bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / exo-tox
Protein
LMU MunichTechnical University of MunichReleased April 2025

Exo-Tox

Bacterial exotoxin classifier over frozen ProtT5 embeddings that separates secreted toxins from non-toxic secreted proteins at MCC 0.94.

78Openness

Where to run it

No providers recorded yet. Browse all providers

Bacterial exotoxins are secreted proteins that damage host cells, and identifying them in a newly sequenced genome is a routine step in pathogen characterization and vaccine target selection. The standard approaches are poorly matched to the task. Homology search finds toxins that resemble known ones and misses the rest, while general-purpose toxin predictors are trained across venoms, plant toxins, and antimicrobial peptides, so the discriminating signal they learn is not the one that separates a bacterial exotoxin from an innocuous bacterial secreted protein.

Exo-Tox, developed at the Walther-Straub Institute of Pharmacology and Toxicology at LMU Munich with the Technical University of Munich, narrows the question deliberately. It is a classifier trained on a single, well-posed contrast: bacterial exotoxins versus secreted bacterial proteins that are not toxic. Sequences are represented by embeddings from ProtT5-XL-UniRef50, one of the ProtTrans protein language models, which are used frozen — the language model is never fine-tuned, and only a small classifier on top of it is trained. The work was posted as a preprint in April 2025 and published in BioData Mining in August 2025.

#Key Features

  • Frozen protein-language-model features: 1024-dimensional per-protein ProtT5 embeddings are compressed by PCA to 20 components retaining 56.46% of the variance, then fed to a support vector classifier.
  • A deliberately narrow decision boundary: Training against secreted non-toxic bacterial proteins, rather than against generic non-toxins, is what produces the performance gap over generalized toxin predictors.
  • Strictly deduplicated benchmark: MMseqs2 removes cross-set identical sequences and reduces within-set identity to 30%, so held-out performance is not inflated by close homologs of the training data.
  • Documented domain limits: The authors characterize where the classifier overpredicts rather than leaving it to users to discover, an unusually candid transferability analysis for a tool of this kind.
  • Alternative encodings benchmarked: Amino acid composition and amphiphilic pseudo-amino-acid composition were evaluated on the same splits, isolating how much of the performance comes from the language model.

#Technical Details

Exotoxins were curated from Swiss-Prot and the literature, restricted to active toxin subunits and excluding fragments, yielding 2,396 sequences; non-toxic secreted proteins were drawn from PSORTb 3.0 predictions, giving 9,082. After redundancy reduction the training and evaluation sets contain 1,069 toxins and 1,308 non-toxins, split with a 15% stratified hold-out. Five classifiers — support vector, k-nearest neighbors, logistic regression, random forest, and XGBoost — were compared across the three encodings; the winning Embs20/SVC combination reaches MCC 0.938 ± 0.036, accuracy 0.969 ± 0.019, precision 0.948 ± 0.035, recall 0.986 ± 0.019, and ROC-AUC 0.995 ± 0.006. BLAST reaches MCC 0.731 on the same task and structure-based Foldseek 0.866, while the generalized predictors CSM-Toxin and MultiToxPred 1.0 reach 0.140 and −0.066.

Out of domain the behavior degrades sharply and the paper says so: 97% of 159,637 bacteriophage proteins and roughly 60% of non-secreted bacterial proteins are called toxins. The authors attribute this to a training-set bias toward phage-associated toxins and conclude the model should be applied only to secreted bacterial proteins.

#Applications

The intended workflow is downstream of a secretion-prediction step: filter a bacterial proteome to predicted secreted proteins, then run Exo-Tox to flag likely exotoxins for experimental follow-up. That suits microbiologists annotating pathogen genomes, groups prioritizing virulence-factor candidates for vaccine or antivirulence programs, and biosafety screening of engineered bacterial strains. Applying it to unfiltered proteomes or to non-bacterial sequences is explicitly outside its validated range.

#Impact

Exo-Tox is a compact demonstration that a narrowly scoped classifier on frozen protein language model embeddings can outperform both homology search and broad-spectrum predictors on a specific biological question, without any additional model training at scale. Its transferability analysis is the more transferable lesson: strong hold-out metrics on a tightly curated contrast say little about behavior on adjacent protein populations, and reporting that gap should be standard. The model, training data, and usage guidelines are deposited at the Open Data LMU repository under CC BY-SA 4.0 rather than on GitHub or HuggingFace, with Python and R code and reproduction instructions included; there is no packaged command-line tool or web server, so use requires working from the deposited archives.

At a glance

Released
April 2025
Category
Protein
Organizations
LMU Munich / Technical University of Munich

Links

Research PaperDatasetDataset

Tags

bacteriaembeddingsprotein_function_predictiontoxin_predictiontransfer_learning

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.