rewirebio.iobenchmarks
Configuration

Evo-2 (frozen embeddings with a trained linear head)

Evo-2 embeddings computed with the backbone frozen and a single trained linear layer on top, which turns the stored embeddings into the task prediction.

10 evaluations · 10 results8 open disagreements between studies involve its model

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 Evo-2 (frozen embeddings with a trained linear 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

10 evaluations · 10 results. Different protocols are not a single leaderboard.

Sorted by AUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes) (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 datasetAUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)AUPRC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)AUPRC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)AUROC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)AUROC (linear head on frozen pair embeddings plus their cosine similarity, genus-level split, mean over held-out test genomes)AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)AUROC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)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)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)
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench essential gene prediction, linear probe AUPRC, BacBench essential gene dataset from the Database of Essential Genes15.37 ± 23.71Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench antibiotic resistance, binary linear probe AUPRC, Antibiotic resistance binary prediction subset, 37 antibioticsNot reportedNot reported38.35 ± 23.08Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench antibiotic resistance, binary linear probe AUROC, Antibiotic resistance binary prediction subset, 37 antibioticsNot reportedNot reportedNot reportedNot reportedNot reportedNot reported66.27 ± 10.19Not reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench antibiotic resistance, MIC regression Pearson correlation, Antibiotic resistance MIC regression subset, 56 antibioticsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported00.53 ± 00.72Not reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench antibiotic resistance, MIC regression R squared, Antibiotic resistance MIC regression subset, 56 antibioticsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported03.95 ± 21.04
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench essential gene prediction, linear probe AUROC, BacBench essential gene dataset from the Database of Essential GenesNot reportedNot reportedNot reportedNot reported56.05 ± 9.46Not reportedNot reportedNot reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench phenotypic trait prediction, linear probe AUPRC, BacBench phenotypic trait datasetNot reportedNot reportedNot reported24.96 ± 20.52Not reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench phenotypic trait prediction, linear probe AUROC, BacBench phenotypic trait datasetNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported55.41 ± 8.85Not reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench protein-protein interaction, linear probe AUPRC, BacBench protein-protein interaction dataset from STRING DBNot reported15.98 ± 7.37Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evo-2 (frozen embeddings with a trained linear head)EvaluationBacBench protein-protein interaction, linear probe AUROC, BacBench protein-protein interaction dataset from STRING DBNot reportedNot reportedNot reportedNot reportedNot reported50.00 ± 0.00Not reportedNot reportedNot reportedNot reported
All 10 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: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench antibiotic resistance, binary linear probe AUPRC
Dataset subset: Antibiotic resistance binary prediction subset, 37 antibiotics
38.35 ± 23.08 AUPRC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)
unitless · higher

Uncertainty: SD 23.08. The caption says the spread is across antibiotics; the Methods state 37 antibiotics in the binary setting and 56 in the regression setting.

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench antibiotic resistance, binary linear probe AUPRC

One linear model per drug on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, binary cross-entropy for the resistant or susceptible setting and mean squared error for the log(1+MIC) regression, at most 100 epochs, early stopping patience 10 on validation AUROC or validation R squared, a learning-rate sweep per model across seeds and drugs, Adam with weight decay 0.01, and three random splits.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S18, row 'Evo-2', column 'Binary AUPRC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench antibiotic resistance, binary linear probe AUROC
Dataset subset: Antibiotic resistance binary prediction subset, 37 antibiotics
66.27 ± 10.19 AUROC (linear head per drug on frozen genome embeddings, random 70/10/20 split, mean over 37 antibiotics)
unitless · higher

Uncertainty: SD 10.19. The caption says the spread is across antibiotics; the Methods state 37 antibiotics in the binary setting and 56 in the regression setting.

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench antibiotic resistance, binary linear probe AUROC

One linear model per drug on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, binary cross-entropy for the resistant or susceptible setting and mean squared error for the log(1+MIC) regression, at most 100 epochs, early stopping patience 10 on validation AUROC or validation R squared, a learning-rate sweep per model across seeds and drugs, Adam with weight decay 0.01, and three random splits.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S18, row 'Evo-2', column 'Binary AUROC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench antibiotic resistance, MIC regression Pearson correlation
Dataset subset: Antibiotic resistance MIC regression subset, 56 antibiotics
00.53 ± 00.72 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)
unitless · higher

Uncertainty: SD 00.72. The caption says the spread is across antibiotics; the Methods state 37 antibiotics in the binary setting and 56 in the regression setting.

