Integrative Genome Annotation: Advanced Pipelines for Structural Discovery and Functional Characterization

Genome annotation is often presented as a linear workflow. In real projects, it behaves more like an arbitration system. The assembly provides sequence, but sequence alone does not tell you which open reading frames are real, where exon boundaries should fall, whether two neighboring coding segments belong to one gene or two, or whether a convincing-looking model is actually repeat-derived noise. Those calls become reliable only when multiple evidence types are forced to agree.

That is why strong annotation programs do not rely on a single predictor. They combine species-aware ab initio modeling, cross-species homology, short-read transcript support, full-length transcript evidence, repeat masking, and targeted human review. Upstream data quality matters just as much. Projects that aim for annotation-ready assemblies often begin with plant and animal whole-genome de novo sequencing, broader whole-genome sequencing support for annotation-ready assemblies, or, when continuity is the limiting factor, telomere-to-telomere sequencing.

The goal is not to produce the highest possible number of gene models. The goal is to produce the highest possible number of defensible gene models. That word matters. A defensible model is one whose structure can be explained by the evidence that supports it, and whose weak points are still visible rather than hidden behind a confident label.

This is the real shift behind modern annotation. The hard problem is no longer generating candidate models at scale. The hard problem is deciding which evidence layer should dominate when the data disagree.

Figure 1: This figure shows that annotation is not a one-direction pipeline but a convergence problem in which assembly quality, repeat masking, structural evidence, functional evidence, and manual review feed into the same final gene set.

Structural annotation starts by defining the gene space

Structural annotation asks two linked questions. Where are the genes, and what is their exon-intron architecture? In eukaryotic genomes, those questions are harder than they look. Real genes are interrupted by introns, surrounded by repeats, blurred by pseudogenic fragments, and complicated by alternative splicing. A useful pipeline therefore has to separate true biological structure from sequence patterns that only look gene-like.

The easiest mistake in this area is to treat all evidence as interchangeable. It is not. Each evidence class is best at resolving a different uncertainty.

That table captures the core rule of modern annotation: do not ask one evidence type to solve a problem that belongs to another.

Ab initio prediction is still essential, but it is only a first hypothesis

Ab initio prediction remains central because it gives full-genome coverage. Tools such as AUGUSTUS and GeneMark scan the assembly and identify regions whose sequence statistics look compatible with coding structure. They are powerful because they do not need every locus to have a close homolog or an expressed transcript in the sampled tissue. Without them, lineage-specific genes and poorly expressed loci would be much harder to recover.

But ab initio prediction is often explained too loosely. These tools are not just "guessing genes." They are scoring a sequence landscape built from gene-shaped signals. Start and stop codons matter, but so do splice donor and acceptor patterns, coding composition, exon length distributions, intron structure, and species-shaped transitions between coding and non-coding states. In practice, the predictor is asking whether the local sequence behaves like a plausible path through gene architecture.

That is exactly why training quality changes the result so much. A well-trained model learns what coding structure looks like in the target lineage. A poorly trained model learns an approximation. The output may still look polished, but the error profile changes fast. Small exons disappear. Neighboring genes fuse. One interrupted locus becomes two artificial genes. A repeat fragment becomes a short coding model because its local signal is statistically convincing enough to pass.

The key point is simple. Ab initio prediction is strongest when it is treated as the first draft of structure, not the final truth.

HMM logic matters because genes are state transitions, not isolated motifs

Your article angle rightly emphasizes HMM-based gene finding. That deserves explicit treatment because it is one of the least well explained parts of public annotation content.

Gene structure is not defined by one motif. It is defined by a sequence of transitions. Coding sequence tends to move into splice boundaries, then introns, then coding sequence again. Intergenic sequence follows a different statistical pattern. Probabilistic models are useful because they do not evaluate each signal in isolation. They evaluate whether the sequence behaves like a believable path through gene states.

That matters in practice for two reasons.

