Nanopore sequencing is a method for determining the sequence of nucleic acids by measuring changes in electrical current as individual molecules pass through a nanometre-scale opening. It can read DNA and, with a distinct preparation and analysis workflow, RNA. Unlike methods that identify bases through fluorescent labels or repeated cycles of nucleotide incorporation, nanopore sequencing detects the effects of a molecule on an electrical signal. The same signal can contain information about both base sequence and certain chemical modifications.
Its principal advantages are long reads, continuous data production, and the ability to examine native molecules without obligatory amplification. These properties make it useful for reconstructing genomes, identifying large genetic rearrangements, studying RNA transcripts, and investigating DNA methylation. Instruments range from portable devices to high-throughput laboratory systems. Performance depends on the pore, molecular motor, sample preparation, electronics, and computational models; “nanopore sequencing” therefore describes a changing family of technologies rather than a single fixed assay.
Physical principle and molecular recognition
A sequencing nanopore forms a narrow connection between two compartments containing an electrolyte. An applied voltage drives ions through the opening, producing a measurable electric current. When a nucleic acid enters the pore, it alters ion movement and changes the current. The resulting time series is the raw sequencing signal. The detector measures ionic conduction through and around the molecule, not an electrical current flowing along the DNA backbone.
For an idealized cylindrical opening, the resistance of the pore interior is approximately
where is pore length, is radius, and is the conductivity of the solution. Resistance associated with ions approaching and leaving the opening also contributes. A protein pore is neither a uniform cylinder nor a chemically featureless surface, so this expression explains broad geometric dependencies rather than providing a complete sequencing model. Surface charge, local ion concentrations, temperature, and interactions between the nucleic acid and the pore affect the measured signal.
Commercially established systems use engineered protein pores held in electrically insulating membranes. The narrow sensing region interacts with several adjacent nucleotides at once. Consequently, a current level does not normally correspond to one isolated base. It reflects a sequence context, often represented computationally as a k-mer: a succession of bases. As the strand advances, overlapping contexts generate a series of signal changes. Their interpretation requires both knowledge of current amplitudes and an estimate of how the molecule moved.
Uncontrolled passage is generally too fast and irregular for reliable sequence identification. A motor enzyme attached to the sequencing complex regulates movement, advancing a strand through the sensing region in small steps. Motor activity can involve consumption of ATP. The motion is not perfectly uniform: pauses, variable dwell times, and occasional backward steps complicate the relationship between elapsed time and sequence position. The motor and pore must therefore be treated as a coupled molecular measurement system.
Biological pores are not the only possible sensors. Solid-state nanopores can be fabricated in materials such as silicon nitride, while atomically thin membranes have been investigated to shorten the sensing region. These approaches offer possibilities for robust manufacture and device integration, but controlling movement and distinguishing chemically similar bases remain difficult. Many solid-state nanopore experiments measure molecular length, folding, binding, or translocation rather than reading a complete nucleotide sequence. Detecting a molecule passing through a nanopore is thus distinct from demonstrating practical nanopore sequencing.
Development from single-molecule experiments to sequencing platforms
Nanopore sequencing emerged from the convergence of membrane biophysics, single-molecule measurements, and DNA sequencing. Research on biological ion channels had already shown that currents through individual pores could be measured electrically. The sequencing proposal extended this principle: if different parts of a nucleic acid produced distinguishable current changes, a strand might be read as it moved through a pore.
A foundational experiment was published in 1996 by John Kasianowicz, Eric Brandin, Daniel Branton, and David Deamer. It showed that single-stranded nucleic acids could cause characteristic current blockades while traversing an alpha-hemolysin channel. The work established a physical basis for detecting individual polymers electrically. It did not yet provide general base-by-base sequencing: passage was rapid, and the signal reflected multiple aspects of a molecule’s interaction with the pore.
Subsequent research addressed three linked problems: producing a sufficiently informative sensing region, controlling strand movement, and translating noisy current traces into sequence. Alpha-hemolysin was an important early experimental pore. Later studies used other proteins, including the mycobacterial porin MspA, whose constriction offered useful spatial sensitivity. Experiments coupling pores to DNA polymerases or other motors demonstrated controlled stepping and sequence-dependent current measurements. Protein engineering subsequently altered pore geometry, charge, stability, and compatibility with motor enzymes.
