# Genome Assembly and Annotation: Methods and Pitfalls

## Introduction to Genome Assembly and Annotation

### What is Genome Assembly?

Genome assembly is the computational process of reconstructing a complete or near-complete genome sequence from the fragmented DNA reads produced by sequencing instruments. A sequencing run does not produce a genome; it produces millions to billions of short or long nucleotide sequences that must be pieced together like a jigsaw puzzle of unknown dimensions. The assembly process involves detecting overlaps between reads, constructing contigs (contiguous consensus sequences), ordering and orienting those contigs into scaffolds, and ultimately generating chromosome-level sequences where possible.

The fundamental challenge is that sequencing reads are far shorter than the chromosomes they derive from. A typical Illumina read is 150 base pairs (bp), while a human chromosome is approximately 250 million bp. Even the longest PacBio reads, which can exceed 25 kilobases (kb), represent only a tiny fraction of a chromosome. Assembly algorithms must therefore infer the original genomic order from local sequence overlaps, a problem complicated by repetitive DNA, sequencing errors, and uneven coverage.

### What is Genome Annotation?

Genome annotation is the process of attaching biological meaning to the assembled sequence. It involves identifying the locations of genes, their exon-intron structures, regulatory elements, repetitive elements, and other functional features, and then assigning putative functions to those features. Structural annotation defines the coordinates of genes and their transcripts; functional annotation assigns biological roles—molecular function, cellular location, participation in pathways—to the predicted gene products.

Annotation is not a single step but an iterative process that integrates multiple lines of evidence. A gene model might be supported by ab initio prediction algorithms, by homology to known genes in related species, and by transcriptome data showing which regions are actually expressed. The quality of annotation depends critically on the quality of the underlying assembly: a misassembled genome will produce misannotated genes, and fragmented assemblies often result in truncated or merged gene models.

### The Genome Project Workflow

A typical genome project proceeds through several stages. First, DNA is extracted and prepared for sequencing. Second, sequencing is performed, producing raw reads. Third, reads are quality-filtered and assembled into contigs and scaffolds. Fourth, the assembly is evaluated for completeness and correctness. Fifth, structural annotation identifies gene models. Sixth, functional annotation assigns biological roles. Finally, the annotated genome is deposited in public databases.

Each stage has its own failure modes, and errors propagate downstream. A contaminated DNA sample will produce a chimeric assembly. A highly repetitive genome assembled only from short reads will be fragmented into thousands of contigs. An annotation pipeline that relies solely on ab initio prediction will overpredict genes in a genome with unusual codon usage. Understanding the methods and their pitfalls is essential for producing a genome resource that the community can trust.

## Sequencing Technologies and Data Types

### Short-Read Sequencing (Illumina)

[Illumina sequencing](/knowledge/diagnostics/molecular/illumina-sequencing-principle-chemistry-and-workflow) by synthesis dominates the short-read market. Genomic DNA is fragmented to 300–500 bp, adapter-ligated, and amplified on a flow cell. Sequencing proceeds by cyclic reversible termination: fluorescently labeled nucleotides are added one at a time, imaged, and cleaved. Read lengths are typically 150 bp paired-end, meaning both ends of each fragment are sequenced, yielding two reads separated by a known insert size.

The key characteristics of Illumina data are high accuracy (Q30 or better, meaning less than 1 error per 1000 bases) and massive throughput (hundreds of gigabases per run). However, the short read length means that reads cannot span most repetitive elements. A read of 150 bp that falls entirely within a repeat family will map ambiguously, and the assembler cannot determine which copy it came from. This limitation is the primary reason why short-read-only assemblies of complex genomes are fragmented at repeat boundaries.

Paired-end information partially mitigates this problem. If the two reads of a pair flank a repeat, the known insert size allows the assembler to bridge the repeat even though the repeat itself is not sequenced. Mate-pair libraries with insert sizes of 2–10 kb provide longer-range information, though they require more input DNA and have higher error rates at the junction.