First, good state modeling improves discrimination between real genes and decoys. A true exon is not only coding-like. It is positioned in a way that makes sense relative to splice signals and neighboring sequence context.

Second, the model becomes highly sensitive to bad priors. If training examples are weak, contaminated, fragmented, or too distant taxonomically, the state transitions lose sharpness. The software still returns gene models, but the biological trustworthiness drops. This is why two projects can both claim to use AUGUSTUS or GeneMark and still end up with gene sets of very different quality.

A stronger way to explain this to readers is not "which predictor did you run?" The better question is "what evidence taught the predictor what a gene looks like before it started scanning the genome?"

Coverage versus credibility: the real trade-off in ab initio modeling

A useful decision rule is to think of ab initio output in two dimensions.

Coverage asks whether the predictor can scan the whole genome and propose candidate loci broadly enough.

Credibility asks whether those proposed loci remain believable after they are confronted with transcript evidence, homology evidence, and repeat-aware filtering.

Strong annotation does not sacrifice one dimension for the other. It uses ab initio prediction to maximize coverage, then uses orthogonal evidence to protect credibility.

That is why purely de novo structural calling almost always inflates confidence. The software is allowed to explain too much with too little constraint. A more disciplined pipeline asks a harsher question: which predicted structures remain intact after evidence layers start disagreeing?

Homology mapping narrows the solution space

Homology-based annotation adds biological restraint. If related species already have curated or high-confidence proteins, those sequences can be aligned to the new assembly to anchor plausible coding regions. This is especially useful when ab initio models start drifting into overcalling, or when expression evidence is incomplete for the tissues, stages, or treatments that matter most.

The value of homology is often described too broadly. Its main strength is not that it proves a gene exists. Its strength is that it makes many implausible models much harder to defend.

A conserved protein can stabilize exon structure, preserve reading-frame expectations, and reduce the chance that a repeat-derived ORF is mistaken for a real gene. This works especially well for conserved enzymes, core cellular machinery, and families with stable domain architecture.

But homology has sharp limits. If the reference is too distant, exon boundaries drift. If the family expanded recently, one reference protein may map across several paralogous loci. If the target lineage gained a novel exon or lost a domain, a homology-first workflow may flatten real biology into an old template.

That is why homology should be treated as a constraint layer, not a mold. It limits bad models. It does not replace organism-specific evidence.

Transcript evidence is the best antidote to structural guesswork

If ab initio prediction provides breadth and homology provides plausibility, transcript evidence provides locality. It tells you where the organism actually transcribed sequence under the sampled conditions. That makes it one of the strongest correctives in the entire structural workflow.

For many projects, standard RNA-Seq analysis is not a side dataset. It is one of the main filters that prevents the structural gene set from drifting away from real splice evidence. Junction-supporting reads can confirm exon boundaries, rescue missed exons, and down-rank models that look statistically plausible but never receive expression support.

Still, short reads do not solve everything. They are strong at coverage, but weaker at transcript continuity. When loci are highly alternative, when exons are short, or when paralogs are very similar, short reads often tell you that transcription happened without telling you exactly which exons belong to the same transcript molecule.

That is where long-read transcript evidence changes the problem.

Iso-Seq for full-length transcript discovery and Nanopore full-length transcript sequencing improve exon chaining, transcript continuity, UTR recovery, and isoform resolution. They do not remove all ambiguity. Expression remains tissue-biased, and low-abundance transcripts can still be missed. But they sharply reduce the amount of inference needed in difficult loci.

A concise way to explain the hierarchy is this:

• RNA-seq shows where transcription support exists
• splice-aware mapping shows which junctions are credible
• full-length transcript data shows which exon chains belong together
• combined evidence shows which structural models survive correction

That last line matters most. Evidence is not simply additive. It is corrective. Each layer edits a different type of mistake.

Figure 2: This figure illustrates how an ab initio draft model is edited by splice support, then refined again by full-length transcript evidence until the final isoform structure becomes defensible.

What to do when the evidence disagrees

Disagreement is normal. The wrong response is to average everything mechanically.