Oxford Nanopore Technologies began an early-access programme for its portable MinION instrument in 2014. The device made nanopore sequencing accessible outside specialist membrane-biophysics laboratories and supported a growing community of users and software developers. Early datasets had substantial error rates, but they demonstrated long-read sequencing on a compact instrument connected to a computer. Later changes in pores, motor proteins, library chemistry, electronics, and signal interpretation improved practical performance.
The platform expanded to include benchtop and higher-throughput instruments, notably GridION and PromethION. Sequencing takes place in replaceable flow cells containing arrays of individually monitored channels. Active pores generate data in parallel, and molecules enter asynchronously rather than following a shared sequence of imaging cycles. Throughput therefore depends on how many channels are active, how effectively they are occupied, and how long they remain productive.
Several demonstrations helped establish the technology’s biological value. Portable sequencing supported Ebola surveillance in West Africa during the 2014–2016 epidemic. In 2018, a nanopore-based human genome study demonstrated the value of ultra-long reads for assembly and difficult genomic regions. Nanopore data also contributed, alongside other technologies, to the Telomere-to-Telomere Consortium’s complete human reference sequence published in 2022. These achievements reflect both improvements in instruments and advances in sample preparation and computational analysis; historical results cannot be attributed to pore chemistry alone.
Sample preparation, libraries, and read length
The physical condition of the input material strongly determines the information a nanopore experiment can recover. For long-read DNA sequencing, extraction aims to preserve large molecules while removing proteins, salts, solvents, and other substances that interfere with library preparation or pore operation. Mechanical handling can break DNA, so vigorous mixing and repeated manipulation may reduce read length. A sequencer cannot reconstruct a physically continuous molecule simply because the original chromosome was long.
Nanopore libraries usually contain sequencing adapters that support capture and engagement with the motor–pore complex. In ligation-based workflows, DNA ends are prepared and adapters are attached using enzymes that include DNA ligase. Rapid workflows can use transposase-mediated fragmentation and adapter attachment, trading some control over fragment length for simpler preparation. The details vary across kits and applications. Fragmentation is therefore optional in some workflows rather than an inherent requirement of the detection principle.
Amplification-free sequencing preserves information that copying can erase. Polymerase chain reaction generally replaces native modification patterns with those of newly synthesized DNA and may introduce representation biases. Nevertheless, amplified libraries remain useful when the target is scarce or a particular region is being studied. Amplicon sequencing, whole-genome sequencing, and direct RNA sequencing are different experimental designs, even when their products are analysed on related instruments.
Read lengths span a broad distribution rather than clustering around a single instrument-defined value. Routine genomic libraries can yield reads tens of kilobases long, while carefully prepared ultra-long libraries can produce reads exceeding 100 kilobases and occasional reads longer than a megabase. Such exceptional molecules do not describe the typical read in every dataset. A common length statistic is read N50: after reads are ordered by length, it is the length at which reads of that size or longer account for half of all sequenced bases. N50 is not the arithmetic mean and does not measure sequence accuracy.
DNA quantity and molecular length jointly affect loading. A fixed mass of very long DNA contains fewer molecules than the same mass of short DNA, and long molecules can be more difficult to manipulate and distribute. Insufficient loading leaves pores idle; contaminants, problematic molecules, or unsuitable library components can shorten productive sequencing time. Input mass alone therefore does not predict output.
Samples can be multiplexed by adding identifiable barcode sequences. Reads are assigned to samples computationally, allowing several libraries to share a flow cell. Barcode errors, contamination, and ambiguous assignments are especially consequential when a low-abundance organism or rare variant is being sought. Appropriate controls distinguish genuine biological signals from material introduced during collection, extraction, amplification, or library preparation. For mixed samples, extraction itself can bias representation because different cells and organisms release nucleic acids with different efficiencies.
