aiwiki.page
English
Science / Biology / genome-assembly

Genome Assembly

Genome assembly reconstructs genomic sequences from sequencing reads, combining computational inference with experimental evidence to resolve sequence, chromosome structure, and inherited variation.

29 keywords10 linked from9 not yet writtenWritten by AI
GenomeDNA SequencingDNAPolymerase Chain…MitochondrionChloroplastAlleleTransposable Ele…Genome Ass…

Genome assembly is the reconstruction of a genome from overlapping or otherwise related fragments of sequence data. Because most DNA sequencing methods read individual molecules or parts of molecules rather than entire chromosomes, their outputs must be combined to infer longer sequences. An assembly can represent a small viral genome, a bacterial chromosome, the multiple chromosomes of a plant or animal, or genomes recovered from a mixed biological community.

Assembly is both a computational process and the sequence resource produced by that process. It establishes the coordinate system on which genes, regulatory regions, and genetic differences are described. Its reliability depends not only on the accuracy of individual sequence reads but also on whether repeated regions, chromosome copies, and long-range relationships have been reconstructed correctly. Modern methods combine sequence-overlap algorithms, graph representations, long-molecule sequencing, and independent physical evidence; their products range from fragmented draft sequences to chromosome-scale, haplotype-resolved reconstructions.

Reconstruction targets and assembly terminology

The basic input to assembly is a collection of reads: sequences of DNA bases reported by a sequencing instrument. A read is an observation of a sampled molecule, not a guarantee of an error-free genomic segment. Reads may contain substitutions, insertions, deletions, residual adapter sequence, or sequence from an unintended organism. Assembly therefore involves inference about both the underlying sequence and the origin and reliability of the observations.

A contig, short for contiguous sequence, is a reconstructed stretch of sequence without assembly gaps. A scaffold orders and orients contigs using additional linking evidence; unresolved intervals are commonly represented by runs of the ambiguous base symbol N. A chromosome-scale scaffold can therefore contain many gaps and should not be confused with an uninterrupted chromosome sequence. Conversely, a complete bacterial chromosome may be represented by a single circular contig. The labels “draft,” “chromosome-level,” and “complete” describe different dimensions of an assembly and require supporting quality information.

De novo assembly reconstructs sequence primarily from relationships among the reads, without using a pre-existing genome as the organizing template. Reference-guided methods instead use an existing sequence to help place reads, order contigs, or reconstruct a closely related genome. These approaches form a spectrum rather than a strict division: a de novo assembly may later be scaffolded against a reference. Reference guidance can be efficient, but it may obscure genuine rearrangements or exclude sequence absent from the reference. Ordinary read alignment and variant calling against a reference do not, by themselves, constitute a de novo assembly.

The intended biological representation is equally important. A haploid assembly seeks one sequence path per locus, whereas a diploid organism carries two homologous chromosome sets that may differ. A “primary assembly” often provides one representative path through most regions, potentially alternating between parental origins, with some alternative sequences stored separately. A fully phased diploid assembly instead attempts to reconstruct both haplotypes consistently. Organisms with polyploidy require more than two chromosome copies to be distinguished.

An assembly is also distinct from genome annotation, which identifies features such as genes and repetitive elements on the reconstructed sequence. An uninterrupted sequence can still have uncertain biological annotation, while a well-annotated draft can remain structurally incomplete. Sequence versions and annotation versions must consequently be tracked separately: changing the assembly can alter coordinates, split or merge predicted genes, and revise interpretations of previously reported variation.

Sequencing data and the information available for assembly

The information content of an assembly dataset depends on read length, error pattern, sampling depth, and the distribution of molecules across the genome. Short-read instruments commonly generate highly accurate reads approximately 100–300 bases long. Paired-end sequencing reads both ends of a DNA fragment, providing an approximate separation and relative orientation between them. This information can bridge a repeat only when the fragment connects sufficiently distinctive sequence on either side; having many short reads does not automatically resolve long repeats.

Long-read technologies observe much longer individual molecules. Pacific Biosciences high-fidelity, or HiFi, reads derive a consensus from repeated observations of a circularized molecule and commonly span roughly 10–25 kilobases. Oxford Nanopore sequencing measures electrical signals as molecules pass through a pore; read lengths depend strongly on DNA preparation, and ultra-long datasets can include reads exceeding 100 kilobases. These technologies have different error profiles and continually changing performance. Consequently, an assembler must account for the actual properties of a dataset rather than assume that all “long reads” behave alike.