A better response is to ask which evidence type is best positioned to resolve the specific uncertainty:

• If the uncertainty is an exon boundary, transcript evidence should usually dominate.
• If the uncertainty is whether a short ORF is real or repeat-derived, repeat context and homology restraint should dominate.
• If the uncertainty is whether several similar models represent one conserved family or a recent expansion, homology and domain structure should dominate.
• If the uncertainty is transcript continuity across a complex locus, full-length transcript evidence should dominate.
• If none of these layers resolves the conflict cleanly, the locus should remain provisional and be sent to manual review.

This decision-first framing is one of the most useful ways to make an annotation article feel expert rather than generic. It shows readers how evidence should be weighted, not just which datasets are fashionable.

The assembly under the annotation still decides the ceiling

A common mistake is to discuss assembly and annotation as if they were separate service boxes. In practice, annotation quality inherits the strengths and weaknesses of the assembly beneath it.

A fragmented assembly breaks loci apart. Repeat collapse distorts local gene density. Residual haplotypic duplication inflates apparent gene number. Misjoins create false proximity between unrelated coding segments. Once those errors enter the substrate, even a sophisticated annotation pipeline can only work around them, not erase them.

That is why annotation-ready projects increasingly treat genome architecture as part of annotation planning rather than a purely upstream task. In larger eukaryotic genomes, chromosome-scale scaffolding from Hi-C sequencing can reduce structural ambiguity, improve locus continuity, and make later gene-model arbitration more reliable.

The practical rule is harsh but useful: annotation can refine a genome, but it cannot fully rescue a weak substrate.

Repeat masking is not housekeeping; it is quality control for the whole pipeline

Repeat masking is one of the most underestimated steps in genome annotation. It is often described in a sentence, then buried under the more visible parts of gene prediction. That treatment is misleading. Repeat handling changes the false-positive environment of the entire workflow.

Eukaryotic genomes are full of repetitive DNA: transposable elements, low-complexity regions, tandem arrays, simple repeats, and lineage-specific repeat families that may not appear in generic libraries. Some are clearly non-coding. Some overlap genes. Some donate fragments that look exon-like. Some generate just enough ORF structure to trick a predictor into calling a coding locus.

Once that happens, the rest of the annotation stack starts wasting effort on artifacts.

RepeatModeler and RepeatMasker are important because generic libraries are not enough

A common workflow pairs RepeatModeler with RepeatMasker. The logic is straightforward. Generic repeat databases do not capture every lineage-specific family, especially in non-model organisms. De novo repeat discovery gives the project a repeat library that actually reflects the genome being annotated. Masking then marks those regions so downstream structural steps can treat them with caution.

The most useful masking mode is usually soft masking. Hard masking removes sequence aggressively and can erase context that is still biologically informative. Soft masking preserves the sequence while flagging it as repeat-derived. That is a better fit for annotation because it reduces false positives without pretending the repeatome is biologically irrelevant.

Weak repeat masking produces an error ecology, not one isolated mistake

When repeat masking is weak, the damage does not show up in one place. It propagates.

That pattern is why repeat masking should be framed as a quality-control gate, not a preprocessing chore. If the repeatome is poorly modeled, every later evidence layer is forced to work in a dirtier search space.

Figure 3: This figure demonstrates how weak masking creates several kinds of structural error at once, then contrasts that with the cleaner interpretation produced by repeat-aware correction.

Repeat-aware annotation requires judgment, not just masking

The repeatome should not be treated only as an obstacle. It is also a biologically meaningful layer of the genome. Repeats shape genome size, local architecture, regulatory innovation, and lineage-specific structure. A mature annotation workflow therefore has to do two things at once.

It must suppress repeat-derived false structure during gene prediction.

It must also preserve repeat annotation as an interpretable genomic feature for downstream analysis.

That dual role is one reason repeat handling deserves its own conceptual space in the article. It is not just there to make the coding annotation cleaner. It also determines how honestly the genome is represented.

