rewirebio.iobenchmarks
Source

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19 · Preprint

DOI: 10.1101/2025.07.20.665723Original source

What this source contributed

Records in this release that cite this source. A record can cite several sources.

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Model or configurationMetricValueDatasetEvaluation
Bacformer Large (frozen embeddings with a trained linear head)AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)76.14 ± 19.59 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsBacformer Large on BacBench antibiotic resistance, binary linear probe AUPRC
Bacformer Large (frozen embeddings with a trained linear head)AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)92.15 ± 4.20 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsBacformer Large on BacBench antibiotic resistance, binary linear probe AUROC
Bacformer Large (frozen embeddings with a trained linear head)Pearson (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)65.73 ± 12.88 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsBacformer Large on BacBench antibiotic resistance, MIC regression Pearson correlation
Bacformer Large (frozen embeddings with a trained linear head)R² (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)42.59 ± 19.54 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsBacformer Large on BacBench antibiotic resistance, MIC regression R squared
Bacformer Large (frozen embeddings with a trained linear head)AUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)55.87 ± 21.38 unitlessBacBench essential gene dataset from the Database of Essential GenesBacformer Large on BacBench essential gene prediction, linear probe AUPRC
Bacformer Large (frozen embeddings with a trained linear head)AUROC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)77.97 ± 11.16 unitlessBacBench essential gene dataset from the Database of Essential GenesBacformer Large on BacBench essential gene prediction, linear probe AUROC
Bacformer Large (frozen embeddings, no trained head)AUPRC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)49.98 ± 26.92 unitlessBacBench operon identification dataset from long-read RNA sequencingBacformer Large on BacBench operon identification, zero-shot AUPRC
Bacformer Large (frozen embeddings, no trained head)AUROC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)77.00 ± 7.46 unitlessBacBench operon identification dataset from long-read RNA sequencingBacformer Large on BacBench operon identification, zero-shot AUROC
Bacformer Large (frozen embeddings with a trained linear head)AUPRC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)48.34 ± 36.39 unitlessBacBench phenotypic trait datasetBacformer Large on BacBench phenotypic trait prediction, linear probe AUPRC
Bacformer Large (frozen embeddings with a trained linear head)AUROC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)73.33 ± 12.13 unitlessBacBench phenotypic trait datasetBacformer Large on BacBench phenotypic trait prediction, linear probe AUROC
Bacformer Large (frozen embeddings with a trained linear head)AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)50.84 ± 10.34 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer Large on BacBench protein-protein interaction, linear probe AUPRC
Bacformer Large (frozen embeddings with a trained linear head)AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)78.35 ± 2.40 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer Large on BacBench protein-protein interaction, linear probe AUROC
Bacformer Large (frozen embeddings, no trained head)AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)37.51 ± 15.20 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer Large on BacBench protein-protein interaction, zero-shot AUPRC
Bacformer Large (frozen embeddings, no trained head)AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)72.09 ± 6.44 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer Large on BacBench protein-protein interaction, zero-shot AUROC
Bacformer Large genome embeddings for strain clustering (checkpoint not resolved)Adjusted rand index (Leiden clustering of mean-pooled genome embeddings scored against species labels, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.970 ± 0.017 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer Large on BacBench strain clustering, adjusted Rand index
Bacformer Large genome embeddings for strain clustering (checkpoint not resolved)Silhouette score (average silhouette width of the Leiden clustering of mean-pooled genome embeddings, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.460 ± 0.031 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer Large on BacBench strain clustering, average silhouette width