### [Long-Read Sequencing](/knowledge/bioinformatics/long-read-sequencing-technologies-pacbio-and-oxford-nanopore) (PacBio, Oxford Nanopore)

Pacific Biosciences (PacBio) single-molecule real-time (SMRT) sequencing produces reads averaging 10–25 kb, with some exceeding 100 kb. The platform works by immobilizing a single DNA polymerase molecule in a zero-mode waveguide and detecting fluorescence as nucleotides are incorporated. The original continuous long-read (CLR) mode had high error rates (approximately 10–15% per base), but the newer circular consensus sequencing (CCS) mode, also called HiFi, sequences the same molecule multiple times to produce reads of 10–20 kb with accuracy exceeding 99.9%.

Oxford Nanopore Technologies (ONT) sequences by threading DNA through a protein nanopore embedded in a membrane. As nucleotides pass through the pore, they modulate an ionic current, and the pattern of current changes is decoded into a sequence. Nanopore reads can be extremely long—the current record exceeds 4 Mb—but raw accuracy is lower than Illumina, typically 95–99% for the latest chemistry. The error profile is also different: errors are predominantly insertions and deletions rather than substitutions, which complicates alignment and assembly.

Long reads solve the repeat-bridging problem because they can span entire repeat units. A 15 kb PacBio read can cross a 5 kb tandem repeat array, providing unambiguous sequence for the flanking regions and allowing the assembler to resolve the copy number and order of the repeats. For this reason, long-read sequencing has become the standard for de novo assembly of complex genomes.

### Linked-Read and Hi-C Technologies

Linked-read technologies, such as 10x Genomics Chromium, barcode fragments derived from the same high-molecular-weight DNA molecule. After sequencing, reads sharing a barcode are known to originate from the same original molecule, providing long-range information without long reads. Linked reads can phase variants and scaffold contigs, but they do not resolve repeats as effectively as true long reads.

Hi-C (high-throughput chromosome conformation capture) measures physical proximity between genomic loci in the nucleus. DNA is cross-linked, digested, and ligated such that fragments that are close in 3D space become joined. Sequencing the ligation junctions reveals which genomic regions are physically associated. In practice, Hi-C data provide a proximity map that can order and orient contigs into chromosome-scale scaffolds, even for genomes assembled from short reads. The technique is now standard for producing chromosome-level assemblies: contigs are placed along chromosomes based on the assumption that intra-chromosomal contacts are more frequent than inter-chromosomal contacts.

## Assembly Algorithms and Strategies

### Overlap-Layout-Consensus (OLC)

The overlap-layout-consensus (OLC) approach was the first assembly paradigm, developed for Sanger sequencing reads of 500–1000 bp. The algorithm proceeds in three stages. First, all pairs of reads are compared to find overlaps—regions where the suffix of one read matches the prefix of another. Second, a graph is constructed in which nodes are reads and edges are overlaps; the layout stage identifies a path through the graph that represents the genomic sequence. Third, the consensus stage aligns all reads along the path and calls a single consensus base at each position.

OLC scales poorly with read number because all-versus-all comparison is O(n²) in the number of reads. For a mammalian genome sequenced at 30× coverage with 150 bp reads, that means comparing hundreds of millions of read pairs, which is computationally prohibitive. However, OLC works well for long reads, where the number of reads is smaller and each read contains enough information to anchor overlaps. Modern long-read assemblers such as Canu, Flye, and Raven use variations of OLC or related graph-based approaches optimized for the error profiles of PacBio and Nanopore data.

### De Bruijn Graph Assembly

De Bruijn graph assemblers, including Velvet, ABySS, and SPAdes, were developed for short reads. Instead of using reads directly, the assembler decomposes each read into k-mers—subsequences of length k. A graph is constructed in which nodes are k-mers and edges connect k-mers that overlap by k−1 bases. The genome corresponds to a path through this graph, and the assembly problem becomes finding that path.