Functional annotation begins when structural confidence is high enough

Once a structural gene set exists, the next question is obvious: what do these genes do? The shallow answer is to run a similarity search, take a top hit, and transfer the label. That approach is fast, familiar, and often too confident.

A better question is this: what combination of similarity, domain architecture, and ortholog context supports the most defensible function call?

That shift matters because function transfer fails in predictable ways. Paralogs look close but behave differently. Partial proteins inherit over-specific names. Multi-domain proteins borrow labels from one preserved domain while ignoring the others. Expanded families create many near-matches, none of which deserves direct one-to-one name transfer.

This is why good functional annotation should behave like layered evidence arbitration, just as structural annotation does.

Fast similarity search is useful because it builds a neighborhood, not because it gives a final answer

Tools such as DIAMOND are valuable because they make proteome-scale similarity searching feasible. They let a project rapidly identify a neighborhood of plausible matches across large protein databases. That is operationally important, but the deeper value is interpretive. Fast search allows the workflow to gather context rather than forcing one top hit to carry the entire meaning of the protein.

Used correctly, similarity search answers questions like these:

• Which known proteins does this sequence resemble?
• Is the similarity broad or narrow?
• Does the match support a family-level label or a precise label?
• Is the sequence well represented in existing databases, or does it appear more weakly conserved?

Those are useful outputs. None of them, by itself, is enough to justify a highly specific name transfer.

That is the point where the second half of the article picks up: domain-level inference with HMMER, ontology mapping with GO, KEGG, and eggNOG, when function calls should be downgraded to broad or provisional labels, how manual curation fits into difficult loci, and how to choose between MAKER, BRAKER, and Ensembl-style annotation logic.

Figure 4: This figure introduces the functional evidence stack by showing how sequence similarity, domain detection, and ortholog context progressively narrow a function call.

Domain-level inference with HMMER: why conserved architecture matters

Fast similarity search gives a useful neighborhood. It does not, by itself, tell you whether the proposed function is structurally coherent. That is the job of domain-level inference.

Tools built around profile Hidden Markov Models, such as HMMER, ask a stricter question: does the predicted protein contain the conserved statistical signature expected for a real member of a domain family? This matters because many annotation errors are not caused by the absence of similarity. They are caused by misplaced specificity. A protein may look broadly similar to a known family, yet lack the catalytic domain, binding module, regulatory tail, or domain order required for the specific function being transferred.

That is why domain analysis should be treated as a checkpoint rather than an accessory step. It helps in at least four ways.

First, it rescues function calls when full-length identity is modest but the core architecture is intact. Second, it rejects overconfident labels when only part of the expected structure is present. Third, it exposes domain shuffling, which is common in eukaryotic genomes and often changes biological interpretation. Fourth, it helps separate a true member of a family from a truncated, fused, or degenerated relative.

The practical value is simple. Similarity gives neighborhood. Domains give mechanism. When the two agree, confidence rises. When they disagree, the annotation should become broader, not more specific.

When function calls should be downgraded

One of the most useful habits in genome annotation is knowing when not to transfer a detailed function name. This is where many pipelines become overconfident. A clean annotation is not the same as an honest one.

A function call should usually be downgraded to a broad or provisional label under the following conditions:

• Partial domain architecture: the sequence matches a known family, but only part of the expected domain structure is present.
• Weak orthology support: the protein has homologs, but its position within orthologous groups is unstable or too broad.
• Family expansion: the gene belongs to a rapidly expanded paralog family where nearest-hit transfer is especially risky.
• Conflicting top hits: different high-scoring matches imply different specific functions.
• Repeat-adjacent or structurally unstable loci: the protein model itself may be incomplete or incorrect.
• Fragmented coding sequence: the predicted protein is truncated, fused, or broken across a difficult assembly interval.

In those cases, a broad label is not a weakness. It is a technical safeguard. It tells downstream users that the sequence belongs to a credible functional neighborhood, but that the current evidence does not justify overclaiming.

