The authors, led by Bishoy Maher Zaki of the Bioscience Research Laboratories in Giza, Egypt, together with colleagues at Adelaide University, the University of Guelph, the University of Virginia, Flinders University, Argonne National Laboratory and Cairo University, frame their guidelines around a single foundational principle: annotate with future applications in mind. Annotating a phage genome to expand knowledge of viral biodiversity demands a different emphasis from annotating one destined for biocontrol, phage therapy, or use as a source of bioactive proteins such as lysins, depolymerases or molecular biology enzymes. Once that purpose is fixed, the thirteen tips follow as pillars supporting a reliable annotation workflow.

The first tip is deceptively simple: learn the basics of phage biology before touching the data. Bioinformaticians experienced with cellular genomes still need to understand phage protein families and modules, when they are expressed, and how they interact. Structural features matter too, because transmembrane domains and signature conserved motifs can decisively aid the annotation of gene products such as lysins and holins. The authors point to established educational resources, including the SEA-PHAGES program, as starting points for building this foundation.

The second tip addresses the raw data itself. Quality control and genome assembly are indispensable, and the authors emphasize a counterintuitive point: too much sequencing depth can be as harmful as too little. Because phage genomes are small, Illumina runs often reach extremely high coverage, and in such datasets sequencing errors, rare contaminants such as a second closely related phage or prophage, and microdiversity in the form of single nucleotide polymorphisms can accumulate enough reads to destabilize assembly, especially in repetitive regions. The recommendation is to subsample short-read data to a depth of 25 to 100 times coverage. The optimal depth for long-read sequencing remains less clearly established and depends on the workflow. The choice of assembler also matters, since the same dataset can yield a complete genome with one algorithm and a fragmented one with another; if an assembly looks wrong, the authors advise comparing several algorithms on completeness, contig number and length, assembly-graph structure, read-mapping depth, and alignment to related genomes.

Coverage patterns can reveal more than assembly quality. Even coverage across a contig suggests reads from a single biological source, whereas uneven or background-level coverage flags host DNA or contaminating viruses, and metagenomic binning can help verify purity. Coverage depth patterns and the orientation of read starts can also be used to predict phage genome termini and DNA packaging mechanisms with tools such as PhageTerm, although transposase-based library preparation methods like Nextera XT complicate this because they modify the natural physical ends of the genome. On sequencing strategy, the authors note that short reads offer higher per-base accuracy while long reads better resolve genome structure and terminal repeats but may need extensive polishing. Hybrid co-assembly can fail on genomes with complex structures such as direct terminal repeats, which is why the authors highlight HYPPA, a workflow that performs long-read assembly first and then applies sequential long- and short-read polishing with multiple tools; HYPPA outperformed traditional hybrid co-assembly and produced complete genomes of DTR-containing T7-like phages.

Tip three concerns something rarely celebrated in methods papers: file organization. Because annotation pipelines require repeated rounds of optimization, the authors recommend creating a folder for each step, such as trimming, assembly, automated annotation and manual verification, with files named and dated according to settings and outcomes. They suggest underscores rather than spaces or hyphens in file names, so names can be used directly in scripts, and the ISO 8601 date format, which sorts unambiguously by date. Spreadsheets, they add, are valuable for tabulating comparisons between tools during manual verification.

Tips four and five tackle the core computational step: gene calling. Most major automated annotation pipelines were built for bacterial genomes, not phages, although phage-aware alternatives have emerged, including PhANOTATE, Phage Commander, Pharokka, PhaGAA, PhageScope, Empathi and the BV-BRC phage pipeline. Automated output is not error-free, so manual verification remains essential. Phage genomes are extremely compact, with coding sequences that often overlap or are separated by very short gaps, and their coding capacity typically reaches 90 to 95 percent; if it appears substantially lower, the genome may use an alternative genetic code. Short coding sequences under 100 base pairs are easily missed by prokaryotic gene callers because the statistical signal distinguishing them from random background is weak. To validate gene calls, the authors recommend comparing results across multiple tools and using closely related reference genomes, screening BLASTn hits with coverage of at least 70 percent and identity of at least 80 percent as a starting point, while noting that no universal cutoffs exist and that plotting the distribution of these values to find statistical outliers is the better approach. Whole-genome intergenomic similarity tools such as VIRIDIC, which applies ICTV clustering thresholds of 70 percent for genera and 95 percent for species, offer a more systematic route. Conserved start sites among homologs, typically preceded by a Shine-Dalgarno ribosome-binding site spaced 3 to 7 bases upstream, provide further checks.