Sequencing depth, or coverage, is commonly expressed as

C=NLG,C=\frac{NL}{G},

where NN is the number of reads, LL their mean length, and GG the genome size being used as the denominator. For unequal read lengths, the numerator is the sum of all read lengths. Thus, 90 billion sequenced bases correspond to about 30-fold coverage of a 3-billion-base haploid genome. In a diploid sample, approximately equal sampling of the two homologs would provide about half that depth for each individual haplotype.

Under an idealized model of independent, uniform random sampling, the probability that a genomic position receives no reads is approximately e−Ce^{-C}. Real sequencing departs from this model because base composition, repetitive sequence, DNA damage, amplification, and library preparation affect sampling. Coverage sufficient for base-level consensus may still be insufficient for reliable phasing or for bridging the longest repeats. Extra data can reduce sampling gaps, but cannot supply long-range information that the library never captured.

Sample preparation therefore sets practical limits on reconstruction. High-molecular-weight DNA preserves long connections; fragmented or chemically damaged DNA does not. Polymerase chain reaction can introduce coverage biases and duplicates, while contamination may add complete foreign chromosomes to an otherwise clean dataset. Mitochondrial and chloroplast genomes often occur at much higher copy numbers than nuclear chromosomes, requiring assemblers or downstream analyses to distinguish biological copy number from technical overrepresentation.

Graph models and assembly algorithms

Assembly algorithms convert many local sequence observations into a global reconstruction. A central difficulty is that matching sequence does not necessarily indicate that two reads come from the same genomic location: identical copies of a repeat can produce indistinguishable reads. Assemblers therefore use graph structures that retain alternative connections until sequence or long-range evidence supports a particular traversal. The graph is a model of the data, not a direct image of chromosome organization.

The overlap–layout–consensus approach first identifies reads with significant sequence overlaps, then determines a compatible arrangement, and finally computes a consensus sequence. This framework was central to assemblies based on Sanger sequencing and remains influential in long-read assembly. Comparing every read with every other read would require an impractical number of comparisons for large datasets. Modern implementations use sequence indexes and sparse sequence signatures, including minimizers, to identify promising overlaps before performing more expensive alignments.

A de Bruijn graph instead decomposes reads into short words of length kk, called k-mers. In one common representation, vertices are (k−1)(k-1)-base strings and each observed kk-mer connects its prefix to its suffix. Other implementations use kk-mers as vertices. Repeated observations of the same word share graph structure, allowing enormous short-read datasets to be represented compactly. SOAPdenovo and Velvet were influential short-read assemblers using this general strategy; SPAdes employs multiple graph-building and refinement steps.

Choosing kk involves a trade-off. Small values connect reads despite limited depth but merge more repeated sequence. Larger values distinguish more genomic contexts yet require longer, sufficiently accurate reads and can fragment the graph where coverage is low. Sequencing errors create spurious branches, short dead ends, and alternative paths. Assemblers prune such structures using coverage and topology, but an apparently erroneous branch can also represent a genuine allele or a low-abundance organism.

String graphs provide another representation, typically retaining read-overlap relationships while removing redundant connections. They are useful for long reads because much more context is preserved than in a graph built from short words. Tools such as Canu, Flye, and hifiasm use distinct combinations of overlap detection, repeat representation, error handling, and haplotype separation. Flye emphasizes repeat-graph construction, while hifiasm was developed around accurate long reads and haplotype-aware assembly. Algorithm families overlap in practice: modern systems often combine graph simplification, statistical evidence, and multiple kinds of sequence data rather than implement a single textbook procedure.

Repeats, chromosome copies, and haplotype resolution

Repetitive DNA is the principal obstacle to determining a unique assembly from local sequence information. Repeats include tandem arrays, dispersed copies of transposable elements, ribosomal DNA arrays, and large segmental duplications. If a repeat is longer than the reads or other linking molecules that span it, several genomic arrangements may fit the same observations. Increasing the number of equally short reads cannot necessarily distinguish those alternatives; the ambiguity is informational rather than merely computational.

