rewirebio.iobenchmarks
Evaluation

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

ESM-2 scored on the protocol 'BacBench protein-protein interaction, zero-shot AUROC' in the Bacformer preprint benchmark.

Evaluation results

1 evaluation · 1 result. 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: 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'

Source checking is not independent reproduction. Release 2026-11-21-c8bea25e5abb.

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-c8bea25e5abb · 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.

Evaluation procedure

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.

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
origin
Author-reported evaluation
configuration
As printed in version 3 of the preprint
adaptation
No training; embeddings scored directly
aggregation
Mean and standard deviation across held-out test genomes
budget
About 90 minutes for the zero-shot evaluation of all models and 91 hours for the finetuning including the learning-rate sweep, on one NVIDIA H100 with 96 GB of memory and 72 CPU cores, excluding the cost of embedding the genomes
dataset version
As described in version 3 of the preprint and published on Hugging Face
inputs
Single protein sequence. Each protein embedded on its own, last-layer hidden states mean-pooled over amino acids
metric implementation
zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes
population
261 genomes across 215 genera, randomly sampled from the bacterial organisms in STRING DB
protocol id
bacformer-20261115-protocol-ppi-zero-shot-auroc
split
Phylogeny-aware split by genus, 60/20/20 train, validation and test.

Metadata review: source checked. Unreported conditions prevent automatic comparisons.

Reproduction

Split
Phylogeny-aware split by genus, 60/20/20 train, validation and test.
Adaptation
No training; embeddings scored directly
Scoring implementation
zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes

No execution recipe has been verified for this exact configuration and evaluation. A benchmark's general instructions may use different inputs, splits or model settings.

Reproducing this published result requires matching its model configuration, data, split and scorer. Source checking or a successful smoke test does not establish score reproduction.

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.

19 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Comparison: adaptation
No training; embeddings scored directly
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.adaptation

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.

Comparison: aggregation
Mean and standard deviation across held-out test genomes
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.aggregation

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.

Comparison: budget
About 90 minutes for the zero-shot evaluation of all models and 91 hours for the finetuning including the learning-rate sweep, on one NVIDIA H100 with 96 GB of memory and 72 CPU cores, excluding the cost of embedding the genomes
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.budget

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.

Comparison: dataset version
As described in version 3 of the preprint and published on Hugging Face
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.dataset_version

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.

Comparison: inputs
Single protein sequence. Each protein embedded on its own, last-layer hidden states mean-pooled over amino acids
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.inputs

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.

Comparison: metric implementation
zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.metric_implementation

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.

Comparison: 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 ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.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.

Comparison: protocol id
bacformer-20261115-protocol-ppi-zero-shot-auroc
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.protocol_id

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.

Comparison: split
Phylogeny-aware split by genus, 60/20/20 train, validation and test.
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

Audit details

Field: attributes.comparison.split

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
2 values
  • Reported by the authors of Bacformer, who also built BacBench and ran every benchmarked model on it.
  • Not reproduced by Rewire; the exact splits, seeds and BacBench commit behind the runs are not published.
Context-only references
Bacformer preprint, bioRxiv 2025.07.20.665723 version 3

Original source ↗

Table S13, row 'ESM-2', column 'Zero-shot AUROC'

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

author reported

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.

How to read the evidence

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

1 source record and release historyDownload this release (gzip)
Technical metadata and extraction receipts

Stable ID: bacformer-20261115-eval-ppi-esm-2-zero-shot-auroc

comparison
adaptation: No training; embeddings scored directly; aggregation: Mean and standard deviation across held-out test genomes; budget: About 90 minutes for the zero-shot evaluation of all models and 91 hours for the finetuning including the learning-rate sweep, on one NVIDIA H100 with 96 GB of memory and 72 CPU cores, excluding the cost of embedding the genomes; dataset version: As described in version 3 of the preprint and published on Hugging Face; inputs: Single protein sequence. Each protein embedded on its own, last-layer hidden states mean-pooled over amino acids; metric implementation: zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes; population: 261 genomes across 215 genera, randomly sampled from the bacterial organisms in STRING DB; protocol id: bacformer-20261115-protocol-ppi-zero-shot-auroc; split: Phylogeny-aware split by genus, 60/20/20 train, validation and test.
limitations
Reported by the authors of Bacformer, who also built BacBench and ran every benchmarked model on it.; Not reproduced by Rewire; the exact splits, seeds and BacBench commit behind the runs are not published.
origin
author_reported
protocol
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.
review
artifact sha256: 86636e89b167d7304b3a83ee7f159cad0a9bdbe97019588ce836087b8fe3b0b2; method: independent-cell-check; deterministic-table-parse; pdf-text-parse; source-hash-verification; ai-assisted-source-review; method note: Independent review pass. Every cell of the six benchmark tables was re-read by review/recheck_cells.py, written for this pass, which uses neither of the extractor's scripts and does not read the pdftotext -layout text layer they read. It takes the word bounding boxes pdftotext -bbox-layout emits for a freshly downloaded version 3 PDF, groups words into printed lines by their y coordinate, derives each table's column boundaries from the x extent of that table's own printed header words, and assigns every number to a column by where it sits on the page, so a value printed under one header cannot be read as a value of another. It asserts the printed row order, that every column of every row is filled, that each cell holds exactly a mean and a spread, that each cell prints its own plus-minus sign, and for Tables S13 and S18 that the group header sits geometrically over the two columns this review names after it. All 186 cells agreed on the printed value, the mean, the spread, the spread's printed string and its source column, with no disagreement, and the six tables print no cell the batch leaves unrecorded. Three negative controls (two cells of one row exchanged, one cell taking its neighbour's value, every Table S18 locator moved one column right) were all rejected, and the unmutated batch accepted; the log is review/negative-controls.txt. npm run batch -- check-values reproduced the extractor's figures exactly, 36 in_cell and 150 in_row with nothing else, against a text layer regenerated from a fresh download. A manual sample of 47 cells was then read by eye against the printed table bodies: 24 drawn with random.sample over all 186 at seed 20261115, all 18 cells of Table S12, and the eight Evo and Evo-2 cells of Tables S1 and S14, where the printed row order differs from the other tables. All matched, column assignment included.; note: Evaluation of one configuration on one protocol, carrying the single printed cell of its result. Its system, assessment and data links, its origin and its comparison fields were checked against the source alongside that cell. Protein-protein interaction, Table S13. The zero-shot and finetuned column group was checked geometrically against the printed group header, not inferred from column order.; 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.
source locator
Table S13, row 'ESM-2', column 'Zero-shot AUROC'
version
As printed in version 3 of the preprint
Related records

Suggest a correction