The de Bruijn approach is efficient because it collapses redundant information: a k-mer present in many reads is represented once. However, the choice of k is critical. Small k values produce graphs that are well-connected but ambiguous in repetitive regions; large k values produce fragmented graphs that may not connect across low-complexity sequence. Many assemblers use multiple k values and merge the results.

The primary weakness of de Bruijn graphs is that repeats longer than k create branches in the graph. If a repeat is longer than k, the graph cannot determine which incoming path connects to which outgoing path, and the assembly breaks at that point. Increasing k helps, but k cannot exceed read length. This is why short-read assemblies of repeat-rich genomes are inherently fragmented.

### Hybrid Assembly and Scaffolding

Hybrid assembly combines short and long reads to exploit the strengths of both. The most common strategy is to use long reads for the initial assembly, producing highly contiguous contigs, and then use short reads to polish the assembly—correcting the residual errors in the long-read consensus. This approach, implemented in pipelines such as MaSuRCA and Unicycler, produces assemblies that are both contiguous and accurate.

Scaffolding is the process of ordering and orienting contigs into larger structures. The scaffold is a superscaffold: contigs connected by gaps of known size but unknown sequence. Scaffolding uses paired-end and mate-pair reads (which link two contigs), linked-read barcodes, or Hi-C contact maps. The gaps between contigs are filled with Ns in the assembly, and gap-closing algorithms attempt to fill them using reads that span the gap.

For example, the Hi-C scaffolding algorithm implemented in tools like HiRise and SALSA2 works as follows: (1) align Hi-C reads to the contigs; (2) build a contact matrix of read pairs between contigs; (3) use the contact frequencies to infer the order and orientation of contigs along chromosomes; (4) break contigs at positions where the contact pattern suggests a misassembly. The result is a chromosome-scale assembly with contigs placed in their correct genomic context.

## Evaluating Assembly Quality

### Contiguity Metrics (N50, L50)

The most widely reported assembly metric is N50, defined as the length L such that 50% of the assembled bases are contained in contigs or scaffolds of length ≥ L. More precisely, if all contigs are sorted by length from longest to shortest, N50 is the length of the contig at which the cumulative sum of lengths first reaches 50% of the total assembly size. L50 is the number of contigs needed to reach that 50% threshold. A higher N50 indicates a more contiguous assembly; a lower L50 indicates fewer, larger pieces.

N50 is a useful but incomplete metric. It says nothing about correctness: a chimeric assembly that incorrectly joins unrelated sequences can have a high N50 while being biologically wrong. It also says nothing about completeness: an assembly that misses entire chromosomes can still have a high N50. N50 should therefore be reported alongside other metrics, and it should be computed separately for contigs and scaffolds.

### Completeness Assessment with BUSCO

The Benchmarking Universal Single-Copy Orthologs (BUSCO) approach assesses assembly completeness by searching for a set of genes that are expected to be present as single copies in the target lineage. The BUSCO database contains orthologs from lineages such as vertebrates, arthropods, fungi, and bacteria. The assembly is searched for each BUSCO gene, and each gene is classified as complete (present and full-length), fragmented (present but partial), or missing.

BUSCO provides a biologically meaningful measure of completeness that complements N50. A genome assembly with a high N50 but missing many BUSCO genes is likely missing real genomic content, either because of assembly gaps or because the DNA was not sequenced. Conversely, an assembly with a low N50 but complete BUSCO genes may be fragmented but content-complete. For example, a typical high-quality insect genome assembly should recover more than 95% of the insect BUSCO set as complete.

### Assembly Validation Methods

Several methods validate assembly correctness beyond contiguity and completeness. One approach is to compare the assembly to a reference genome from a closely related species, using whole-genome alignment to identify structural errors. Another is to check the consistency of paired-end read alignments: if reads that should map to adjacent positions map to distant locations, the assembly is likely misassembled at that point.

The software QUAST (QUality ASsembly Tool) computes a suite of metrics, including N50, number of misassemblies, and coverage of a reference genome. REAPR and FRCbam use read mapping to detect assembly errors. For Hi-C-based assemblies, the contact map itself is a validation tool: a correct assembly should show a smooth diagonal of high contact frequency, with no abrupt discontinuities that would indicate misjoins.