This is one reason why functional annotation should be written as a confidence ladder rather than a binary call. A good output system distinguishes between:

• high-confidence specific function
• family-level function
• domain-containing protein
• hypothetical or uncharacterized protein

That hierarchy is far more useful than forcing every sequence into a confident-looking name.

GO, KEGG, and eggNOG: turning genes into interpretable systems

Once sequence-level and domain-level evidence are strong enough, the next step is to connect genes to larger biological structure. This is where ontology and orthology mapping become central.

GO is useful because it organizes annotation into molecular function, biological process, and cellular component. That gives the gene set a controlled vocabulary. Instead of carrying only free-text protein names, the annotation begins to support enrichment analysis, process-level comparison, and more stable cross-project interpretation.

KEGG adds pathway logic. This matters when the real biological question is not "what is this protein called?" but "does this genome encode the components needed for a pathway, module, or metabolic branch?" Pathway mapping turns a list of gene products into a systems-level picture.

eggNOG adds orthology-aware structure. That is especially important when a protein belongs to a large family with many paralogs. Straight similarity transfer can overfit to the nearest sequence. Orthology-aware grouping gives a more conservative framework and often improves the discipline of downstream GO and pathway assignment.

A strong functional workflow therefore moves in layers:

1. Use similarity search to identify a plausible functional neighborhood.
2. Use domain models to test whether the architecture supports that interpretation.
3. Use orthology to decide whether the label should stay broad or can become more specific.
4. Map the sequence into GO, KEGG, and related systems only after that evidence stack is coherent.
5. Preserve uncertainty when the stack does not fully converge.

Figure 5: This figure illustrates that functional annotation is a layered narrowing process, not a one-step label transfer.

The manual curation paradox

Automation is essential in genome annotation. It is also incomplete by design. The largest share of loci can be processed well enough by automated pipelines, but the loci that matter most to biological interpretation are often the ones that automation handles worst.

That is the manual curation paradox.

The hardest cases usually include tandemly duplicated families, repeat-adjacent loci, microexon-containing genes, pseudogene-rich neighborhoods, long and variable UTRs, nested transcription, and families with rapid lineage-specific expansion. These are exactly the regions where a polished automated gene model may still be wrong in a biologically important way.

This is why browser-based review environments such as Apollo and JBrowse remain important. They allow a curator to inspect the evidence stack in context rather than treating the final annotation file as a sealed product. A curator can ask practical questions:

• Do the splice junctions have real support?
• Does the predicted coding frame remain stable across the locus?
• Does the homology evidence support one gene or several?
• Is the apparent model crossing into repeat-derived sequence?
• Do the long-read isoforms agree with the short-read splice structure?
• Is the locus biologically important enough to justify review even if the model is only moderately uncertain?

A strong annotation program does not send every disagreement to a curator. That does not scale. Instead, it triages disagreements according to their likely impact on biological interpretation.

A useful escalation rule looks like this:

• Keep the locus automated when ab initio structure, homology, transcript evidence, and functional interpretation all agree.
• Escalate the locus when one evidence layer breaks sharply from the others.
• Prioritize human review when the disagreement affects a project-critical target, a high-priority biological family, or a highly visible deliverable in the study.

That last point matters. Annotation quality is not measured only by global completeness metrics. It is also measured by whether the loci that matter most to the project were handled with enough care.

Figure 6: This figure reveals which kinds of loci move from automated annotation into manual review, and why.

MAKER vs. BRAKER vs. Ensembl: three different annotation logics

The common question is which pipeline is best. The more useful question is which annotation logic best matches the project.

MAKER: modular evidence integration and iterative refinement

MAKER is often the better fit when the project needs flexibility. It is designed to combine multiple evidence streams in a modular way, and it works well when annotation improves over rounds rather than in one pass. That makes it attractive for projects that expect iterative updates, changing evidence inputs, or repeated refinement of training and filtering choices.

Its strength is not that it automatically solves every problem. Its strength is that it gives the project room to evolve. Teams can compare predictor behavior, incorporate new transcript evidence, and improve the annotation set without rebuilding the workflow philosophy from scratch.