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench antibiotic resistance, MIC regression Pearson correlation

One linear model per drug on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, binary cross-entropy for the resistant or susceptible setting and mean squared error for the log(1+MIC) regression, at most 100 epochs, early stopping patience 10 on validation AUROC or validation R squared, a learning-rate sweep per model across seeds and drugs, Adam with weight decay 0.01, and three random splits.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S18, row 'Evo-2', column 'Regression Pearson'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench antibiotic resistance, MIC regression R squared
Dataset subset: Antibiotic resistance MIC regression subset, 56 antibiotics
03.95 ± 21.04 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)
unitless · higher

Uncertainty: SD 21.04. The caption says the spread is across antibiotics; the Methods state 37 antibiotics in the binary setting and 56 in the regression setting.

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench antibiotic resistance, MIC regression R squared

One linear model per drug on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, binary cross-entropy for the resistant or susceptible setting and mean squared error for the log(1+MIC) regression, at most 100 epochs, early stopping patience 10 on validation AUROC or validation R squared, a learning-rate sweep per model across seeds and drugs, Adam with weight decay 0.01, and three random splits.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S18, row 'Evo-2', column 'Regression R2'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench essential gene prediction, linear probe AUPRC
Dataset: BacBench essential gene dataset from the Database of Essential Genes
15.37 ± 23.71 AUPRC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)
unitless · higher

Uncertainty: SD 23.71

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench essential gene prediction, linear probe AUPRC

The backbone is frozen and a single linear layer, preceded by dropout of 0.2 and layer normalisation, is trained on the gene embeddings with binary cross-entropy loss, at most 100 epochs, early stopping patience 10 on macro AUROC over validation genomes, Adam with weight decay 0.02, and a learning-rate sweep per model. Evo and Evo-2 are reported without layer normalisation, which the authors state worked better for them.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S14, row 'Evo-2', column 'AUPRC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench essential gene prediction, linear probe AUROC
Dataset: BacBench essential gene dataset from the Database of Essential Genes
56.05 ± 9.46 AUROC (linear head on frozen gene embeddings, genus-level split, median over held-out test genomes)
unitless · higher

Uncertainty: SD 9.46

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench essential gene prediction, linear probe AUROC

The backbone is frozen and a single linear layer, preceded by dropout of 0.2 and layer normalisation, is trained on the gene embeddings with binary cross-entropy loss, at most 100 epochs, early stopping patience 10 on macro AUROC over validation genomes, Adam with weight decay 0.02, and a learning-rate sweep per model. Evo and Evo-2 are reported without layer normalisation, which the authors state worked better for them.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S14, row 'Evo-2', column 'AUROC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench phenotypic trait prediction, linear probe AUPRC
Dataset: BacBench phenotypic trait dataset
24.96 ± 20.52 AUPRC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)
unitless · higher

Uncertainty: SD 20.52

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench phenotypic trait prediction, linear probe AUPRC

One linear model per trait on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, cross-entropy loss, at most 100 epochs, early stopping patience 10 on validation AUROC, a learning-rate sweep per model across seeds and traits, and no class weighting. Scores are macro-averaged across the classes of each trait.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S16, row 'Evo-2', column 'AUPRC'
Configuration: Evo-2 (frozen embeddings with a trained linear head)Protocol: BacBench phenotypic trait prediction, linear probe AUROC
Dataset: BacBench phenotypic trait dataset
55.41 ± 8.85 AUROC (linear head per trait on frozen genome embeddings, macro-averaged across classes of each trait, genus-level split, mean over traits and seeds)
unitless · higher

Uncertainty: SD 8.85

Coverage: Not reported scored / Not reported eligible

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

Evo-2 on BacBench phenotypic trait prediction, linear probe AUROC

One linear model per trait on mean-pooled whole-genome embeddings: a single layer preceded by dropout of 0.1 and layer normalisation, cross-entropy loss, at most 100 epochs, early stopping patience 10 on validation AUROC, a learning-rate sweep per model across seeds and traits, and no class weighting. Scores are macro-averaged across the classes of each trait.

Aggregation: Not reported

Bacformer preprint, bioRxiv 2025.07.20.665723 version 3 · Table S16, row 'Evo-2', column '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'

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

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-adc15377fab2
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-adc15377fab2 · 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-evo-2-linear-probe.

Source locator: Table S9, row 'Evo-2'; Methods 'Benchmarking' and 'General modelling details'

Related records

How to cite

Evo-2 (frozen embeddings with a trained linear head). Rewire biological model benchmark database, release 2026-11-21-adc15377fab2. https://benchmarks.rewirebio.io/database/configuration/bacformer-20261115-config-evo-2-linear-probe/

Release 2026-11-21-adc15377fab2. 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