## Repeat Resolution and Complex Regions

### Types of Repetitive DNA

Repetitive DNA constitutes a substantial fraction of most eukaryotic genomes. In the human genome, approximately 50% of the sequence is repetitive. The major classes are:

- **Tandem repeats**: arrays of identical or nearly identical units arranged head-to-tail. These include microsatellites (1–6 bp units), minisatellites (10–100 bp units), and satellite DNA (100 bp to several kb units). Ribosomal RNA genes are present in large tandem arrays of ~43 kb units.
- **Interspersed repeats**: [transposable elements](/knowledge/molecular-biology/transposable-element) that are dispersed throughout the genome. These include retrotransposons (LINEs, SINEs, LTR retrotransposons) and DNA transposons. The human genome contains approximately 500,000 LINE-1 elements and over 1 million Alu elements.
- **Segmental duplications**: blocks of 1–200 kb that are duplicated to different genomic locations, often with >90% sequence identity. These are particularly problematic for assembly because they are long and highly similar.

### Long-Read Advantages for Repeats

Short reads cannot span most repeats, so assemblies from short reads break at repeat boundaries. The result is that each copy of a repeat is collapsed into a single consensus sequence, and the flanking unique sequence is left unassembled or misassembled. Long reads solve this problem by spanning entire repeat units.

Consider a tandem array of 5 kb units. A 15 kb PacBio read can span two full units plus flanking sequence, allowing the assembler to determine the exact copy number and the sequence of each unit. For interspersed repeats, a long read that contains a full LINE-1 element (approximately 6 kb) plus flanking sequence on both sides can be placed uniquely in the genome. The key is that the read must be long enough to include unique sequence on at least one side of the repeat.

The accuracy of long reads also matters. PacBio HiFi reads with >99.9% accuracy can resolve repeats that differ by a single nucleotide, whereas the higher error rate of raw Nanopore reads makes it difficult to distinguish between repeat copies that are nearly identical. For this reason, HiFi reads are now the preferred data type for assembling complex genomes.

### Optical Mapping and Hi-C for Scaffolding

Optical mapping is a technique that produces genome-wide restriction maps without sequencing. High-molecular-weight DNA is stretched on a surface, digested with a restriction enzyme, and imaged to produce a map of restriction sites. These maps are then used to scaffold sequence contigs by matching the predicted restriction pattern of the contigs to the observed pattern of the optical map.

Optical mapping is particularly useful for resolving large structural variants and for ordering contigs in regions that are difficult to assemble. However, it has limited resolution: the maps show restriction sites but not the underlying sequence, so they cannot resolve small repeats or sequence-level errors. Hi-C provides complementary information, giving a global view of [chromosome structure](/knowledge/molecular-biology/chromosome-structure) that can order and orient contigs even in the absence of sequence similarity.

## Gene Prediction and Structural Annotation

### Ab Initio Gene Prediction

Ab initio gene prediction uses statistical models of gene structure to identify genes directly from the genomic sequence, without external evidence. The most widely used tools are AUGUSTUS, GENSCAN, and SNAP. These programs use hidden Markov models (HMMs) that encode the typical features of genes: promoter elements, start codons, splice donor and acceptor sites, exons, introns, and stop codons.

The models are trained on a set of known genes, usually from the same species or a close relative. The training set provides the parameters: the frequency of each codon, the length distribution of exons and introns, the consensus sequences of splice sites, and the probability of each nucleotide at each position in the gene. During prediction, the HMM finds the most probable gene structure given the sequence and the model parameters.

Ab initio prediction is fast and requires no experimental data beyond the genome sequence itself, making it attractive for newly sequenced genomes. However, its accuracy is limited. In the human genome, ab initio predictions miss approximately 20% of exons and incorrectly predict many false exons. The accuracy is even lower for genomes with unusual features, such as very long introns, non-canonical splice sites, or biased codon usage. For this reason, ab initio prediction is rarely used alone; it is typically combined with evidence-based methods.