Bacformer Large genome embeddings for strain clustering (checkpoint not resolved)Normalized mutual information (Leiden clustering of mean-pooled genome embeddings scored against species labels, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.985 ± 0.006 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer Large on BacBench strain clustering, normalised mutual information
Bacformer (frozen embeddings with a trained linear head)AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)67.97 ± 20.73 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsBacformer on BacBench antibiotic resistance, binary linear probe AUPRC
Bacformer (frozen embeddings with a trained linear head)AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)87.61 ± 5.68 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsBacformer on BacBench antibiotic resistance, binary linear probe AUROC
Bacformer (frozen embeddings with a trained linear head)Pearson (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)57.53 ± 15.43 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsBacformer on BacBench antibiotic resistance, MIC regression Pearson correlation
Bacformer (frozen embeddings with a trained linear head)R² (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)33.84 ± 18.60 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsBacformer on BacBench antibiotic resistance, MIC regression R squared
Bacformer (frozen embeddings with a trained linear head)AUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)50.09 ± 22.00 unitlessBacBench essential gene dataset from the Database of Essential GenesBacformer on BacBench essential gene prediction, linear probe AUPRC
Bacformer (frozen embeddings with a trained linear head)AUROC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)79.49 ± 10.48 unitlessBacBench essential gene dataset from the Database of Essential GenesBacformer on BacBench essential gene prediction, linear probe AUROC
Bacformer (frozen embeddings, no trained head)AUPRC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)50.31 ± 26.61 unitlessBacBench operon identification dataset from long-read RNA sequencingBacformer on BacBench operon identification, zero-shot AUPRC
Bacformer (frozen embeddings, no trained head)AUROC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)77.59 ± 6.64 unitlessBacBench operon identification dataset from long-read RNA sequencingBacformer on BacBench operon identification, zero-shot AUROC
Bacformer (frozen embeddings with a trained linear head)AUPRC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)47.60 ± 36.13 unitlessBacBench phenotypic trait datasetBacformer on BacBench phenotypic trait prediction, linear probe AUPRC
Bacformer (frozen embeddings with a trained linear head)AUROC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)71.35 ± 13.01 unitlessBacBench phenotypic trait datasetBacformer on BacBench phenotypic trait prediction, linear probe AUROC
Bacformer (frozen embeddings with a trained linear head)AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)45.73 ± 9.29 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer on BacBench protein-protein interaction, linear probe AUPRC
Bacformer (frozen embeddings with a trained linear head)AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)73.79 ± 4.69 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer on BacBench protein-protein interaction, linear probe AUROC
Bacformer (frozen embeddings, no trained head)AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)31.43 ± 12.61 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer on BacBench protein-protein interaction, zero-shot AUPRC
Bacformer (frozen embeddings, no trained head)AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)64.93 ± 4.96 unitlessBacBench protein-protein interaction dataset from STRING DBBacformer on BacBench protein-protein interaction, zero-shot AUROC
Bacformer genome embeddings for strain clustering (checkpoint not resolved)Adjusted rand index (Leiden clustering of mean-pooled genome embeddings scored against species labels, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.857 ± 0.035 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer on BacBench strain clustering, adjusted Rand index
Bacformer genome embeddings for strain clustering (checkpoint not resolved)Silhouette score (average silhouette width of the Leiden clustering of mean-pooled genome embeddings, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.250 ± 0.022 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer on BacBench strain clustering, average silhouette width
