Whole-genome prediction of bacterial pathogenic capacity on novel bacteria using protein language models, with PathogenFinder2
Alfred Ferrer FlorensaiD, José Juan Almagro ArmenterosiD, Rolf Sommer KaasiD, Philip T. L. C. ClauseniD, Henrik Vedel NielseniD, Burkhard RostiD, Frank M. AarestrupiD
Abstract
Infectious diseases continue to be a leading cause of mortality and pose a significant global health threat. Thus the development of tools for surveillance and early detection of emerging pathogens is needed. In this study, we introduce PathogenFinder2, a novel predictor of bacterial pathogenic capacity in humans, available through an online server (http://genepi.food.dtu.dk/pathogenfinder2), or as a standalone program (https://github.com/genomicepidemiology/PathogenFinder2). The model, using protein language models for whole-genome phenotype prediction, surpasses the performance of previous methods, especially for novel bacterial taxa, while being taxonomy-agnostic and alignment-free. At the same time, it predicts the importance of each protein for the pathogenic capacity. This output might aid in characterizing potential pathogens, it readily identifies new candidates for virulence factors and vaccine targets, and offers insights into infection metabolic pathways. Furthermore, we introduce the Bacterial Pathogenic Landscape, revealing distributions related to the host conditions, antagonist bacteria, infection site, or habitat.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 50
- FWCI
- 84.14
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (18 Apr 2025) · CC BY 4.0
A ProtT5-embedding, ConvNext-plus-attention ensemble predicts human pathogenic capacity from whole bacterial genomes, outperforming prior predictors on species absent from training and flagging influential proteins.
Virulence genes are often shared with commensals, and existing predictors depend on assembled genomes, database alignment, or read-based models that cannot explain mechanisms. PathogenFinder2 uses protein language model embeddings without alignment or taxonomy, aiming to generalize to novel taxa while reporting per-protein importance.
Key results
ShownOn bacterial species excluded from training, PathogenFinder2 reached a higher Matthews correlation coefficient (MCC) and AUC than both retrained baselines and all previously published predictors.
Text reports higher MCC and AUC than PathogenFinder1DB2, KmerPatho, PathogenFinder1, WSPC, BacPacs and DeePac on Test-NovelSpecies (3,310 pathogenic, 976 non-pathogenic genomes), though the legend says PathogenFinder1 was excluded after failing on 1,660 genomes. No numeric MCC or AUC values are given. · 3.1 PathogenFinder2 excels in predicting novel…; Figure legends
ShownWhen test genomes may resemble training data (Test-2024Strains), PathogenFinder2 was comparable to the alignment-based retrained PathogenFinder1DB2, with no significant MCC gain but a better true positive rate at near-zero false positive rate.
MCC higher than PathogenFinder1DB2 but not significant by 95% CI (±1.96 standard errors); Test-2024Strains has 3,704 pathogenic and 176 non-pathogenic genomes. · 3.1 PathogenFinder2 excels in predicting novel…; Figure legends
ShownThe sensitivity/specificity balance was favorable: only DeePac had higher sensitivity on novel species, at the cost of many false positives. Only BacPacs had higher specificity, without a significant difference. Outputs were well calibrated in the extreme probability ranges.
DeePac sensitivity above PathogenFinder2 on Test-NovelSpecies but very low specificity; BacPacs specificity difference not significant; calibration good in output ranges 0-0.3 and 0.8-1. · 3.1 PathogenFinder2 excels in predicting novel…; 4 Discussion
SuggestedApplied to metagenome-assembled genomes (MAGs) from sewage, most were predicted to lack pathogenic capacity, a much lower pathogen fraction than in the training set; embeddings grouped potential pathogens into seven clusters.
Of 2,739 MAGs, 1,839 predicted non-pathogenic, 370 potentially pathogenic, 530 uncertain (near the 0.5 threshold); 13% of sewage genomes predicted pathogenic vs 76.9% in the training dataset. · 3.4 Practical case: bacterial pathogenicity…; 4 Discussion
SuggestedAttention weights highlight proteins in known virulence categories (adhesion, invasion, toxins, iron and glucose metabolism, mobile elements). Genome embeddings also reflect infection site and habitat, though neither was a training label; no ground-truth validation of either is reported.
Top-attention proteins mapped to UniRef50 for Test-NovelSpecies and Test-2024Strains; UMAP of embeddings colored by HDBSCAN clusters and manually annotated from literature (e.g., Streptococcus pneumoniae grouped with lung-infection bacteria). · 3.2 Attention highlights proteins related to…; 3.3 The Bacterial Pathogen Capacity landscape; 4 Discussion
Methods, data and reproducibility
Prodigal-predicted proteins are embedded with ProtT5 and summed per protein into 1024-feature vectors, then stacked as a sequence. An ensemble of four ConvNext-style 1D convolutional networks with Bahdanau attention and a sigmoid output predicts pathogenic capacity.