BRAKER: automated structural prediction with evidence-guided training

BRAKER is often the stronger choice when the immediate need is a fast, solid structural baseline in a eukaryotic genome with transcript evidence available. Its main advantage is that it automates one of the hardest parts of prediction: shaping species-aware models using evidence rather than assuming that generic parameters are good enough.

That makes BRAKER especially useful for non-model eukaryotes where training quality is a major determinant of annotation quality. It reduces manual burden without pretending that training does not matter.

Its limitation is that it remains part of a larger system. It gives a powerful structural backbone, but repeat handling, difficult-locus review, and downstream function transfer still need separate discipline.

Ensembl-style annotation: standardized production logic

Ensembl-style annotation is best understood as a production framework rather than a lightweight standalone choice. It emphasizes standardization, repeat-aware preprocessing, evidence integration, stable releases, and, in selected cases, manual curation layered into the build process.

Its strength is consistency. That is crucial when the goal is not only to annotate a genome, but also to maintain comparability across species, builds, or release cycles.

Its limitation is that this style of annotation usually makes the most sense in reference-oriented programs rather than smaller projects that mainly need a fast, adaptable answer.

Practical comparison

A practical decision rule works well:

Choose MAKER when the project is likely to change as new evidence arrives and iterative refinement is part of the plan.

Choose BRAKER when the priority is a fast, evidence-guided structural baseline for a eukaryotic genome.

Choose an Ensembl-style approach when the priority is release discipline, consistency across builds, and reference-oriented annotation quality.

This is not a winner-takes-all comparison. It is a fit-to-purpose comparison.

Annotation quality is inherited from upstream design

By this stage, one principle should be clear. Annotation quality does not begin with annotation software. It begins with project design.

A fragmented assembly constrains structural confidence. Weak repeat modeling inflates false-positive space. Poor transcript sampling narrows isoform recovery. Weak homology sets reduce biological restraint. Careless function transfer inflates specificity. No amount of polishing at the end can fully erase those upstream decisions.

That is why strong projects are increasingly designed backward from the biological question.

If the main question depends on isoform structure, then full-length transcript sequencing or Nanopore direct RNA sequencing may be central rather than optional.

If the main question depends on chromosome context and locus continuity, then Hi-C sequencing becomes part of annotation readiness, not a separate downstream convenience.

If the main question depends on regulatory interpretation, then annotation may need to be paired with ATAC-Seq or ChIP-Seq so gene models can be interpreted alongside chromatin state and binding context.

The best workflow is not the one with the longest methods list. It is the one where each evidence type is present because it removes a known uncertainty.

Final perspective

Integrative genome annotation is not the mechanical act of stacking tools until a GFF file appears. It is the disciplined process of deciding which evidence is strong, which evidence is weak, and which loci still require human judgment.

Ab initio prediction gives coverage. Homology gives restraint. RNA-seq gives splice support. Iso-Seq gives transcript continuity. Repeat masking reduces false structure. Domain analysis prevents careless name transfer. GO, KEGG, and eggNOG connect gene models to systems-level interpretation. Manual curation protects the project from the small number of mistakes that can distort a very large biological conclusion.

That is the modern workflow. Not a straight line, but a controlled negotiation among evidence layers.

FAQ

What evidence combination usually produces the most defensible gene models?
For most eukaryotic genomes, the strongest baseline comes from a high-quality assembly, repeat-aware preprocessing, ab initio prediction, protein homology, and transcript evidence. Confidence improves further when full-length transcript data is available for complex loci.

How much transcript evidence is enough for a new eukaryotic annotation project?
There is no single threshold, because the answer depends on genome complexity, tissue diversity, and the project question. Short-read RNA-seq may be sufficient for broad splice support, but full-length transcript data becomes much more important when isoform structure is central to the study.

What is the difference between structural annotation and functional annotation?
Structural annotation defines where genes are and how their exon-intron architecture is organized. Functional annotation assigns probable biological roles to the resulting proteins and pathways.