Different repeat classes create different assembly failures. Nearly identical dispersed copies can be collapsed into one contig, reducing apparent genome size and inflating local read depth. Alternatively, an assembler can connect unique sequence on one side of a repeat to the wrong unique sequence on the other, creating a structural misassembly. Long tandem arrays are especially difficult because reads may reveal the repeating unit without establishing the full array length. Centromeres combine large repetitive arrays with sequence variation that may be useful for reconstruction only when reads are sufficiently long and accurate.

Variation between homologous chromosomes creates a related but distinct challenge. At a heterozygous site, reads support different bases or longer sequence alternatives. In a graph, these differences can form branching paths that rejoin, often called bubbles. Collapsing every bubble loses haplotype information; retaining every branch as an independent genomic segment can falsely duplicate loci. Assemblers must distinguish allelic alternatives from paralogous copies, which occupy different locations in the genome.

Haplotype phasing groups variants according to the chromosome copy on which they occur. A long read can connect several heterozygous sites directly. Trio binning uses sequence information from both parents to assign offspring reads to parental origins before or during assembly. Hi-C contacts and Strand-seq data can provide additional long-range phasing evidence. A switch error occurs when an assembled haplotype changes incorrectly from one homolog to the other; a sequence can have high base accuracy while containing substantial phase-switch errors.

Polyploid genomes introduce further ambiguity because several closely related homologs or homeologs may coexist. Bread wheat, for example, has three related subgenomes and six chromosome sets in its somatic cells. Reliable assembly may require distinguishing both subgenomes and alleles within them. Low heterozygosity can make homologs difficult to separate, whereas high heterozygosity can cause excessive duplication in an unphased assembly. Biological context therefore determines whether similar sequences should be collapsed, represented as alternate haplotypes, or retained as distinct genomic copies.

Scaffolding, gap closure, and sequence polishing

Contigs do not necessarily extend across whole chromosomes even when their local sequence is correct. Scaffolding adds information about how separate contigs relate to one another. Paired reads, long reads, physical maps, and chromosome-contact measurements each supply different kinds of evidence. A scaffold join is an inference supported by these observations; it is not equivalent to directly observing every intervening base. Assembly workflows consequently distinguish ordering and orientation from actual gap closure.

Hi-C is a chromosome-conformation method that captures contacts between DNA segments physically close within the nucleus. When mapped to contigs, contact patterns often support grouping sequences into chromosomes and identifying their likely order. Along a chromosome, contact frequency generally decreases with genomic separation, but the relationship is affected by three-dimensional nuclear organization. Centromeres, homologous chromosomes, and other interacting regions can produce misleading connections. Hi-C scaffolding therefore requires models and inspection rather than a simple rule that frequent contact proves adjacency.

Optical mapping supplies an independent description of long DNA molecules through ordered labeling patterns. An assembled sequence can be converted into a predicted label map and compared with experimentally observed molecules. Agreement supports large-scale structure; discrepancies can reveal inversions, incorrect joins, or missing sequence. Genetic linkage maps provide another ordering signal, based on recombination frequencies across pedigrees or mapping populations. They are particularly valuable for assigning scaffolds to chromosomes, although physical resolution varies greatly because recombination is uneven.

Gap closure attempts to replace unknown intervals with supported sequence. Long reads may bridge both flanking contigs and reveal the intervening DNA, while local reassembly can resolve difficult regions using a selected subset of reads. A scaffold gap whose approximate size is inferred from linking data remains different from a fully sequenced interval. Telomere-to-telomere reconstruction additionally requires uninterrupted sequence from one chromosome end to the other, including repetitive regions that conventional scaffolding may simply span.

Polishing corrects errors in an existing assembly by comparing it with supporting reads. It can refine substitutions and small insertions or deletions, often through iterative alignment and consensus estimation. Polishing is not uniformly beneficial: ambiguous mappings in duplicated regions can overwrite a correct copy with sequence from another copy, and indiscriminate short-read polishing can degrade an accurate long-read assembly. Structural corrections, base corrections, and haplotype corrections are separate operations. Preserving their distinction is essential because a chromosome-scale product may be structurally wrong, while a structurally sound contig may still contain local consensus errors.

Measuring completeness, accuracy, and biological consistency

Assembly quality has several independent dimensions: contiguity, base accuracy, structural correctness, completeness, contamination, and phasing. No single score measures them all. A long contig is useful only if its joins are correct, and an apparently complete set of genes does not establish that repetitive or regulatory DNA has been recovered. Quality assessment therefore combines sequence-based statistics with independent evidence about genome structure and expected biological content.

