GPTomics/bio-genome-annotation-prokaryotic-annotation

Version Compatibility

Reference examples tested with: Bakta 1.9+, Prokka 1.14.6 (legacy), tRNAscan-SE 2.0+, CheckM2 1.0+, gffutils 0.12+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags
• Python: pip show <package> then help(module.function) to check signatures

Annotation content tracks the database version, not just the binary: a Bakta full DB (schema v5+, ~30 GB zipped) and a Bakta light DB give different functional calls, and two runs months apart can differ purely from DB updates. Record the Bakta DB version in methods. Prokka's bundled databases are frozen (~2019-2021). If code throws an error, introspect the installed tool and adapt rather than retrying.

Prokaryotic Genome Annotation

"Annotate my bacterial genome" -> Call protein-coding genes, tRNAs/rRNAs/ncRNAs, and other features, then assign function by database identity, and emit submission-ready files.

• CLI: bakta --db db/ assembly.fa (default), prokka --outdir out assembly.fa (legacy), NCBI PGAP (submission/RefSeq-grade)

The Single Most Important Modern Insight -- Gene Calling Is Near-Solved; Function Is Not, and Wrong Labels Propagate

For a typical isolate, Prodigal/Pyrodigal recovers >95-99% of true coding genes. The unsolved problems are the start codon (translation initiation site), the small/overlapping/recoded ORFs, and above all the function. Three consequences a postdoc must internalize:

1. A confidently wrong product name is worse than "hypothetical protein." Names assigned by loose homology transfer across distant lineages and self-amplify through databases - there is no mechanism to retract a correction once it spreads, so annotation accuracy has gone down, not up, as sequencing scaled (Salzberg 2019 Genome Biol 20:92). Bakta's design counters this: gene-level identity only via exact (MD5/UniRef100) match, falling back to cluster level, then "hypothetical" rather than guessing. 25-50% hypothetical is healthy for a non-model organism; near 0% means over-confident transfer.

2. Comparing gene counts across tools/versions/DBs is invalid. Tool agreement is "highly dependent on the organism of study" and biased toward model organisms (Dimonaco 2022 Bioinformatics 38:1198) - there is no universal best tool. Differences between two annotations are mostly tool artifacts, not biology. For any collection (pangenomics), re-annotate every assembly from FASTA with one pipeline + one pinned DB; merging published annotations inflates the accessory genome ~10× (Tonkin-Hill 2020 Genome Biol 21:180).

3. Annotation completeness ≠ assembly completeness. A fragmented/contaminated assembly produces truncated partial CDS at every contig break, inflated counts, and missing rRNA operons. Run CheckM2 before trusting any annotation QC number.

Tool Taxonomy

| Tool | Maintained | Database approach | Output | When | |------|-----------|-------------------|--------|------| | Bakta | Yes (active) | Curated, versioned, alignment-free (UniRef + AMRFinderPlus + expert systems) | GFF3/GBFF/EMBL/FASTA/TSV/JSON + plot | Default for new work; reproducible, MAG-aware | | Prokka | Frozen ~2021 | BLAST hierarchy vs frozen UniProt/RefSeq + HMM | GFF3/GBK/FAA + tbl | Legacy only; pangenome pipelines (Roary) expect Prokka GFF | | NCBI PGAP | Yes (NCBI) | RefSeq protein-family models + ProSplign | ASN.1/GenBank, submission-ready | GenBank/RefSeq submission; best pseudogene/frameshift/selenoprotein handling | | DFAST | Yes (DDBJ) | DFAST reference DBs + swappable refs | INSDC files | DDBJ submission; fast, flexible reference swap | | RAST / BV-BRC | Yes | SEED subsystems | GenBank/GFF + subsystems | Subsystem/metabolic framing; web; not for direct INSDC submission |

Prokka uses Aragorn (tRNA) + barrnap (rRNA); Bakta uses tRNAscan-SE 2.0 (tRNA, domain-specific) + Infernal/Rfam (rRNA/ncRNA) + PILER-CR (CRISPR arrays) + DeepSig (signal peptides). Bakta is deliberately conservative on naming, so it reports "hypothetical" where Prokka confidently (sometimes wrongly) names a gene - Bakta looking "less annotated" is it being honest.