### Homology-Based Annotation

Homology-based annotation identifies genes by comparing the genome to known genes from other species. The basic approach is to align protein sequences from a reference database (such as Swiss-Prot or the NCBI non-redundant database) to the genome using a tool like TBLASTN, which translates the genome in all six reading frames and searches for matches to the query protein. Regions with significant matches are then refined to produce complete gene models.

The advantage of homology-based annotation is that it is highly specific: a gene that matches a known protein is almost certainly a real gene. The disadvantage is that it cannot find genes that have no homologs in the database—lineage-specific genes, fast-evolving genes, or genes that have been lost from the reference species. In a typical bacterial genome, homology-based methods identify 70–90% of genes; in a eukaryotic genome, the fraction is lower because of the greater evolutionary distance and the presence of more lineage-specific genes.

Tools such as Exonerate and GeneWise align proteins to genomic DNA while accounting for introns, using dynamic programming to find the optimal alignment that includes splice sites. The resulting gene models can be used directly or as evidence for other prediction methods.

### Transcriptome-Guided Annotation (RNA-seq, Iso-seq)

Transcriptome-guided annotation uses RNA sequencing data to identify genes experimentally. RNA-seq reads are aligned to the genome, and the alignment pattern reveals the exon-intron structure of expressed genes. Reads that span exon-exon junctions provide direct evidence for splice sites; the depth of coverage provides a measure of expression level.

The standard workflow is as follows:

1. Align RNA-seq reads to the genome using a splice-aware aligner such as HISAT2 or STAR.
2. Assemble transcripts from the aligned reads using Cufflinks, StringTie, or similar tools.
3. Compare the assembled transcripts to the genome to produce gene models.
4. Merge the transcript models with ab initio and homology-based predictions using an evidence combiner such as EvidenceModeler or MAKER.

Iso-seq (PacBio full-length cDNA sequencing) provides an important complement to short-read RNA-seq. Iso-seq reads span entire transcripts, from the 5' cap to the 3' poly-A tail, so they directly reveal the full exon structure of each isoform. This is particularly valuable for genes with complex [alternative splicing](/blog/guides/alternative-splicing), where short reads cannot reliably reconstruct the full transcript.

The integration of multiple evidence sources is critical. A gene model supported by both transcriptome data and homology to a known protein is far more reliable than one supported by either alone. The MAKER pipeline is designed for this purpose: it runs ab initio predictors, aligns evidence (proteins and transcripts), and produces a consensus set of gene models with quality scores.

## Functional Annotation and Data Resources

### Sequence Similarity Searches

Functional annotation begins with sequence similarity searches. The predicted protein sequences are compared to databases of known proteins using BLAST (Basic Local Alignment Search Tool). BLASTP compares a protein query to a protein database; BLASTX compares a translated nucleotide query to a protein database. The best hits provide a first-pass functional assignment: if a predicted protein matches a characterized protein with high similarity, it is likely to have a similar function.

The choice of database matters. The NCBI non-redundant (nr) database is comprehensive but contains many uncharacterized proteins. Swiss-Prot is smaller but curated, with each entry manually reviewed. The UniProt Knowledgebase combines both. For a focused analysis, species-specific databases may be more useful.

E-value thresholds must be set carefully. A stringent threshold (e.g., E < 1e-50) reduces false positives but may miss distant homologs. A relaxed threshold (E < 1e-5) captures more hits but includes many spurious matches. The appropriate threshold depends on the evolutionary distance between the query species and the database species.

### Domain and Motif Analysis

Sequence similarity to a whole protein can be complemented by domain analysis. Protein domains are conserved structural and functional units, typically 50–200 amino acids in length, that fold independently and often have a specific function such as DNA binding, kinase activity, or transmembrane transport. Domain databases include Pfam (protein families), InterPro (a meta-database that integrates Pfam, SMART, and others), and CDD (Conserved Domain Database).

