Skip to content

baktfold Output Files

Main Outputs

  • .faa which holds all amino acid sequences of CDSs (not only hypotheticals) (like Bakta)
  • .ffn which holds all nucletotide sequences of CDSs (like Bakta)
  • .fna which holds the input genome FASTA (parsed from the JSON) (like Bakta)
  • _3di.fasta which holds all Foldseek 3Di sequences of predicted CDSs as predicted by ProstT5
  • .gbff .gff3 and .embl are GenBank, GFF3 and EMBL format output files (like Bakta) with full genome annotations including the input annotations combined with Baktfold's additional annotations
  • {prefix}.tsv is the same as Bakta's .tsv output, containing all full genome annotations including the input annotations combined with Baktfold's additional annotations
  • .json which is a Bakta JSON format output file with all genome annotations including the input annotations combined with Baktfold's additional annotations
  • *_tophit.tsv which have detailed alignment statistics for the Foldseek output tophits for each of the four consituent Baktfold databases (Swiss-Prot, AFDB clusters, CATH and PDB)
    • For example:
query   target  bitscore    fident  evalue  qStart  qEnd    qLen    qCov    tStart  tEnd    tLen    tCov
MEGJMN_070  AF-A0A1I3V7E0-F1-model_v6   292 0.41    2.619e-06   1   91  93  0.97    1   95  99  0.95

  • There will also be a custom_database_tophit.tsv if you have used --custom-db

  • baktfold.inference.tsv which contains the new annotated funciton (Product column), an Annotation_Confidence column and all annotated tophits (if they exist) for all proteins that were attempted to be annotated by Baktfold

  • Note: This will contain an extra column Custom_DB if you have used --custom-db with a custom database, and extra TMscore and LDDT columns if you ran with user-provided structures (--structure-dir)
    • For example:
ID  Length  Product Annotation_Confidence   Swissprot   AFDBClusters    PDB CATH
MEGJMNBEGN_27   162 HTH-type quorum-sensing regulator RhlR  high    swissprot_P54292    afdbclusters_A0A9E1VSB0 pdb_5l09    cath_3sztB01
MEGJMNBEGN_30   68  hypothetical protein                    
MEGJMNBEGN_70   94  hypothetical protein    low     afdbclusters_A0A1I3V7E0

Annotation confidence

Each Baktfold annotation is labelled high, medium or low confidence (in the Annotation_Confidence column of the TSV and the annotation_confidence field of the JSON — it is intentionally not written to the GFF3/GenBank/EMBL outputs) to make the structural-alignment results more interpretable. The heuristics mirror Phold:

  • High: both query and target have ≥80% reciprocal alignment coverage, and one of (i) >30% amino acid identity (the "light zone" of sequence homology), (ii) mean ProstT5 confidence ≥60%, or (iii) alignment E-value < 1e-10.
  • Medium: either the query or target has ≥80% coverage, and one of (i) >30% amino acid identity or (ii) ProstT5 confidence between 45–60%, and E-value < 1e-5.
  • Low: all other hits below the E-value threshold that do not meet the above (low coverage, low identity, low ProstT5 confidence, or an E-value near the 0.001 threshold).

If Baktfold is run with user-provided structures (--structure-dir) instead of ProstT5, the ProstT5-confidence criteria are dropped and the output additionally includes Foldseek TMscore (template modeling score) and LDDT (local distance difference test) values to help assess quality.

A low annotation is not necessarily wrong — only that it should be trusted less than a high (extremely trustworthy) or medium (very trustworthy) annotation, and likely warrants a bit more manual curation.

  • baktfold.summary.txt which contains basic summary statistics showing how many CDS baktfold was able to annotate
  • CDS beginning hypotheticals is the amount of hypotheticals parsed from the input JSON
  • CDS annotated with Baktfold database hit is the number of CDS that have at least 1 hit to a constituent Baktfold DB
  • CDS annotated with Baktfold function is the number of CDS whose Baktfold tophit has a non-hypothetical function transferred
  • CDS remaining hypotheticals is the number of CDS remaining hypothetical most baktfold
    • For example
Annotation:
CDS count: 2635
CDS beginning hypotheticals: 55
CDS annotated with Baktfold database hit: 12
CDS annotated with Baktfold function: 7
CDS remaining hypotheticals: 48

Baktfold:
Software: v0.2.0
Database: v0.1.0
DOI: https://doi.org/10.64898/2026.03.31.715528
URL: github.com/gbouras13/baktfold

Supplementary Outputs

  • _prostT5_3di_mean_probabilities.csv - contains the mean ProstT5 probability score for each CDS. These are equivalent to the probability of how similar the overall ProstT5 3Di sequence is predicted to be compared to its Alphafold2 baseline
  • _prostT5_3di_all_probabilities.json - contains the ProstT5 probabilities for each residue for each CDS, in the json format

Optional Outputs

  • _embeddings_per_protein.h5 - contains the ProstT5 embeddings for each protein in the h5 format
  • _embeddings_per_residue.h5 - contains the ProstT5 embeddings for each residue in the h5 format

Reconstituting outputs from the JSON

The .json is the authoritative record of a baktfold run. You can regenerate every standard output above from it (without re-running ProstT5 or Foldseek) with baktfold json:

  • Reconstitutable from the .json: .gff3, .gbff, .embl, .tsv, .inference.tsv, .faa, .ffn, .fna, .summary.txt (and a fresh .json).
  • Not reconstitutable (only produced by an actual compare/run): the Foldseek *_tophit.tsv, foldseek_results_*.tsv, _3di.fasta, the ProstT5 probability files and the embedding .h5 files.