Tip six warns that automated tools routinely overlook features that are characteristic of phages: embedded coding sequences, programmed ribosomal frameshifting, introns and inteins. Some spanins, for example, encode the outer-membrane spanin entirely within the gene for the inner-membrane spanin. Programmed ribosomal frameshifting, documented in coliphages T7 and lambda, uses a slippery sequence motif that redirects ribosomes into an alternative reading frame, producing an additional longer protein; BLASTx analysis, conservation of the frameshift motif across related phages, and well-mapped reads in the region can support such a prediction, which would otherwise be dismissed as a sequencing or assembly error. Introns and inteins, frequent among some crAss-like phages, can be detected with tools such as PRFect, and hits matching different regions of a single query protein, or a mismatch between open reading frame length and expected protein size, signal their presence.

Tips seven through ten concern functional assignments, where the authors are at their most emphatic. They urge annotators to avoid non-descriptive labels such as arbitrary numbers or generic terms like gp##, which refer to different functions in different phages, and to follow ICTV naming guidelines and controlled vocabularies from UniProtKB and ViralZone. Because any functional assignment will be propagated to all homologs, they recommend thinking globally: never embed a phage’s name or geographic origin in a protein’s functional role. Most phage genes have no database homologs, so it is entirely normal for most coding sequences to be labeled hypothetical proteins; the authors explicitly reject the temptation to inflate functional assignments without evidence, a practice they describe as muddling databases with misleading annotations. Conservative use of BLASTp, with attention to protein length, E-values below 1 times 10 to the minus 5, alignment coverage and conserved domains, is advised, and if confidence is lacking, the label hypothetical is the safer choice. Gene synteny adds another layer of evidence, since phage genomes are organized into functional modules and operons; tools such as progressiveMauve, clinker, LoVis4u and Phynteny, along with the PHROGS database, exploit conserved gene order to improve assignments, particularly for hypothetical proteins.

The final tips address therapeutic safety, structural prediction and cross-verification. Phages intended for therapy should lack virulence-associated toxin genes and antimicrobial resistance proteins, and current protocols require them to be strictly lytic. The authors caution that clear plaques do not settle the question: lysogeny markers such as integrases, repressors and excisionases may be non-functional remnants inherited from temperate ancestors, and some temperate phages require host-specific factors such as attB attachment sites to lysogenize. Evidence of gene decay, including premature stop codons, frameshift indels and loss of conserved domains, should be sought, and diversity-generating retroelements, which can unpredictably alter phage-host interactions, should be screened for. Structure prediction has transformed the field, with AlphaFold-based tools such as Phold improving annotation of head, neck, baseplate and tail proteins, while topology predictors like DeepTMHMM and SignalP 6.0 identify transmembrane and signal peptide features in evolutionarily distant proteins. The closing message is redundancy: integrate outputs from tools built on different algorithms, including InterProScan, HHpred, UniProt BLAST and Foldseek, but never rely on a single tool. The authors provide a summary table of tools, an annotator’s checklist and an interactive HTML guide, and they close by inviting researchers to add their own tips and build their own phage annotation workflows.

Subject of Research: Guidelines for accurate bacteriophage genome annotation

Article Title: Thirteen quick tips for bacteriophage genome annotation

Article References: Zaki, B. M., Bouras, G., Kropinski, A. M., Wattam, A. R., Edwards, R. A., & Aziz, R. K. (2026). Thirteen quick tips for bacteriophage genome annotation. PLOS Computational Biology, 22(10), e1014887. https://doi.org/10.1371/journal.pcbi.1014887

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014887

Keywords: bacteriophages, genome annotation, bioinformatics, Caudoviricetes, gene calling, phage therapy, AlphaFold, sequencing, hypothetical proteins, database quality, gene synteny, PLoS Computational Biology