rewirebio.iobenchmarks
Dataset

BacBench protein-protein interaction dataset from STRING DB

Protein pairs of 261 bacterial genomes sampled from STRING DB across 215 genera, labelled by binarising the STRING combined interaction score at 0.6, with DNA matched from GenBank.

Evaluation results

48 evaluations · 48 results. Different protocols are not a single leaderboard.

Filter evaluations

Applied filters: All linked evaluations

Exact evaluated configurations and original reported results
Tested configurationProtocol and datasetFindingEvidence and details
Configuration: Bacformer Large (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
50.84 ± 10.34 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 10.34

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer Large on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer Large', column 'Finetuned AUPRC'
Configuration: Bacformer Large (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
78.35 ± 2.40 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 2.40

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer Large on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer Large', column 'Finetuned AUROC'
Configuration: Bacformer Large (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
37.51 ± 15.20 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 15.20

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer Large on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer Large', column 'Zero-shot AUPRC'
Configuration: Bacformer Large (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
72.09 ± 6.44 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 6.44

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer Large on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer Large', column 'Zero-shot AUROC'
Configuration: Bacformer (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
45.73 ± 9.29 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.29

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer', column 'Finetuned AUPRC'
Configuration: Bacformer (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
73.79 ± 4.69 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.69

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer', column 'Finetuned AUROC'
Configuration: Bacformer (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
31.43 ± 12.61 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 12.61

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer', column 'Zero-shot AUPRC'
Configuration: Bacformer (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
64.93 ± 4.96 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.96

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Bacformer on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Bacformer', column 'Zero-shot AUROC'
Configuration: DNABERT-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
28.30 ± 9.55 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.55

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

DNABERT-2 on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'DNABERT-2', column 'Finetuned AUPRC'
Configuration: DNABERT-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
61.64 ± 3.58 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 3.58

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

DNABERT-2 on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'DNABERT-2', column 'Finetuned AUROC'
Configuration: DNABERT-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
18.76 ± 8.84 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 8.84

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

DNABERT-2 on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'DNABERT-2', column 'Zero-shot AUPRC'
Configuration: DNABERT-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
53.25 ± 4.93 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.93

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

DNABERT-2 on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'DNABERT-2', column 'Zero-shot AUROC'
Configuration: ESM-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
44.24 ± 8.64 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 8.64

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-2 on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-2', column 'Finetuned AUPRC'
Configuration: ESM-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
72.83 ± 4.53 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.53

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-2 on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-2', column 'Finetuned AUROC'
Configuration: ESM-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
22.08 ± 9.15 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.15

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-2 on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-2', column 'Zero-shot AUPRC'
Configuration: ESM-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
56.94 ± 2.44 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 2.44

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-2 on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-2', column 'Zero-shot AUROC'
Configuration: ESM-C (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
44.75 ± 9.08 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.08

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-C on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-C', column 'Finetuned AUPRC'
Configuration: ESM-C (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
73.20 ± 4.29 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.29

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-C on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-C', column 'Finetuned AUROC'
Configuration: ESM-C (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
20.81 ± 9.24 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.24

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-C on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-C', column 'Zero-shot AUPRC'
Configuration: ESM-C (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
57.44 ± 3.13 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 3.13

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

ESM-C on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'ESM-C', column 'Zero-shot AUROC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
15.98 ± 7.37 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 7.37

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evo-2 on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Evo-2', column 'Finetuned AUPRC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
50.00 ± 0.00 AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 0.00

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evo-2 on BacBench protein-protein interaction, linear probe AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Evo-2', column 'Finetuned AUROC'
Configuration: Evo-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
17.80 ± 9.10 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 9.10

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evo-2 on BacBench protein-protein interaction, zero-shot AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Evo-2', column 'Zero-shot AUPRC'
Configuration: Evo-2 (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
51.52 ± 3.16 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 3.16

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evo-2 on BacBench protein-protein interaction, zero-shot AUROC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Evo-2', column 'Zero-shot AUROC'
Configuration: Evo (frozen embeddings with a trained linear head)Protocol: BacBench protein-protein interaction, linear probe AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
20.93 ± 8.80 AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 8.80

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evo on BacBench protein-protein interaction, linear probe AUPRC

Zero-shot columns score each protein pair by the cosine similarity of the two embeddings. The finetuned columns train a single linear layer, preceded by layer normalisation and dropout of 0.2, on the average of the two frozen protein embeddings together with their cosine similarity, so the input is of dimension D plus one; binary cross-entropy loss, Adam with weight decay 0.01, at most 20 epochs with early stopping patience 5 on AUROC, and a learning-rate sweep per model. Metrics are computed per genome on the test split.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S13, row 'Evo', column 'Finetuned AUPRC'

Source checking is not independent reproduction. Release 2026-11-21-adc15377fab2.

Research readiness

0 of 4 readiness checks met. These checks assess whether the evidence supports a reproducible investigation; a source-checked score alone does not meet them.

Readiness checks, gaps and artifacts

Release 2026-11-21-adc15377fab2 · Evidence verified: Not verified

Evidence incomplete

Replay metrics

Exact outcomes, predictions, identifiers and evaluator are connected.

Missing or unresolved evidence

  • No verified artifact manifest is linked to this exact record.
  • File checksums match the recorded files: not yet verified
  • Predictions are matched to the right samples: not yet verified
  • Score meaning and direction are confirmed: not yet verified
  • Metrics are recomputed from the saved predictions: not yet verified

Verified: Not verified

Evidence incomplete

Investigate discrepancies

Replay evidence includes annotations and an assessment of dependence. Unknown independence permits descriptive analysis only.