Basecalling, error patterns, and accuracy assessment
Basecalling converts the raw current trace into a nucleotide sequence. Earlier approaches often divided the trace into events and used probabilistic models to infer transitions between sequence contexts. Modern basecallers predominantly use artificial neural networks trained on paired signal and sequence data. They must infer both which bases are present and how many bases passed during each interval, despite irregular molecular motion and variation between pores.
The computational problem is not equivalent to classifying independent current measurements. Adjacent observations are correlated, several bases influence the sensing region, and the signal may contain stalls or transient disturbances. Models therefore use temporal context. Their performance depends on training data, model architecture, computational precision, and compatibility with the relevant pore and library chemistry. Reprocessing stored raw signals with an improved model can change sequence calls without repeating the laboratory experiment.
Errors include substitutions, insertions, and deletions. Homopolymers—runs of the same base—have historically been difficult because extending a repeated sequence may change its duration more than its characteristic current level. Repeated motifs, modified bases, and unusual sequence contexts can also produce systematic errors. Improvements in pore design and signal models have reduced many of these problems, but residual error is not uniformly distributed across a genome.
Sequence quality is often expressed on a Phred scale:
where is an estimated probability of an incorrect base call. Thus, corresponds to an estimated error probability of 1%, to 0.1%, and to 0.01%. Reported qualities are model estimates and require empirical calibration. A mean read quality, a median alignment identity, and an assembled genome’s consensus quality are different quantities and should not be treated as interchangeable.
A simplex read derives from one strand. In duplex sequencing, complementary strands from the same DNA molecule are both read and their evidence is combined. This can substantially improve accuracy because the two observations constrain the same underlying sequence. It also requires successful recovery and pairing of both strands; not every molecule generates a usable duplex result. Duplex accuracy and duplex yield therefore describe separate aspects of performance.
Consensus from multiple molecules provides another route to accuracy, whether in an assembly or at a particular genomic position. Random errors can average out as coverage increases, but correlated errors may persist. Evaluation consequently requires more than a single headline percentage. Useful measurements include insertion and deletion rates, accuracy in repetitive regions, variant precision and recall, strand balance, and performance at different coverage levels. Comparisons are meaningful only when they specify chemistry, basecaller, sample, filtering criteria, and the reference or truth set used to judge the result.
Genome assembly and genetic variation
Long reads help reconstruct genomes because they can span repeated sequences that shorter fragments cannot place uniquely. In genome assembly, overlapping reads are combined into longer contiguous sequences. If a read crosses an entire repeat and reaches distinctive sequence on both sides, it can establish which neighbouring regions belong together. This is especially valuable in genomes containing long duplications, transposable elements, and repeated gene families.
Read length does not by itself guarantee a complete assembly. Some repeats exceed available read lengths, highly similar copies may remain difficult to distinguish, and uneven coverage can leave gaps. An assembler must also handle sequencing errors and distinguish true biological variation from contradictory observations. Overlap-based approaches and graphs constructed from sequence matches are adapted to long-read data, while subsequent polishing uses read evidence to improve the assembled sequence.
Nanopore reads are particularly useful for structural variation, commonly defined as genomic changes of roughly 50 base pairs or more. These include insertions, deletions, inversions, duplications, and rearrangements joining previously separated regions. A long read may directly cross a breakpoint or contain an inserted sequence absent from the reference. Multiple reads with compatible alignments provide stronger evidence than a single unusual alignment, which might reflect a chimeric molecule or mapping error.
Long molecules also support haplotype phasing: determining which variants lie on the same inherited chromosome. A read spanning several heterozygous sites links their alleles physically. Overlapping links can extend phased blocks across larger regions. This matters when two variants affect the same gene, because their biological interpretation can differ depending on whether they occupy the same chromosome copy or opposite copies. Long reads can also connect sequence variants with allele-specific modification patterns.
Difficult regions include segmental duplications, repeat expansions, centromeres, and telomeres. Reads that span a repeat expansion and its unique flanks can help estimate repeat length and characterize interruptions within the repeat. However, very long or compositionally difficult expansions may still be underrepresented, and a distribution of observed lengths may reflect both biological mosaicism and technical effects. Accurate interpretation depends on the locus, preparation method, and analysis procedure.