- Development set: nonredundant genomes labeled pathogenic or non-pathogenic, from NCBIPathogen, ENAPathogen, PATRIC keyword mining, Thermobase, Tempura and Human Microbiome Project · Bacteria (multiple species) · n = 16,297 pathogenic and 4,882 nonpathogenic genomes (full dataset)
- Test-NovelSpecies: held-out set whose species are absent from training, split with SpanSeq · Bacteria (multiple species) · n = 3,310 pathogenic, 976 non-pathogenic
- Test-2024Strains: genomes submitted in 2024, built with the same procedure; 15 PATRIC entries removed after manual review · Bacteria (multiple species) · n = 3,704 pathogenic, 176 non-pathogenic
- Sewage metagenome-assembled genomes from a previous study, used as a practical application · Bacteria (sewage community) · n = 2,739 MAGs
- Methods & tools
- Prodigal, ProtT5, PyTorch, Optuna (TPE), KMA (Hobohm homology reduction), SpanSeq, MMseqs2, KmerFinder, PathogenFinder1, BacPacs, WSPC, DeePaC
- Data & accessions
- https://www.ncbi.nlm.nih.gov/pathogens/https://www.pathogensportal.org/https://www.bv-brc.org/
- Software & environment
- Implemented in PyTorch; the text gives no package versions or container.
- Assay conditions & compute
- Training, hyperparameter search and testing on Tesla NVidia A100 GPUs (40 GB) with AMD Epyc 7H12 CPUs; inference fits in a 24GB GPU with maximum protein length 9920 amino acids.
- Reproducibility gaps
- Listed data URLs are source databases; no deposited processed dataset, train/test split files or accession is stated. Keyword vocabulary is in the Supplementary Material.
Limitations (6)
- Pathogenic labels come from isolation at infection sites, so commensals may be mislabeled. The broad definition (any human infection) also means predictions may include bacteria needing specific host conditions. · stated by the authors
- The non-pathogenic subset (extremophiles plus bacteria frequently in contact with humans without reported infection) may be biased, and the pathogenic/non-pathogenic ratio may not reflect real ecosystems. · stated by the authors
- Predictions for E. coli from diarrheal and non-diarrheal patients did not differ. The authors suggest the predictors may not detect minimal genetic differences, or that the outcome depends on the host, and note that attention scores pose challenges as explanations. · stated by the authors
- The authors state that the 13% pathogen fraction found in sewage genomes cannot be generalized to other ecosystems. · stated by the authors
- The text reports no numeric MCC or AUC values, and the original PathogenFinder1 failed on 1,660 genomes in Test-NovelSpecies and 2,189 in Test-2024Strains and was excluded in the legend, yet appears in the comparison text. · noted in this summary
- Sewage MAG predictions and attention-based protein annotations lack independent experimental validation, and Test-2024Strains has only 176 non-pathogenic genomes, giving imprecise specificity estimates. · noted in this summary
Who should read:Read this if you need a taxonomy-agnostic, alignment-free predictor of bacterial human pathogenic capacity for novel taxa or metagenome-assembled genomes, with protein-level attention output.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Identification and characterisation of bacterial pathogens through Large Language Model-assisted text mining
Michiel Vos, Markus Göker, … Fabrizio Costa·bioRxiv·Jul 29, 2025
- Commensal or pathogen: computational vectorizing microbial genomes for de novo risk assessment and virulence feature discovery in Klebsiella pneumoniae
Kristen L. Beck, Akshay Agarwal, … Ed Seabolt·Communications BiologyPeer-reviewed·Apr 17, 2025
- Taxonomic propagation of phenotypic features predict host pathogen interactions
Wang Liu-Wei, Şenay Kafkas, Robert Hoehndorf·bioRxiv·Apr 28, 2019
- DefensePredictor: A Machine Learning Model to Discover Novel Prokaryotic Immune Systems
Peter C. DeWeirdt, Emily M. Mahoney, Michael T. Laub·SciencePeer-reviewed·Apr 2, 2026
- PathoFact 2.0: An Integrative Pipeline for Predicting Antimicrobial Resistance Genes, Virulence Factors, Toxins and Biosynthetic Gene Clusters in Metagenomes
Júlia Ortís Sunyer, Luis F. Delgado, … Paul Wilmes·bioRxiv·Dec 9, 2024
- PathoFact: A pipeline for the prediction of virulence factors and antimicrobial resistance genes in metagenomic data
L. de Nies, S. Lopes, … P. Wilmes·bioRxiv·Mar 24, 2020