Decision Tree by Scenario

| Scenario | Recommended | Why | |----------|-------------|-----| | Routine isolate, reproducible | Bakta, full DB | standardized, versioned, current functional calls | | Submitting to GenBank/RefSeq | PGAP | RefSeq re-annotates with PGAP regardless; best pseudogene handling | | Submitting to DDBJ | DFAST | INSDC-ready, DDBJ-aligned | | MAG / metagenome bin | Bakta --meta + CheckM2 gate | anonymous-mode calling; QC before trust | | Mycoplasma / Mollicutes | Bakta --translation-table 4 | UGA=Trp; table 11 splits genes at every UGA | | Archaeon | Bakta + verify tRNAscan-SE archaeal model; PGAP if N-termini matter | leaderless mRNAs degrade Prodigal TIS | | Inherited Prokka pangenome pipeline | pin Prokka version, or re-call all in Bakta | tool consistency vs current biology | | Subsystem/metabolic view | RAST / BV-BRC | SEED subsystem categories | | Genome not yet assembled / poor QC | -> genome-assembly/assembly-qc | fix assembly before annotating | | AMR for clinical report | -> run AMRFinderPlus/CARD-RGI directly | --organism context, point mutations, partials matter |

Bakta (Default)

# Database (record the version): full ~30 GB for publishable annotation; light for triage/CI
bakta_db download --output db/ --type full

bakta --db db/ --output bakta_out --prefix ecoli_k12 \
    --genus Escherichia --species coli --strain K-12 \
    --locus-tag ECK12 --gram - --complete --threads 16 \
    assembly.fasta

Key flags: --translation-table {11,4,25} (default 11), --gram {+,-,?} (gates DeepSig signal-peptide calls; default ?), --complete (all sequences are finished replicons; enables oriC detection - do NOT use on draft contigs), --meta (metagenome/MAG mode), --compliant (enforce INSDC structure), --keep-contig-headers, --proteins <faa> (trusted-protein transfer). Set --genus/--species from a GTDB-Tk classification, not a guess. Primary outputs: .gff3, .gbff, .faa, .ffn, .fna, .tsv, plus .hypotheticals.tsv and .inference.tsv (open the inference column to ask why a product was assigned).

Prokka (Legacy)

prokka --outdir prokka_out --prefix my_genome --locustag MYORG \
    --genus Escherichia --species coli --cpus 8 --rfam assembly.fasta

Use only for tool-chain consistency with an existing Prokka-based pangenome workflow, and pin the version. Its bundled databases are frozen: a gene family characterized after ~2019 is "hypothetical" in Prokka but named by current Bakta/PGAP, so the same gene flips core/accessory purely on DB vintage.

Coding Density and CDS Extraction with Python

Goal: Load Bakta/Prokka GFF3 into a queryable database and compute coding density, the first sanity number.

Approach: Build a gffutils database, sum CDS lengths, divide by genome length; flag values outside the expected band.

import gffutils

CODING_DENSITY_LOW = 0.85   # <0.85 in a free-living bacterium suggests wrong table, fragmentation, or heavy pseudogenization
CODING_DENSITY_HIGH = 0.93  # >0.93 suggests ORF over-calling (spurious short hypotheticals)

def coding_density(gff_file, genome_length):
    db = gffutils.create_db(gff_file, ':memory:', merge_strategy='merge')
    coding_bp = sum(c.end - c.start + 1 for c in db.features_of_type('CDS'))
    density = coding_bp / genome_length
    if density < CODING_DENSITY_LOW or density > CODING_DENSITY_HIGH:
        print(f'WARNING: coding density {density:.1%} outside expected 85-93%')
    return density