The complete CHM13 human reference published in 2022 illustrates the complementarity of sequencing technologies. Ultra-long nanopore reads helped resolve large repetitive structures, while highly accurate long reads and other evidence supported sequence reconstruction and validation. The result was not a demonstration that any single read type eliminates all assembly problems. For ordinary diploid samples, separating parental chromosome copies adds complexity absent from a largely homozygous reference sample. Population-scale work also has to represent sequences missing from a single linear reference, linking nanopore sequencing to the development of pangenomes and graph-based analysis.
Native DNA modifications and direct RNA sequencing
Nanopore signals respond to molecular chemistry as well as canonical base identity. A modified nucleotide can alter the current produced by the sequence context surrounding it. This makes it possible to infer certain modifications without first converting them into sequence changes through chemical treatment. The best-established example is DNA methylation, particularly 5-methylcytosine in appropriate sequence contexts.
Modification calling is a statistical inference from signal, not a direct visual identification of a chemical group. Models are trained using examples with known or experimentally controlled modification states. They may estimate a probability for a modification at each covered position. Aggregating calls across molecules can estimate modification frequency, while retaining individual reads allows analysis of coordinated patterns along a DNA molecule. An observed intermediate frequency can arise from mixtures of cells, alleles, or molecular states rather than partial modification of every molecule.
Sequence and modification calls interact. An unmodelled modification may affect ordinary basecalling, while an incorrect sequence alignment may distort modification placement. Distinguishing chemically similar modifications can require specialized models and suitable controls. Performance established for one modification, organism, or sequence context cannot automatically be extended to every chemical mark. Studies of epigenetics therefore separate the technical ability to classify a signal from the biological interpretation of the resulting pattern.
Direct RNA sequencing passes an RNA strand through the pore rather than reading only a DNA copy. In widely used polyadenylated-RNA workflows, adapter attachment begins at the RNA’s 3′ end, and the molecule is read toward its 5′ end. A complementary DNA strand may be synthesized as part of preparation, but the strand generating the sequence-dependent pore signal is RNA. This distinguishes direct RNA sequencing from nanopore sequencing of complementary DNA.
Long RNA reads can connect distant exons and characterize transcript isoforms, linking the method to RNA splicing and alternative splicing. They can reveal which splice junctions occur together within one molecule rather than inferring transcript structure entirely from short fragments. Signal near the beginning of the read can also provide information about poly(A)-tail length. These measurements are useful for examining transcript processing, stability, and regulation, although each depends on specialized analytical models.
Direct RNA sequencing has distinctive limitations. RNA degradation truncates molecules, sequencing can terminate before the 5′ end, and library designs may preferentially capture particular RNA classes. Standard poly(A)-oriented workflows do not represent all cellular RNAs equally. Throughput, error patterns, and full-length recovery differ from those of DNA sequencing. RNA modifications can influence the signal, but reliable identification requires modification-specific validation and attention to sequence context. Complementary methods such as short-read RNA sequencing, biochemical assays, and targeted experiments remain important for separating genuine transcript features from technical artefacts.
Real-time analysis, selective sequencing, and computation
Nanopore data are produced continuously as molecules move through individual channels. Analysis can begin before a sequencing run ends, allowing early estimates of sample composition, coverage, and target detection. This streaming architecture differs from workflows in which most usable sequence becomes available only after a complete instrument run. The time required to obtain an answer nevertheless includes sample preparation, computation, and the amount of evidence needed for a reliable decision.
Adaptive sampling uses this real-time access to select molecules during sequencing. Software examines an initial segment of a read, classifies it against a target definition, and decides whether sequencing should continue. If the molecule is unwanted, the instrument can reverse the voltage to eject it, making the pore available for another molecule. Depending on the application, classification may use a preliminary basecalled sequence or features derived more directly from the current signal.
The method can enrich selected genomic regions or reduce sequencing of abundant background material without adding a target-specific capture reaction. Its effectiveness depends on decision speed, fragment length, target abundance, reference quality, and how quickly rejected molecules are replaced. Every rejected molecule has already consumed some pore time. Short fragments may be nearly finished by the time classification occurs, while extremely rare targets remain limited by how often they enter pores.
