rewirebio.iobenchmarks
Configuration

ProtBERT (frozen embeddings, no trained head)

ProtBERT embeddings used without any trained parameters, scored directly by cosine similarity between embeddings or clustered as whole-genome vectors.

7 evaluations · 7 results

Overview

An explanatory profile has not yet been reviewed for this record.

Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.

Model Capabilities & Hardware Requirements

Deployment facts recorded for ProtBERT (frozen embeddings, no trained head). Fields the record does not carry are marked not recorded.

ParametersNot recorded
Minimum VRAMNot recorded
Inference LatencyNot recorded
Context WindowNot recorded
AccessNot recorded
LicenseNot recorded

Evaluated Checkpoint: As benchmarked in version 3 of the preprint

Evaluations and results

7 evaluations · 7 results. Different protocols are not a single leaderboard.

Sorted by 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) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.

Tested configurationProtocol and datasetAdjusted 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)AUPRC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)AUROC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)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)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)
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench strain clustering, adjusted Rand index, BacBench strain clustering dataset of 6,071 MGnify MAGs0.720 ± 0.028Not reportedNot reportedNot reportedNot reportedNot reportedNot reported
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench operon identification, zero-shot AUPRC, BacBench operon identification dataset from long-read RNA sequencingNot reported47.58 ± 27.98Not reportedNot reportedNot reportedNot reportedNot reported
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench operon identification, zero-shot AUROC, BacBench operon identification dataset from long-read RNA sequencingNot reportedNot reportedNot reported75.11 ± 8.65Not reportedNot reportedNot reported
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench protein-protein interaction, zero-shot AUPRC, BacBench protein-protein interaction dataset from STRING DBNot reportedNot reported18.38 ± 8.86Not reportedNot reportedNot reportedNot reported
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench protein-protein interaction, zero-shot AUROC, BacBench protein-protein interaction dataset from STRING DBNot reportedNot reportedNot reportedNot reported50.69 ± 4.11Not reportedNot reported
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench strain clustering, average silhouette width, BacBench strain clustering dataset of 6,071 MGnify MAGsNot reportedNot reportedNot reportedNot reportedNot reportedNot reported0.326 ± 0.037
ProtBERT (frozen embeddings, no trained head)EvaluationBacBench strain clustering, normalised mutual information, BacBench strain clustering dataset of 6,071 MGnify MAGsNot reportedNot reportedNot reportedNot reportedNot reported0.866 ± 0.012Not reported
All 7 result rows with coverage, uncertainty and sources
Filter evaluations

Applied filters: All linked evaluations

Exact evaluated configurations and original reported results
Tested configurationProtocol and datasetFindingEvidence and details
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench operon identification, zero-shot AUPRC
Dataset: BacBench operon identification dataset from long-read RNA sequencing
47.58 ± 27.98 AUPRC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)
unitless · higher

Uncertainty: SD 27.98

Coverage: Not reported scored / Not reported eligible

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

ProtBERT on BacBench operon identification, zero-shot AUPRC

Zero-shot. For each pair of contiguous genes the score is the cosine similarity between the two gene embeddings multiplied by an indicator that both genes are on the same strand; the intergenic-distance term of the Bacformer operon model is deliberately left out so that the comparison is about the embeddings. Performance is averaged across all five sequenced strains.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S1, row 'ProtBERT', column 'AUPRC'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench operon identification, zero-shot AUROC
Dataset: BacBench operon identification dataset from long-read RNA sequencing
75.11 ± 8.65 AUROC (zero-shot operon boundary prediction scored by embedding cosine similarity times a same-strand indicator, mean over the five sequenced strains)
unitless · higher

Uncertainty: SD 8.65

Coverage: Not reported scored / Not reported eligible

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

ProtBERT on BacBench operon identification, zero-shot AUROC

Zero-shot. For each pair of contiguous genes the score is the cosine similarity between the two gene embeddings multiplied by an indicator that both genes are on the same strand; the intergenic-distance term of the Bacformer operon model is deliberately left out so that the comparison is about the embeddings. Performance is averaged across all five sequenced strains.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S1, row 'ProtBERT', column 'AUROC'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUPRC
Dataset: BacBench protein-protein interaction dataset from STRING DB
18.38 ± 8.86 AUPRC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 8.86