InterProScan is the standard tool for domain annotation. It scans a protein sequence against multiple domain databases and reports the matches, along with the associated Gene Ontology (GO) terms. For example, a protein that matches the Pfam domain "WD40" and the SMART domain "WD_REPEATS" would be annotated as a WD40-repeat protein, with the GO term "protein binding."

Domain analysis can assign function to proteins that have no overall similarity to known proteins. A protein that contains a kinase domain but otherwise has no matches in the database is likely a kinase, even if its specific substrates are unknown. Domain analysis also provides structural information: the presence of a transmembrane domain suggests a membrane protein, and a signal peptide suggests secretion.

### Pathway and GO Enrichment

The Gene Ontology (GO) provides a controlled vocabulary for describing gene function in three categories: molecular function (e.g., "ATP binding"), biological process (e.g., "DNA repair"), and cellular component (e.g., "nucleus"). GO terms are assigned to genes based on sequence similarity, domain analysis, and experimental evidence. The assignment is made by tools such as Blast2GO, which combines BLAST results with InterProScan output to produce GO annotations.

Pathway databases such as KEGG (Kyoto Encyclopedia of Genes and Genomes) and MetaCyc provide a higher level of functional context. KEGG maps genes to metabolic and regulatory pathways, allowing the reconstruction of the organism's metabolic capabilities from its genome. For example, the presence of all enzymes in the glycolysis pathway suggests that the organism can metabolize glucose.

Pathway and GO enrichment analysis asks whether a set of genes (e.g., genes upregulated under a condition) is enriched for particular functions. The analysis uses a hypergeometric test or Fisher's exact test to compare the frequency of each GO term or pathway in the gene set to its frequency in the genome as a whole. The result is a list of overrepresented functions, which can guide biological interpretation.

## Common Pitfalls and Best Practices

### Misassembly and Chimeric Contigs

Misassembly occurs when the assembler joins sequences that are not adjacent in the genome. The most common cause is repetitive DNA: if a repeat is longer than the reads, the assembler cannot determine which flanking sequence belongs to which copy of the repeat, and it may join the wrong pairs. The result is a chimeric contig that contains sequence from two different genomic locations.

Misassembly can also result from sequencing errors that create false overlaps, from contamination that introduces foreign DNA, or from structural variation that confuses the assembler. The consequences are severe: a chimeric contig will produce incorrect gene models, false structural variants, and misleading comparative analyses.

Detection of misassembly requires independent evidence. Paired-end reads that map to distant locations in the assembly indicate a problem. Hi-C contact maps that show abrupt discontinuities indicate misjoins. Comparison to a closely related reference genome can reveal structural inconsistencies. The best practice is to validate the assembly with multiple independent methods before proceeding to annotation.

### Overprediction and Underprediction of Genes

Gene prediction errors come in two directions. Overprediction produces false genes: predictions that do not correspond to real transcripts. Underprediction misses real genes. Both are common, and both have distinct causes.

Overprediction is most common with ab initio methods, which can identify spurious open reading frames in intergenic regions, especially in AT-rich genomes where stop codons are rare. Pseudogenes—nonfunctional copies of real genes—are also frequently predicted as functional genes. The best defense is to require transcriptome support: a gene model that is not expressed in any RNA-seq dataset should be viewed with suspicion.

Underprediction is most common for genes with unusual features: very short genes, genes with non-canonical splice sites, genes embedded in repetitive sequence, and genes that are expressed only in specific tissues or conditions. Genes with large introns are also frequently missed because the statistical models underestimate the probability of long introns. The best defense is to use multiple prediction methods and to include RNA-seq data from multiple tissues and conditions.

### Contamination and Quality Control

Contamination is a pervasive problem in genome projects. DNA extracted from a cultured organism may contain DNA from the culture medium, from symbionts, or from other organisms in the sample. DNA extracted from a tissue sample may contain DNA from blood parasites, gut contents, or environmental microbes. The contaminating sequences are assembled along with the target genome, inflating the assembly size and producing spurious gene models.