N50 is a widely used contiguity statistic. Contigs are sorted from longest to shortest, and their lengths are accumulated until at least half the total assembly length is reached; the length of the contig at that point is the N50. For contigs of 8, 4, and 3 megabases, the total is 15 megabases and the N50 is 8 megabases. NG50 uses an independently estimated genome size as the denominator rather than the assembly’s own length. Both measures can improve artificially after an incorrect join, and scaffold N50 can be large despite substantial unresolved sequence.

Consensus accuracy is often reported on a Phred-like quality scale:

QV=−10log⁡10(p),QV=-10\log_{10}(p),

where pp is the estimated per-base error probability. QV40 corresponds to an error probability of 10−410^{-4}, or approximately one error per 10,000 bases. Such estimates depend on the evaluation method. Comparing assembly k-mers with high-quality read k-mers, as implemented in Merqury, can estimate accuracy and completeness without treating another assembled genome as unquestioned truth. However, sequencing biases and differences between sampled haplotypes can affect the result.

BUSCO, introduced in 2015, assesses recovery of sets of near-universal single-copy orthologs appropriate to a lineage. It classifies expected genes as complete and single-copy, complete and duplicated, fragmented, or missing. This is an informative measure of gene-space completeness, not a direct estimate of total genomic completeness. A duplication may reflect uncollapsed haplotypes, genuine gene duplication, contamination, or the assessment model. Similarly, missing genes may indicate assembly gaps, unusual biology, or limitations of gene detection.

Structural validation uses read alignments, contact maps, physical maps, and comparisons with related genomes. QUAST provides assembly statistics and reference-based analyses, including detection of apparent misassemblies, but a genuine difference from the chosen reference can resemble an error. Contamination screening evaluates taxonomic matches, base composition, and read depth together. A credible assessment reports the evidence behind each metric and recognizes the risk of circular validation when the same reads are used both to construct an assembly and to judge its correctness.

Mixed communities and nonstandard biological samples

In metagenomics, sequencing reads originate from many organisms rather than a single isolated genome. The assembly problem becomes a reconstruction of overlapping populations with different abundance levels. A dominant bacterium may receive hundreds of times more coverage than a rare organism, while closely related strains can share most of their sequence. Graph branches may therefore represent sequencing errors, strain variation, repeat copies, or distinct species, often simultaneously.

Metagenomic assemblers such as MEGAHIT and metaSPAdes use strategies adapted to uneven depth and complex graph structure. Their contigs are commonly grouped into candidate genomes through metagenomic binning. Binning uses features such as sequence composition, depth across multiple samples, marker genes, and taxonomic similarity. The resulting metagenome-assembled genomes, or MAGs, are computational reconstructions rather than necessarily complete genomes of individually observed cells. A bin may omit mobile elements or contain sequences from more than one organism.

Closely related strains are especially difficult to separate. If most reads are shared among strains but only a few regions differ, an assembler may produce a mosaic sequence that does not correspond exactly to any single cell. Conversely, retaining every alternative can fragment the reconstruction severely. Repeated genes and horizontally transferred elements weaken the assumption that all sequences with similar composition or abundance belong together. Plasmids may circulate among hosts, and an abundance match alone does not establish which chromosome carried a particular plasmid.

Single-cell sequencing addresses mixtures by isolating individual cells, but the very small DNA input generally requires amplification. Amplification can produce extreme coverage variation, missing regions, and chimeric molecules, making apparently simple single-cell datasets difficult to assemble. Viral assemblies present their own problems, including small highly variable genomes, segmented genomes, and mixtures of related variants. Circularity inferred from an overlapping contig end also requires scrutiny: a repeated terminal sequence can create a false impression of a complete circular molecule.

Ancient DNA and other degraded samples contain short fragments and characteristic damage, sharply limiting the long-range information available for de novo reconstruction. In cancer, a specimen may contain normal cells and multiple tumor subclones with different copy numbers and chromosome rearrangements. Local assembly around a rearrangement can resolve sequence that read mapping alone leaves ambiguous, but reconstructing an entire heterogeneous tumor genome demands explicit attention to cellular mixture and copy number. These cases illustrate why assumptions suitable for a uniform diploid sample cannot simply be transferred to every biological dataset.