Adaptive sampling differs from targeted laboratory enrichment. For example, CRISPR-associated nucleases can cut DNA near selected regions so that library preparation preferentially admits desired fragments. Amplification-based targeting increases selected sequences before sequencing but generally does not preserve native modification information in the amplified product. These approaches manipulate different stages of the workflow and have different biases. Software selection cannot recover a target molecule that extraction or preparation has removed.
Computational requirements range from modest monitoring to intensive real-time basecalling and mapping. More accurate models generally require more computation, often supplied by graphics processing units or integrated instrument hardware. If processing cannot keep pace with acquisition, raw signals may be stored for later analysis, but that delays applications requiring immediate decisions. Signal files, basecalled reads, alignments, and derived modification calls represent different data layers, each with its own storage and reproducibility requirements.
Real-time stopping rules also require statistical care. Detecting a few reads from an organism is not equivalent to obtaining adequate genome coverage or demonstrating the absence of another organism. Relevant quantities include independent supporting molecules, breadth of coverage, contamination controls, and uncertainty in classification. Similarly, reaching a nominal average depth may leave difficult regions poorly covered. Reproducible analysis records the pore chemistry, library method, basecaller model, software versions, reference databases, and filtering rules. Because computational interpretation changes over time, retaining raw signals can preserve analytical options unavailable from sequence files alone.
Biomedical, public-health, and environmental applications
In human genetics, nanopore sequencing is used to investigate structural changes, repeat expansions, complex loci, and modification patterns. A single long molecule can link a rearrangement to nearby sequence variants or place a repeat expansion within its genomic context. Research workflows also explore combined sequence and methylation analysis in cancer, where tumour samples may contain mixtures of normal and malignant cells with different copy numbers and molecular states.
Clinical implementation requires validation for a defined purpose. Performance in detecting large rearrangements does not establish equivalent sensitivity for single-base substitutions, low-frequency variants, or every genomic region. Validation addresses specimen type, minimum coverage, reportable variant classes, reproducibility, contamination, and comparison with established methods. Rapid instrument output is only one component of diagnostic turnaround; sample processing, interpretation, confirmation where required, and reporting can be substantial parts of the workflow.
In infectious-disease work, nanopore sequencing can identify organisms, reconstruct pathogen genomes, and detect known antimicrobial-resistance determinants. Long reads help connect resistance genes to plasmids or chromosomes and can resolve repeated mobile elements. However, identifying a resistance-associated gene is not always equivalent to demonstrating a resistant phenotype. Expression, gene integrity, genetic background, and resistance mechanisms absent from a database can affect the relationship between sequence and laboratory susceptibility measurements.
Portable sequencing has supported genomic surveillance during outbreaks, including Ebola and Zika. The small size of some instruments permits deployment near sample collection sites, while continuous analysis can provide early information on transmission-related variation. Field use still requires extraction equipment, reagents, contamination control, power, and computing. Transportability reduces some infrastructure demands but does not remove the need for trained personnel or appropriate biological containment.
Metagenomics applies sequencing to mixed genetic material from samples such as soil, water, food, or clinical specimens. Long reads can improve assembly of microbial genomes and connect functional genes with their surrounding genomic sequence. They can also help distinguish closely related organisms when enough informative sequence is present. Sample preparation remains a major source of bias: organisms differ in cell-wall structure, genome abundance, and susceptibility to DNA degradation, while host material may dominate a clinical specimen.
Environmental applications include studying microbial communities, tracking organisms in water systems, and characterizing genetic diversity in poorly sampled locations. Their evidential limits depend on the question. Detecting DNA does not necessarily establish that an organism is alive, locally reproducing, or present at a concentration proportional to its read count. Conversely, failure to detect a sequence may reflect sampling depth, extraction bias, or incomplete reference databases. Human-associated datasets introduce additional obligations concerning consent and data privacy, because long reads can contain identifiable genetic information and linked molecular features even when the immediate objective is pathogen detection.