Bacformer genome embeddings for strain clustering (checkpoint not resolved)Normalized mutual information (Leiden clustering of mean-pooled genome embeddings scored against species labels, at the resolution and neighbour count that maximise the mean of the three reported metrics)0.924 ± 0.015 unitlessBacBench strain clustering dataset of 6,071 MGnify MAGsBacformer on BacBench strain clustering, normalised mutual information
DNABERT-2 (frozen embeddings with a trained linear head)AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)52.35 ± 23.55 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsDNABERT-2 on BacBench antibiotic resistance, binary linear probe AUPRC
DNABERT-2 (frozen embeddings with a trained linear head)AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)79.88 ± 7.26 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsDNABERT-2 on BacBench antibiotic resistance, binary linear probe AUROC
DNABERT-2 (frozen embeddings with a trained linear head)Pearson (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)45.89 ± 19.27 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsDNABERT-2 on BacBench antibiotic resistance, MIC regression Pearson correlation
DNABERT-2 (frozen embeddings with a trained linear head)R² (linear head per drug on frozen genome embeddings predicting log(1+MIC), random 70/10/20 split, mean over 56 antibiotics, printed as a percentage)23.61 ± 18.33 unitlessAntibiotic resistance MIC regression subset, 56 antibioticsDNABERT-2 on BacBench antibiotic resistance, MIC regression R squared
DNABERT-2 (frozen embeddings with a trained linear head)AUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)30.21 ± 19.81 unitlessBacBench essential gene dataset from the Database of Essential GenesDNABERT-2 on BacBench essential gene prediction, linear probe AUPRC
DNABERT-2 (frozen embeddings with a trained linear head)AUROC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)64.97 ± 8.29 unitlessBacBench essential gene dataset from the Database of Essential GenesDNABERT-2 on BacBench essential gene prediction, linear probe AUROC
DNABERT-2 (frozen embeddings, no trained head)AUPRC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)47.53 ± 30.07 unitlessBacBench operon identification dataset from long-read RNA sequencingDNABERT-2 on BacBench operon identification, zero-shot AUPRC
DNABERT-2 (frozen embeddings, no trained head)AUROC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)74.13 ± 10.14 unitlessBacBench operon identification dataset from long-read RNA sequencingDNABERT-2 on BacBench operon identification, zero-shot AUROC
DNABERT-2 (frozen embeddings with a trained linear head)AUPRC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)41.93 ± 34.31 unitlessBacBench phenotypic trait datasetDNABERT-2 on BacBench phenotypic trait prediction, linear probe AUPRC
DNABERT-2 (frozen embeddings with a trained linear head)AUROC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)63.81 ± 11.15 unitlessBacBench phenotypic trait datasetDNABERT-2 on BacBench phenotypic trait prediction, linear probe AUROC
DNABERT-2 (frozen embeddings with a trained linear head)AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)28.30 ± 9.55 unitlessBacBench protein-protein interaction dataset from STRING DBDNABERT-2 on BacBench protein-protein interaction, linear probe AUPRC
DNABERT-2 (frozen embeddings with a trained linear head)AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)61.64 ± 3.58 unitlessBacBench protein-protein interaction dataset from STRING DBDNABERT-2 on BacBench protein-protein interaction, linear probe AUROC
DNABERT-2 (frozen embeddings, no trained head)AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)18.76 ± 8.84 unitlessBacBench protein-protein interaction dataset from STRING DBDNABERT-2 on BacBench protein-protein interaction, zero-shot AUPRC
DNABERT-2 (frozen embeddings, no trained head)AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)53.25 ± 4.93 unitlessBacBench protein-protein interaction dataset from STRING DBDNABERT-2 on BacBench protein-protein interaction, zero-shot AUROC
ESM-2 (frozen embeddings with a trained linear head)AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)62.04 ± 23.64 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsESM-2 on BacBench antibiotic resistance, binary linear probe AUPRC
ESM-2 (frozen embeddings with a trained linear head)AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)85.05 ± 6.63 unitlessAntibiotic resistance binary prediction subset, 37 antibioticsESM-2 on BacBench antibiotic resistance, binary linear probe AUROC

