at several levels: damaged starting molecules, contamination, library amplification, signal detection, sequence inference, alignment, and biological interpretation. Some are approximately random, whereas others recur in particular sequence contexts or genomic regions. Repeated observation can suppress independent random errors but may reinforce systematic ones. For example, a consistently misaligned read family can create strong-looking support for a false variant. Conversely, filtering that removes all ambiguous evidence may also exclude real variants in duplicated or repetitive regions.\n\nSequencing benchmarks therefore require more than a single genome-wide accuracy figure. Sensitivity and precision differ among substitutions, small insertions and deletions, repeat expansions, copy-number changes, and complex rearrangements. Benchmark truth sets themselves have boundaries and uncertainties. The Genome in a Bottle program, led by the U.S. National Institute of Standards and Technology with collaborators, develops reference materials and benchmark datasets for human genome analysis. Comparisons are meaningful only when the evaluated regions, variant definitions, matching rules, and confidence intervals are specified. Agreement between two pipelines is not proof that both are correct.\n\nNew references and computational approaches address some longstanding limitations. Human pangenome references represent multiple haplotypes and structural arrangements, reducing dependence on one linear sequence. Improved long reads help resolve repeats and structurally complex regions, while machine learning contributes to base calling, variant detection, and modification analysis. These approaches introduce their own validation requirements. A model trained mainly on one set of instruments, genomes, or sample types may perform differently on another, and software updates can alter both reported variants and quality scores.\n\nGenomic information raises distinctive data privacy concerns because it is persistent, potentially identifying, and informative about biological relatives. Removing names does not necessarily make genomic data anonymous. Research governance therefore addresses consent, controlled access, permitted reuse, security, and procedures for sharing information across institutions and jurisdictions. Clinical and research sequencing may also reveal findings unrelated to the original question, creating policy issues concerning disclosure and participant preferences. The appropriate framework depends on the study, the clinical setting, and applicable law rather than on sequencing technology alone.\n\nRepresentation in reference databases affects both scientific inference and interpretation. Underrepresentation of particular populations can increase uncertainty about whether a rare sequence difference is benign or disease-associated and can reduce the portability of statistical predictions. Infrastructure also shapes access: instruments are only one component of the required system, alongside trained personnel, quality assurance, storage, computing, and interpretation. Long-term stewardship must preserve sufficient metadata to explain how a dataset was produced, while controlling access to sensitive information and documenting the conditions under which future analyses are permitted."
    }
  ],
  "links": [
    "dna",
    "rna",
    "dna-sequencing",
    "genetics",
    "molecular-biology",
    "cancer",
    "human-genome-project",
    "dna-ligase",
    "polymerase-chain-reaction",
    "exome-sequencing",
    "dna-polymerase",
    "nucleotide",
    "ph",
    "enzyme",
    "genome",
    "chromosome",
    "artificial-neural-network",
    "dna-methylation",
    "probability-distribution",
    "gene-expression",
    "rna-sequencing",
    "ribosomal-rna",
    "alternative-splicing",
    "cell",
    "protein",
    "chromatin",
    "epigenetics",
    "genetic-diversity",
    "allele",
    "population-genetics",
    "genome-wide-association-study",
    "causation",
    "pangenome",
    "metagenomics",
    "bacteria",
    "archaea",
    "ancient-dna",
    "human-evolution",
    "diagnosis",
    "public-health",
    "machine-learning",
    "data-privacy"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}