Why is repeat masking necessary before gene prediction?
Because repetitive DNA can mimic coding signals, attract misleading alignments, and inflate false gene counts. Repeat-aware masking reduces that background before structural prediction begins.

Is ab initio prediction enough for a new eukaryotic genome?
Usually not. It provides genome-wide coverage, but accuracy improves when transcript evidence, homology evidence, and repeat-aware filtering are added.

Why does long-read transcript data matter so much?
Because it improves transcript continuity, isoform resolution, UTR recovery, and exon chaining in loci where short reads leave ambiguity.

When should a locus be manually curated?
When major evidence layers disagree, or when the locus belongs to a high-priority family and a modeling error would materially affect the biological conclusion.

Where does eggNOG fit into annotation?
It provides ortholog-aware context, which helps transfer function more conservatively than plain similarity alone.

Can a strong annotation compensate for a weak assembly?
Only partly. Good annotation can reduce some ambiguity, but fragmentation, repeat collapse, and unresolved duplication still limit the confidence of the final gene set.

References

1. Bruna T, Hoff KJ, Lomsadze A, Stanke M, Borodovsky M. BRAKER2: automatic eukaryotic genome annotation with GeneMark-EP+ and AUGUSTUS supported by a protein database. NAR Genomics and Bioinformatics. 2021;3(1):lqaa108. DOI: 10.1093/nargab/lqaa108
2. Hoff KJ, Lange S, Lomsadze A, Borodovsky M, Stanke M. BRAKER1: unsupervised RNA-Seq-based genome annotation with GeneMark-ET and AUGUSTUS. Bioinformatics. 2016;32(5):767-769. DOI: 10.1093/bioinformatics/btv661
3. Holt C, Yandell M. MAKER2: an annotation pipeline and genome-database management tool for second-generation genome projects. BMC Bioinformatics. 2011;12:491. DOI: 10.1186/1471-2105-12-491
4. Campbell MS, Holt C, Moore B, Yandell M. Genome Annotation and Curation Using MAKER and MAKER-P. Current Protocols in Bioinformatics. 2014;48:4.11.1-39. DOI: 10.1002/0471250953.bi0411s48
5. Hoff KJ, Stanke M. Predicting Genes in Single Genomes with AUGUSTUS. Current Protocols in Bioinformatics. 2019;65(1):e57. DOI: 10.1002/cpbi.57
6. Smit AFA, Hubley R, Green P. RepeatMasker Open-4.0. Software and project documentation. Available from the RepeatMasker project site.
7. Buchfink B, Reuter K, Drost HG. Sensitive protein alignments at tree-of-life scale using DIAMOND. Nature Methods. 2021;18(4):366-368. DOI: 10.1038/s41592-021-01101-x
8. Eddy SR. Accelerated Profile HMM Searches. PLoS Computational Biology. 2011;7(10):e1002195. DOI: 10.1371/journal.pcbi.1002195
9. Huerta-Cepas J, Szklarczyk D, Heller D, et al. eggNOG 5.0: a hierarchical, functionally and phylogenetically annotated orthology resource. Nucleic Acids Research. 2019;47(D1):D309-D314. DOI: 10.1093/nar/gky1085
10. Kanehisa M, Furumichi M, Sato Y, Kawashima M, Ishiguro-Watanabe M. KEGG for taxonomy-based analysis of pathways and genomes. Nucleic Acids Research. 2023;51(D1):D587-D592. DOI: 10.1093/nar/gkac963
11. The Gene Ontology Consortium. The Gene Ontology knowledgebase in 2023. Genetics. 2023;224(1):iyad031. DOI: 10.1093/genetics/iyad031
12. Korlach J, Gedman G, Kingan SB, et al. De novo PacBio long-read and phased avian genome assemblies correct and add to genes important in neuroscience research. Gigascience. 2017;6(10):1-16. DOI: 10.1093/gigascience/gix085