Claims

Statements this source supports, each with where in the source it was checked.

  • Apache-2.0 covers the Bacformer code and checkpoints: Licence Apache License 2.0 for the code and models in the Bacformer repository (Preprint 'Code and data availability': 'publicly available at https://github.com/macwiatrak/Bacformer under a permissive Apache 2.0 license'; LICENSE file at the pinned commit; model card front matter 'license: apache-2.0')
  • Bacformer base parameter count printed in preprint Table S9: Parameter count 27M (Table S9, row 'Bacformer', column 'Params')
  • Bacformer base parameter count printed in preprint Table S11: Parameter count 26M (Table S11, row 'Bacformer (26M)')
  • Bacformer context limits in proteins and amino acids: Context limits At most 6,000 proteins per genome, which the authors state covers more than 98% of the genomes in the training corpus, and at most 1,024 amino acids per protein (Hugging Face model card for macwiatrak/bacformer-masked-MAG, 'Preprocessing' and 'Pretraining details'; preprint Table S10 footnote)
  • Bacformer Large parameter count printed in the preprint: Parameter count 332M (Table S9, row 'Bacformer Large', column 'Params'; Table S11, row 'Bacformer Large (332M)')
  • Bacformer pretraining corpus: Training data A MAG corpus of more than 1.3 million metagenome-assembled genomes, 3 billion proteins and 225 million contigs, collected across about 25,000 species and 70 environments from MGnify and SPIRE (Preprint Figure 1 legend, panel A; Methods 'Data'; model card 'Training Data')
  • Bacformer's authors also built and ran the benchmark: Independence The same group built Bacformer, built BacBench, curated every task dataset and ran all twelve benchmarked models, including the comparators (Preprint 'Author Contributions' and Methods 'Benchmarking'; BacBench README, which is in the same author's namespace)
  • Base protein language model of Bacformer: Base protein language model ESM-2 t12 35M (facebook/esm2_t12_35M_UR50D), whose 480-dimensional mean-pooled per-protein embeddings are the input tokens (Bacformer README 'Requirements'; model card 'Architecture'; preprint Table S9 and Methods 'Bacformer')
  • Base protein language model of Bacformer Large, as the cards name it: Base protein language model ESM-C 300M (EvolutionaryScale/esmc-300m-2024-12), with 960-dimensional per-protein embeddings (Hugging Face model cards for the Bacformer Large checkpoints, 'Architecture'; preprint Table S9 row 'ESM-C' checkpoint 'esmc_300m')
  • Clustering hyperparameters were selected on the reported metrics: Selection bias The Leiden resolution and neighbour count were chosen as the combination that maximised the mean of the adjusted Rand index, normalised mutual information and average silhouette width at the species level, on the same genomes the table reports (Preprint Methods 'Strain clustering': 'We then retained the combination of resolution and number of neighbours that maximised the mean of the three metrics described above (ARI, NMI and ASW) at the species level')
  • Model table differences between the BacBench README and the preprint: Model table differences The BacBench README and preprint Table S9 disagree on Nucleotide Transformer (250M against 235M), ProkBERT (27M against 26M), ESM-C (300M against 332M), ProtBert (420M against 412M), Bacformer Large (27M against 332M) and on the ESM-C context length (2,048 against 1,024 tokens) (BacBench README 'Benchmarked models' table against preprint Table S9)
  • Operon labels come from the authors' own long-read RNA sequencing: Label provenance Operon annotations come from long-read RNA sequencing triplicates of five strains, mapped to annotated genes with minimap2 and SAMtools and merged across replicates; the preprint calls the sequencing proprietary while publishing the derived annotations on Hugging Face (Preprint Methods 'Zero-shot operon prediction using long-read RNA sequencing' and 'Benchmarking'; 'Code and data availability')
  • Preprint versions 2 and 3 print different values: Version differences Version 3 (posted 17 August 2026) reports Bacformer Large, which version 2 (posted 13 August 2025) does not mention anywhere; the two versions are separate reports of the same study and their numbers must not be pooled. Where the two were compared, on the genus-level nearest-neighbour retrieval figure of almost 95% balanced accuracy and the 0.81 plus or minus 0.07 median essentiality AUROC, both versions print the same figures (Version 3 Results sections against version 2 of the same preprint)
  • The phenotypic trait table and Methods disagree on the number of repeats: Aggregation conflict Table S16 says its values are the mean and standard deviation across 5 seeds, while the Methods say the per-trait split was repeated across three independent runs with different data splits (Preprint Table S16 caption against Methods 'Phenotypic traits prediction')
  • The released PPI checkpoint has no model card content: Documentation The checkpoint resolves and serves weights, but its card is the auto-generated transformers template, with the model description, licence and training data all left as 'More Information Needed' (Hugging Face model card for macwiatrak/bacformer-for-ppi-prediction, whole card)
  • What Bacformer takes as input and returns: Inputs and outputs Inputs are the protein sequences encoded by a genome, embedded one by one with a base protein language model and ordered by their coordinates on the chromosome and plasmids; outputs are contextual protein embeddings, a mean-pooled genome embedding, and protein-family predictions from the pretraining head (Bacformer README opening paragraph and 'Usage'; model card 'Architecture'; preprint Methods 'Bacformer')
  • Where each BacBench task's printed results are: Result coverage All five active tasks have printed numeric results in the supplementary tables of version 3 of the Bacformer preprint: operon identification in Table S1, protein-protein interaction in Table S13, essential genes in Table S14, phenotypic traits in Table S16 and antibiotic resistance in Table S18; the deprecated strain clustering task is in Table S12 and a protein function prediction task that BacBench does not list is in Table S15 (Preprint Tables S1, S12, S13, S14, S15, S16 and S18)

Models, benchmarks and other records

Models (4)

Configurations (26)

Benchmarks (1)

Tasks (6)

Protocols (17)

Datasets (6)

Dataset subsets (2)

Evaluations (186)

Sources (12)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

33 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-adc15377fab2
Property and statementOriginal source and locationReview and provenance
Archive sha256
0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Article licence
CC-BY 4.0 International, stated in the banner on every page
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.article_licence

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Artifact SHA-256
d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Doi
10.1101/2025.07.20.665723
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.doi

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Extraction method
pdftotext -layout text layer, parsed by extract/parse_tables.py and rechecked by extract/check_tables.py
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.extraction_method