Missing or unresolved evidence

  • No verified artifact manifest is linked to this exact record.
  • File checksums match the recorded files: not yet verified
  • Predictions are matched to the right samples: not yet verified
  • Score meaning and direction are confirmed: not yet verified
  • Metrics are recomputed from the saved predictions: not yet verified
  • Sample annotations are recorded: not yet verified
  • Dependence between samples is assessed: not yet verified

Verified: Not verified

Evidence incomplete

Run locally

A pinned recipe describes the inputs, environment and resource requirements.

Missing or unresolved evidence

  • No verified artifact manifest is linked to this exact record.
  • File checksums match the recorded files: not yet verified
  • Predictions are matched to the right samples: not yet verified
  • Score meaning and direction are confirmed: not yet verified
  • A pinned run recipe exists: not yet verified
  • Compute requirements are estimated: not yet verified

Verified: Not verified

Evidence incomplete

Validate independently

Separate data and exposure records support an independent test.

Missing or unresolved evidence

  • No verified artifact manifest is linked to this exact record.
  • File checksums match the recorded files: not yet verified
  • Predictions are matched to the right samples: not yet verified
  • Score meaning and direction are confirmed: not yet verified
  • Independent validation data exist: not yet verified
  • Overlap with training data is checked: not yet verified

Verified: Not verified

Readiness describes the evidence in this release. Availability on your computer is checked separately when an investigation runs. Existing data exposure can prevent independent validation even when files are available.

Artifacts and reproduction

No verified artifact manifest is connected to this record yet. The gaps above identify what is needed before analysis can begin.

Dataset and evaluation context

A dataset supplies biological observations. The evaluation protocol defines how those observations are split, used and scored.

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.

24 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
Access
Published as Hugging Face datasets in protein-sequence and DNA modalities, with a larger protein-only version of more than 10 thousand genomes.
Context-only references
BacBench repository README at commit 64b7d45

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Commit 64b7d45823b2ff2ef88458e3aa235caa461a12aa
Retrieved: 2026-11-15T23:09:36Z

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.access

Source artifact SHA-256: c9d28efef5b4cc30adbcc9de224c350100f5aa2a15a025beaf10dcdeb86cd93e

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Access
Published as Hugging Face datasets in protein-sequence and DNA modalities, with a larger protein-only version of more than 10 thousand genomes.
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.access

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.

Assay
STRING DB combined interaction scores, which pool evidence channels, binarised at a threshold of 0.6.
Context-only references
BacBench repository README at commit 64b7d45

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Commit 64b7d45823b2ff2ef88458e3aa235caa461a12aa
Retrieved: 2026-11-15T23:09:36Z

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.assay

Source artifact SHA-256: c9d28efef5b4cc30adbcc9de224c350100f5aa2a15a025beaf10dcdeb86cd93e

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Assay
STRING DB combined interaction scores, which pool evidence channels, binarised at a threshold of 0.6.
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.assay

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.

Label semantics
A pair is positive when its STRING DB combined score is at least 0.6, which the authors state leaves about 10% of pairs positive; the threshold was chosen on validation AUROC and AUPRC.
Context-only references
BacBench repository README at commit 64b7d45

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Commit 64b7d45823b2ff2ef88458e3aa235caa461a12aa
Retrieved: 2026-11-15T23:09:36Z

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.label_semantics

Source artifact SHA-256: c9d28efef5b4cc30adbcc9de224c350100f5aa2a15a025beaf10dcdeb86cd93e

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Label semantics
A pair is positive when its STRING DB combined score is at least 0.6, which the authors state leaves about 10% of pairs positive; the threshold was chosen on validation AUROC and AUPRC.
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.label_semantics

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.

Population
261 genomes across 215 genera, randomly sampled from the bacterial organisms in STRING DB
Context-only references
BacBench repository README at commit 64b7d45

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Commit 64b7d45823b2ff2ef88458e3aa235caa461a12aa
Retrieved: 2026-11-15T23:09:36Z

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.population

Source artifact SHA-256: c9d28efef5b4cc30adbcc9de224c350100f5aa2a15a025beaf10dcdeb86cd93e

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Population
261 genomes across 215 genera, randomly sampled from the bacterial organisms in STRING DB
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.population

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.

Repository
https://huggingface.co/datasets/macwiatrak/bacbench-ppi-stringdb-protein-sequences-small
Context-only references
BacBench repository README at commit 64b7d45

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

Version: Commit 64b7d45823b2ff2ef88458e3aa235caa461a12aa
Retrieved: 2026-11-15T23:09:36Z

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.repository

Source artifact SHA-256: c9d28efef5b4cc30adbcc9de224c350100f5aa2a15a025beaf10dcdeb86cd93e

Hash scope: Hash scope not separately documented; inspect source record

Inspected artifact

Repository
https://huggingface.co/datasets/macwiatrak/bacbench-ppi-stringdb-protein-sequences-small
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Shared locator for this statement’s cited sources; not a separate locator for each citation.

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

not individually reviewed

No individual claim review recorded

Audit details

Field: attributes.repository

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

2 source records and release historyDownload this release (gzip)
Technical metadata and record identifiers

Technical metadata for this record is tracked under stable ID bacformer-20261115-data-ppi-stringdb.

Source locator: Methods 'Benchmarking' under protein-protein interaction prediction; BacBench README 'Datasets'

Related records

Suggest a correction