Hard Cases the Caller Gets Wrong

• Wrong genetic code is silent and looks like fragmentation. A Mycoplasma run under table 11 yields anomalously low coding density + many short "hypothetical" fragments because every internal UGA split a gene. Confirm the table from taxonomy (GTDB-Tk), never a guess. Table 4 (UGA=Trp, Mollicutes), table 25 (UGA=Gly, Gracilibacteria/SR1).
• Pseudogene over-calling in reductive genomes is real biology. Mycobacterium leprae has ~1,116 pseudogenes vs ~1,604 intact CDS (Cole 2001 Nature 409:1007) - high pseudogene fraction in a symbiont/host-restricted pathogen (Rickettsia, Sodalis) is a lifestyle signal, not a defect. PGAP (ProSplign aligns through frameshifts, emits /pseudo) is the best automated arbiter.
• Phase variation is not a pseudogene. A frameshift inside a homopolymer/SSR tract in a contingency locus (Neisseria ~65 candidate loci, Haemophilus, Campylobacter poly-G tracts) means the assembled cell was in the OFF phase - do NOT "correct" or polish it away. This confounds with ONT homopolymer indel error: a pseudogene spike is ambiguous between reductive biology, phase variation, and basecalling error; disambiguate with orthogonal (short-read/HiFi) data.
• Programmed frameshifts make one gene look like two. prfB/RF2 (a +1 frameshift at a slippery CTTT + internal SD), dnaX (−1), and IS-element transposases encode one protein across two frames; naive callers emit two short ORFs. PGAP recognizes a curated set.
• Small ORFs (<~50 aa) are systematically missed - callers impose a ~30 aa minimum because short ORFs arise by chance and coding statistics are unreliable there. E. coli K-12 was missing dozens of 16-50 aa proteins. Treat the gene count as a lower bound on the small proteome; for sORF biology use a dedicated caller (smORFer, smORFinder) and, ideally, Ribo-seq (the experimental arbiter of translation).
• Archaea and Actinobacteria use leaderless mRNAs (no 5' UTR, no Shine-Dalgarno) - Prodigal's RBS-first scoring degrades TIS placement; GeneMarkS-2 (and PGAP) model leaderless transcription. Use tRNAscan-SE's archaeal model for archaeal intron-containing tRNAs.

Per-Method Failure Modes

Cross-tool gene-count comparison

Trigger: comparing counts from Bakta vs Prokka vs old RefSeq, or mixed-vintage records. Mechanism: tool/DB differences dominate biological differences. Symptom: "novel genes" or accessory-genome inflation. Fix: re-annotate every genome from FASTA with one pipeline + pinned DB.

Wrong translation table

Trigger: table 11 on a Mollicute. Mechanism: UGA read as stop. Symptom: low coding density, doubled short gene count, high hypothetical fraction. Fix: --translation-table 4; confirm from GTDB-Tk.

Annotating an unvetted assembly

Trigger: running Bakta before CheckM2. Mechanism: contig breaks truncate CDS; contamination mixes organisms. Symptom: partial CDS at ends, inflated/chimeric gene set, missing rRNA operons. Fix: CheckM2 first; contamination >5% -> decontaminate.

Trusting an over-specific product name

Trigger: reading the product column as ground truth. Mechanism: loose-homology transfer below ~40% identity is unreliable, especially for promiscuous folds. Symptom: a "histidine-kinase expansion" that is one over-called domain. Fix: open .inference.tsv; prefer InterPro/Pfam architecture over free-text product when they conflict.

Single-mode calling on a metagenome or tiny replicon

Trigger: Prodigal single mode on a MAG/plasmid/phage. Mechanism: self-training needs a full single-organism genome. Symptom: poor calls on short/mixed input. Fix: Bakta --meta (anonymous mode).

Missing signal peptides on macOS

Trigger: Bakta on macOS. Mechanism: DeepSig was dropped from the default mac conda env (~v1.9.4). Symptom: silently absent signal-peptide calls. Fix: verify the run log; run on Linux if SPs matter.