The first line of defense is careful laboratory practice: use clean reagents, minimize handling, and sequence a negative control. The second line is computational screening. The assembly can be searched for sequences that match known contaminants (e.g., the human genome, common laboratory bacteria) using BLAST. The GC content and coverage depth can be examined for anomalies: a contaminant with different GC content will appear as a distinct peak in the GC distribution. Tools such as BlobTools and Kraken can classify contigs by taxonomic origin based on sequence composition and similarity.

The best practice is to screen the raw reads before assembly and the assembled contigs after assembly. Contigs that match contaminants should be removed, but the decision should be documented and the raw data should be retained for reanalysis.

## Frequently Asked Questions

### What is the difference between genome assembly and genome annotation?

Genome assembly is the process of reconstructing the genome sequence from sequencing reads. It produces contigs and scaffolds—long stretches of nucleotide sequence. Genome annotation is the process of interpreting that sequence: identifying genes, their structures, and their functions. Assembly answers the question "What is the sequence?" Annotation answers the question "What does it mean?" Assembly must precede annotation, and the quality of annotation depends on the quality of assembly.

### What is N50 in genome assembly?

N50 is a statistic that describes the contiguity of an assembly. It is the length L such that 50% of the assembled bases are in contigs or scaffolds of length ≥ L. To compute it, sort all contigs by length from longest to shortest, then find the point at which the cumulative length reaches 50% of the total. The length of the contig at that point is the N50. A higher N50 means a more contiguous assembly. N50 is a useful metric but does not measure correctness or completeness.

### Why are long reads better for genome assembly?

Long reads are better because they can span repetitive elements. A read that spans a repeat and includes unique flanking sequence on both sides can be placed uniquely in the genome. Short reads that fall entirely within a repeat cannot be placed uniquely, and the assembly breaks at repeat boundaries. Long reads also provide better phasing of variants and can resolve complex structural variation. The trade-off is that long-read sequencing is more expensive per base and has historically had higher error rates, though HiFi reads have largely solved the accuracy problem.

### How do you evaluate the completeness of a genome assembly?

The standard approach is BUSCO (Benchmarking Universal Single-Copy Orthologs), which searches the assembly for a set of genes expected to be present as single copies in the target lineage. Each gene is classified as complete, fragmented, or missing. A high-quality assembly should recover >95% of the BUSCO set as complete. Other approaches include comparing the assembly size to the expected genome size, checking the coverage of conserved syntenic regions relative to a reference genome, and verifying that all expected chromosomes are represented.

### What is ab initio gene prediction?

Ab initio gene prediction uses statistical models of gene structure to identify genes directly from genomic sequence, without external evidence. The models, typically hidden Markov models, encode the features of genes: start codons, splice sites, exon and intron length distributions, and codon usage. The model is trained on known genes from the same or a related species, then applied to the genome to find the most probable gene structures. Ab initio prediction is fast but less accurate than evidence-based methods.

### How does RNA-seq help in genome annotation?

RNA-seq provides direct experimental evidence for gene expression. By aligning RNA-seq reads to the genome, you can identify which regions are transcribed, determine the exon-intron structure of transcripts, and quantify expression levels. Splice-aware aligners can detect reads that span exon-exon junctions, providing direct evidence for splice sites. RNA-seq data are used to refine ab initio predictions, to identify genes missed by other methods, and to validate gene models. The best annotation pipelines integrate RNA-seq with ab initio and homology-based evidence.

### What are common causes of misassembly?

The most common cause is repetitive DNA. When a repeat is longer than the sequencing reads, the assembler cannot determine which flanking sequence belongs to which copy, and it may join the wrong sequences. Other causes include sequencing errors that create false overlaps, contamination that introduces foreign DNA, and structural variation that confuses the assembler. Misassembly can be detected by checking the consistency of paired-end read alignments, examining Hi-C contact maps, and comparing to reference genomes.