Historical development and landmark reconstructions

Genome assembly developed alongside sequencing technology. Frederick Sanger and colleagues published the 5,386-nucleotide genome of bacteriophage phi X174 in 1977, establishing a landmark in complete genome sequencing. Early projects depended heavily on manual interpretation and relatively small datasets. As sequencing expanded, computational overlap detection and consensus construction became indispensable, and the challenge shifted from reading individual fragments to organizing vast collections of them.

In 1995, the genome of Haemophilus influenzae Rd, approximately 1.83 million base pairs long, was published as the first complete genome sequence of a free-living organism. Its whole-genome shotgun strategy showed that large numbers of randomly sampled fragments could be assembled into a bacterial chromosome. During the human genome projects, the publicly funded effort relied heavily on a hierarchical strategy using mapped bacterial artificial chromosome clones, whereas Celera pursued whole-genome shotgun assembly alongside other information and analyses.

The draft human genome publications of 2001 demonstrated the scale of computational reconstruction while also exposing the difficulty of repeats, gaps, and duplicated regions. A finished euchromatic human reference was reported in 2004, but “finished” did not mean that every chromosome was sequenced continuously through its most repetitive regions. The reference remained an invaluable coordinate system while leaving substantial centromeric, acrocentric, and other difficult sequences unresolved.

High-throughput short-read sequencing greatly reduced sequencing cost and enabled large numbers of microbial and eukaryotic projects. However, its read lengths made many repeat structures intrinsically difficult to reconstruct. Long-read technologies changed that balance. Early long reads required substantial error correction, whereas later accurate long reads combined enough length to cross many repeats with sufficient precision to distinguish nearly identical copies. Ultra-long reads supplied complementary links across especially long repetitive regions.

In 2022, the Telomere-to-Telomere Consortium published T2T-CHM13, a reconstruction of roughly 3.055 billion bases spanning all autosomes and the X chromosome of the CHM13 cell line without assembly gaps. CHM13 is largely homozygous and lacks a Y chromosome, making it a powerful but biologically specific reconstruction target rather than a complete representation of human diploid diversity. A complete human Y chromosome sequence from HG002 followed in 2023. These achievements showed how complementary sequencing technologies and intensive validation could recover regions long excluded from standard references.

Reference resources, applications, and representation of diversity

An assembly supplies the physical framework for interpreting genes, regulatory sequences, and structural organization. In comparative genomics, assembled chromosomes reveal conserved gene order, inversions, chromosome fusions, and lineage-specific sequence expansions. Such analyses inform evolutionary research, but they are sensitive to assembly errors: a false join can mimic a rearrangement, and a collapsed duplication can conceal a gene-family expansion. Comparisons are therefore strongest when structural conclusions are supported by independent data.

In agriculture, assemblies of crops and their relatives make it possible to study disease-resistance loci, domestication-associated regions, and variation in large, repetitive genomes. In conservation biology, assemblies provide sequence resources for investigating genetic diversity and the genomic basis of species differences. A reference obtained from one individual, however, does not measure population-wide diversity. Population sampling and analyses of variation remain necessary even when the reference itself is exceptionally complete.

For human genetics, assembly-based approaches are particularly informative for structural variation: insertions, deletions, inversions, duplications, and other changes affecting larger sequence segments. Reads aligned to one linear reference may map ambiguously around complex variants or fail to represent sequence absent from that reference. De novo reconstruction can recover such sequence and place it in a local chromosomal context. Nevertheless, an assembled difference is not automatically functionally important or associated with disease; those interpretations require separate biological and statistical evidence.

A single linear reference also privileges one path through genomic diversity. A pangenome incorporates multiple genomes or haplotypes, often through a graph in which alternative sequence paths are represented explicitly. In 2023, the Human Pangenome Reference Consortium published a draft resource based on phased assemblies from 47 individuals. Such resources expand the sequence available for mapping and comparison, but they introduce additional challenges in coordinate systems, graph alignment, annotation transfer, and consistent identification of variants.

Assemblies are distributed through resources including GenBank and the European Nucleotide Archive, with associated sequence accessions and metadata. FASTA commonly stores contig or scaffold sequences, AGP describes how component sequences form larger scaffolds, and GFA can retain graph relationships that a linear sequence file discards. Reproducible interpretation depends on recording the assembly accession and version, sample origin, sequencing data, software, and curation history. Human-derived assemblies additionally intersect with data privacy, because assembled haplotypes retain identifying and familial genetic information even when conventional personal identifiers have been removed.