Quantitative Thresholds

| Threshold | Source | Rationale | |-----------|--------|-----------| | Coding density ~88-90% (band 85-93%) | bacterial genome norm | <85% = wrong table/fragmentation/decay; >93% = over-calling | | ~850-1,000 genes/Mb (~1 gene/kb) | bacterial gene density | far higher = over-call; far lower = under-call/wrong table | | Hypothetical 25-50% | annotation norm | ~0% = over-confident transfer; >60-70% = novel lineage or wrong tool/table | | tRNA count ≥ ~20 (often 40-60) | one isoacceptor set minimum | far fewer = fragmented assembly broke tRNA regions | | rRNA operons ~1-15 (e.g. ~7 in E. coli) | copy-number norm | zero/fractional in a "complete" genome = short-read repeat collapse | | Prodigal minimum ~30 aa / 90 nt | caller default | shorter ORFs need a dedicated sORF caller + Ribo-seq | | CheckM2 contamination ≤5%, completeness ≥90% | MIMAG-aligned | above/below -> annotation QC numbers uninterpretable, fix assembly first |

Common Errors

| Error / symptom | Cause | Solution | |-----------------|-------|----------| | Genes split, low coding density | wrong translation table | --translation-table 4/25; confirm from GTDB-Tk | | Many hypothetical proteins | novel organism / light DB | full Bakta DB; add InterProScan/eggNOG; normal if 25-50% | | Low gene / tRNA count, missing rRNA | fragmented assembly | CheckM2; prefer long-read/complete assembly | | RefSeq record differs from the local GFF | RefSeq is PGAP, GenBank keeps the submitter's annotation | report WP_ accession + locus_tag; expect divergence | | Pseudogene spike | ONT homopolymer indels vs real decay vs phase variation | check homopolymer context; corroborate with short-read/HiFi | | Submission rejected on locus tags | invented prefix | register the prefix via BioSample (3-12 alnum, starts with a letter) |

References

• Schwengers O, et al. 2021. Bakta: rapid and standardized annotation of bacterial genomes via alignment-free sequence identification. Microb Genom 7:000685.
• Seemann T. 2014. Prokka: rapid prokaryotic genome annotation. Bioinformatics 30:2068-2069.
• Hyatt D, et al. 2010. Prodigal: prokaryotic gene recognition and translation initiation site identification. BMC Bioinformatics 11:119.
• Larralde M. 2022. Pyrodigal: Python bindings and interface to Prodigal. J Open Source Softw 7:4296.
• Lomsadze A, et al. 2018. Modeling leaderless transcription and atypical genes results in more accurate gene prediction in prokaryotes (GeneMarkS-2). Genome Res 28:1079-1089.
• Tatusova T, et al. 2016. NCBI prokaryotic genome annotation pipeline. Nucleic Acids Res 44:6614-6624.
• Li W, et al. 2021. RefSeq: expanding the Prokaryotic Genome Annotation Pipeline reach with protein family model curation. Nucleic Acids Res 49:D1020-D1028.
• Tanizawa Y, et al. 2018. DFAST: a flexible prokaryotic genome annotation pipeline for faster genome publication. Bioinformatics 34:1037-1039.
• Chan PP, et al. 2021. tRNAscan-SE 2.0: improved detection and functional classification of transfer RNA genes. Nucleic Acids Res 49:9077-9096.
• Chklovski A, et al. 2023. CheckM2: a rapid, scalable and accurate tool for assessing microbial genome quality using machine learning. Nat Methods 20:1203-1212.
• Dimonaco NJ, et al. 2022. No one tool to rule them all: prokaryotic gene prediction tool annotations are highly dependent on the organism of study. Bioinformatics 38:1198-1207.
• Tonkin-Hill G, et al. 2020. Producing polished prokaryotic pangenomes with the Panaroo pipeline. Genome Biol 21:180.
• Salzberg SL. 2019. Next-generation genome annotation: we still struggle to get it right. Genome Biol 20:92.
• Cole ST, et al. 2001. Massive gene decay in the leprosy bacillus. Nature 409:1007-1011.

Related Skills

• functional-annotation - Add GO/KEGG/Pfam to hypothetical proteins with eggNOG-mapper/InterProScan
• ncrna-annotation - Detailed ncRNA identification with Infernal/Rfam beyond the built-in callers
• annotation-qc - CheckM2 completeness/contamination gate and gene-set sanity metrics
• genome-assembly/assembly-qc - Assess assembly quality before annotation
• genome-intervals/gtf-gff-handling - Parse and manipulate GFF3 output
• comparative-genomics/pangenome-analysis - Uniform re-annotation before pangenome clustering