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Hash scope
The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.hash_scope

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Licence
CC BY 4.0
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Limitations
4 values
  • Preprint; not certified by peer review.
  • Version 2 (posted 13 August 2025) does not include Bacformer Large; no value in this batch comes from it.
  • The tables carry no per-cell counts for some aggregations, so several denominators stay unreported.
  • bioRxiv regenerates this PDF on request, so its artifact_sha256 is one download's hash and no refetch reproduces it. The stable identity is the pdftotext -layout text layer, SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2.
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Media type
application/pdf
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.media_type

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

Publication status
preprint
Source metadata
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

No field-specific location recorded

Version: Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Retrieved: 2026-11-15T22:55:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd

Hash scope: The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.

How to read the evidence

Sources and history

Release 2026-11-21-adc15377fab2 · Record review: source checked

Preprint describing Bacformer, a contextualised protein language model of bacterial genomes, and its evaluation against other genomic language models on the BacBench tasks.

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: bacformer-20261115-source-preprint-v3

areas
microbes-communities
contexts
research
archive sha256
0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa
article licence
CC-BY 4.0 International, stated in the banner on every page
artifact sha256
d86de3315063e038ec8d5bed60b7b1a916c137ee60973e55d9d5c82f078d8dfd
doi
10.1101/2025.07.20.665723
extraction method
pdftotext -layout text layer, parsed by extract/parse_tables.py and rechecked by extract/check_tables.py
hash scope
The hash is of one download. bioRxiv regenerates this PDF on request, so the container bytes differ between downloads: a second download minutes later gave 08fcc83af256845d942345fc6b63c64911878733acd028edab5e9307d63319db. The pdftotext -layout text layer of both downloads is byte-identical (SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2), and that text layer, archived gzipped in artifacts/ with SHA-256 0afdefd78a410d036d1c4f17e33bdf481e8d76241ee02f4c84c7c4be91ee72aa, is what every extracted value was read from.
licence
CC BY 4.0
limitations
Preprint; not certified by peer review.; Version 2 (posted 13 August 2025) does not include Bacformer Large; no value in this batch comes from it.; The tables carry no per-cell counts for some aggregations, so several denominators stay unreported.; bioRxiv regenerates this PDF on request, so its artifact_sha256 is one download's hash and no refetch reproduces it. The stable identity is the pdftotext -layout text layer, SHA-256 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2.
media type
application/pdf
publication status
preprint
retrieval note
The issue reported HTTP 403 for a version 2 retrieval during discovery. On 2026-11-15 the bioRxiv API listed versions 1 (2025-07-20), 2 (2025-08-13) and 3 (2026-08-19); the version 3 PDF was served over plain curl without a 403. Europe PMC also serves this study as PPR1052405, whose full-text XML carries article-version number 3 but holds no tables.
retrieved at
2026-11-15T22:55:00Z
review
artifact sha256: 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2; method: transcription; pdf-text-parse; source-hash-verification; ai-assisted-source-review; method note: Independent review pass. Each source artifact was retrieved again at review time and its SHA-256 compared with the source record; the quoted statement was then read at the recorded locator and compared with the recorded value. Access states were rechecked rather than taken from notes.md.; note: Statement re-read against the version 3 preprint. The PDF's container bytes differ on every download, so the artifact checked is its pdftotext -layout text layer, which two fresh downloads reproduce byte for byte and which the archived copy in artifacts/ decompresses to.; retrieval url: https://www.biorxiv.org/content/10.1101/2025.07.20.665723v3.full.pdf; reviewed at: 2026-11-15T23:55:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent, independent of the agent that extracted this batch; no human review claimed. Transcription was checked against the pinned sources. Nothing was reproduced and no value was recomputed.
review artifact
artifacts/biorxiv-665723v3-text-layer.txt.gz, whose contents hash to 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2
review artifact matches original artifact sha256
false
source locator
Supplementary Tables S1, S9, S10, S11, S12, S13, S14, S15, S16, S17 and S18; Methods sections 'Downstream tasks & analysis' and 'Code and data availability'
url
https://www.biorxiv.org/content/10.1101/2025.07.20.665723v3.full.pdf
version
Version 3, posted 17 August 2026 according to the page banner; the bioRxiv API lists version 3 with date 2026-08-19
Related records

Suggest a correction