Coverage: Not reported scored / Not reported eligible

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

ProtBERT 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 'ProtBERT', column 'Zero-shot AUPRC'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench protein-protein interaction, zero-shot AUROC
Dataset: BacBench protein-protein interaction dataset from STRING DB
50.69 ± 4.11 AUROC (zero-shot pair scoring by embedding cosine similarity, genus-level split, mean over held-out test genomes)
unitless · higher

Uncertainty: SD 4.11

Coverage: Not reported scored / Not reported eligible

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

ProtBERT 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 'ProtBERT', column 'Zero-shot AUROC'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench strain clustering, adjusted Rand index
Dataset: BacBench strain clustering dataset of 6,071 MGnify MAGs
0.720 ± 0.028 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)
unitless · higher

Uncertainty: SD 0.028

Coverage: Not reported scored / Not reported eligible

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

ProtBERT on BacBench strain clustering, adjusted Rand index

Final-layer hidden states are mean-pooled into one vector per genome and clustered with Leiden over a grid of resolutions 0.1, 0.25, 0.5 and 1.0 and neighbour counts 5, 10 and 15; the combination that maximises the mean of the adjusted Rand index, normalised mutual information and average silhouette width at the species level is retained. Clustering is run with Scanpy.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S12, row 'ProtBERT', column 'ARI'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench strain clustering, average silhouette width
Dataset: BacBench strain clustering dataset of 6,071 MGnify MAGs
0.326 ± 0.037 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)
unitless · higher

Uncertainty: SD 0.037

Coverage: Not reported scored / Not reported eligible

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

ProtBERT on BacBench strain clustering, average silhouette width

Final-layer hidden states are mean-pooled into one vector per genome and clustered with Leiden over a grid of resolutions 0.1, 0.25, 0.5 and 1.0 and neighbour counts 5, 10 and 15; the combination that maximises the mean of the adjusted Rand index, normalised mutual information and average silhouette width at the species level is retained. Clustering is run with Scanpy.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S12, row 'ProtBERT', column 'ASW'
Configuration: ProtBERT (frozen embeddings, no trained head)Protocol: BacBench strain clustering, normalised mutual information
Dataset: BacBench strain clustering dataset of 6,071 MGnify MAGs
0.866 ± 0.012 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)
unitless · higher

Uncertainty: SD 0.012

Coverage: Not reported scored / Not reported eligible

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

ProtBERT on BacBench strain clustering, normalised mutual information

Final-layer hidden states are mean-pooled into one vector per genome and clustered with Leiden over a grid of resolutions 0.1, 0.25, 0.5 and 1.0 and neighbour counts 5, 10 and 15; the combination that maximises the mean of the adjusted Rand index, normalised mutual information and average silhouette width at the species level is retained. Clustering is run with Scanpy.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S12, row 'ProtBERT', column 'NMI'

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

Use this configuration

How it works, versions and access

Additional methodological details, versions and access terms have not yet been extracted into a structured profile for this configuration. Consult the linked sources and evaluation records below.

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.

0 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

No evidence rows match these filters. Choose another scope or clear the search.

Sources and history

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

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

Technical metadata for this record is tracked under stable ID bacformer-20261115-config-protbert-zero-shot.

Source locator: Table S9, row 'ProtBERT'; Methods 'Benchmarking' and 'General modelling details'

Related records

How to cite

ProtBERT (frozen embeddings, no trained head). Rewire biological model benchmark database, release 2026-11-21-c8bea25e5abb. https://benchmarks.rewirebio.io/database/configuration/bacformer-20261115-config-protbert-zero-shot/

Release 2026-11-21-c8bea25e5abb. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.

The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence

Suggest a correction