### What is the best approach for annotating a new genome?

The best approach integrates multiple lines of evidence. First, generate a high-quality assembly using long reads (PacBio HiFi or Oxford Nanopore) with short-read polishing. Second, generate transcriptome data from multiple tissues and conditions, including both short-read RNA-seq and long-read Iso-seq if possible. Third, run multiple gene prediction methods: ab initio, homology-based, and transcriptome-guided. Fourth, combine the evidence using a pipeline like MAKER or EvidenceModeler. Fifth, assign functional annotations using BLAST, InterProScan, and pathway databases. Finally, manually curate the most important gene families.

## Key Takeaways

- Genome assembly reconstructs the sequence from reads; genome annotation interprets that sequence biologically. Both are required for a useful genome resource.
- Long-read sequencing (PacBio HiFi, Oxford Nanopore) is essential for assembling complex genomes because it can span repetitive elements that break short-read assemblies.
- Assembly quality must be evaluated with multiple metrics: N50 for contiguity, BUSCO for completeness, and read-mapping consistency for correctness.
- Gene prediction should integrate ab initio, homology-based, and transcriptome-guided methods; no single method is sufficient.
- Functional annotation relies on sequence similarity (BLAST), domain analysis (InterProScan), and pathway mapping (KEGG, GO), with confidence increasing when multiple lines of evidence agree.
- The most common pitfalls are misassembly at repeats, gene overprediction from ab initio methods, and contamination from foreign DNA. Each can be mitigated with appropriate validation and quality control.
- A high-quality genome project is iterative: assembly informs annotation, and annotation problems often trace back to assembly errors that must be corrected first.

## Further Reading

- Michael TP. *Plant genome assembly and annotation*. Current opinion in plant biology. 2026. [PubMed 41678991](https://doi.org/10.1016/j.pbi.2026.102859)
- Narh Mensah DL, Wingfield BD, Coetzee MP. *A practical approach to genome assembly and annotation of Basidiomycota using the example of Armillaria*. BioTechniques. 2023. [PubMed 37681497](https://doi.org/10.2144/btn-2023-0023)
- Zhang S et al. *[Whole genome sequencing](/blog/guides/whole-genome-sequencing) and annotation of Scleroderma yunnanense, the only edible Scleroderma species*. Genomics. 2023. [PubMed 37839651](https://doi.org/10.1016/j.ygeno.2023.110727)
- Seppey M, Manni M, Zdobnov EM. *BUSCO: Assessing Genome Assembly and Annotation Completeness*. Methods in [molecular biology](/blog/careers/molecular-biology) (Clifton, N.J.). 2019. [PubMed 31020564](https://doi.org/10.1007/978-1-4939-9173-0_14)
- Simão FA et al. *BUSCO: assessing genome assembly and annotation completeness with single-copy orthologs*. Bioinformatics (Oxford, England). 2015. [PubMed 26059717](https://doi.org/10.1093/bioinformatics/btv351)
- Kimbrel JA, Jeffrey BM, Ward CS. *Prokaryotic Genome Annotation*. Methods in [molecular biology](/blog/careers/molecular-biology) (Clifton, N.J.). 2022. [PubMed 34718997](https://doi.org/10.1007/978-1-0716-1585-0_10)

## Related Topics

- [Genome Sequencing and Assembly](/knowledge/molecular-biology/genome-sequencing-and-assembly)
- [Process RNA-seq Data](/knowledge/molecular-biology/process-rna-seq-data)
- [Sanger Sequencing Protocol](/knowledge/molecular-biology/sanger-sequencing-protocol)
- [Bisulfite Sequencing](/knowledge/molecular-biology/bisulfite-sequencing)
- [ATAC Sequencing](/knowledge/molecular-biology/atac-sequencing)

## Related Clinical & Scientific Guides

* [MAPK Pathway: Mechanism, Function, and Clinical Relevance](/knowledge/molecular-biology/mapk-pathway)
* [Mammalian Cell Culture Bioreactors: A Practical Guide](/knowledge/molecular-biology/mammalian-cell-culture-bioreactor)
* [Nucleotide Formation: Biosynthesis and Assembly of DNA/RNA Building Blocks](/knowledge/molecular-biology/nucleotide-formation)