Whole-genome bacterial pathogenicity prediction from ProtT5 embeddings, alignment-free and taxonomy-agnostic, with per-protein attention scores.
No providers recorded yet. Browse all providers
Deciding whether a newly sequenced bacterium can cause disease in humans is routine in outbreak response and food-safety surveillance, and surprisingly hard. The usual answers work by proxy: assign the genome to a species and inherit that species' reputation, or align its proteins against a curated virulence-factor database and count hits. Both break down exactly where the question matters most — on organisms that are new, poorly represented in reference databases, or that belong to species containing both harmless and dangerous strains.
PathogenFinder2, from the Research Group for Genomic Epidemiology at the Technical University of Denmark with collaborators at the Technical University of Munich and Bristol Myers Squibb, replaces both proxies with learned protein representations. Every open reading frame in an assembled genome is embedded with ProtT5 from the ProtTrans suite, and the stacked embeddings — the genome as an ordered bag of protein vectors — are read by a trained classifier. Nothing in the pipeline consults taxonomy or performs an alignment, so the model degrades gracefully on taxa it has never seen.
It is a full redesign rather than an update: the original PathogenFinder, released in 2013, was an alignment-based tool built on protein family databases and shares only the name and the hosting group. PathogenFinder2 is published in Bioinformatics following an April 2025 preprint, and ships as an Apache-2.0 standalone package and a free web server.
The pipeline runs Prodigal to call proteins, embeds each with ProtT5, and sums over the residue dimension to produce one 1,024-feature vector per protein. Stacking these in genome order yields a P x 1,024 matrix that feeds ConvNeXt-style convolutional blocks with layer normalization, followed by the attention layer; a positional encoding of the protein count is concatenated before classification. The released model averages four such networks trained on different training/validation splits. Protein length is capped so inference fits in a 24 GB GPU.
Training used the largest bacterial pathogenicity-annotated dataset assembled to date: 16,297 pathogenic and 4,882 non-pathogenic non-redundant genomes, labeled by mining isolation and host-health metadata from public genome databases, with the two subsets deduplicated separately so that near-identical genomes carrying opposite phenotypes were retained. Splits were made with SpanSeq so that genomes from a species stay together, hyperparameters were tuned with Optuna's TPE sampler against Matthews correlation coefficient, and each network trained for 60 epochs on A100 GPUs. On Test-NovelSpecies (3,310 pathogenic and 976 non-pathogenic genomes from species held out of training), PathogenFinder2 reached a higher MCC and ROC AUC than PathogenFinder1, WSPC, BacPacs, DeePaC, and two baselines retrained on the new dataset. On Test-2024Strains (3,704 pathogenic, 176 non-pathogenic, annotated a year after dataset creation) it matched the strongest alignment baseline while achieving a higher true positive rate at near-zero false positive rate.
The model suits surveillance workflows where isolates arrive faster than they can be characterized: screening clinical, food-chain, or environmental assemblies for pathogenic potential, triaging organisms with no close database relative, and separating pathogenic from commensal strains within a species. Because attention scores nominate the proteins behind a call, the same run generates hypotheses about virulence factors, vaccine targets, and infection-associated metabolic pathways.
PathogenFinder2 shows that a frozen protein language model plus a modest task head can outperform decades of alignment-based tooling on the case that matters — unfamiliar organisms — while producing protein-level evidence rather than a bare probability. The release is unusually complete for an academic bioinformatics tool: Apache-2.0 code, in-repo ensemble weights, training and dataset-construction scripts, and a hosted server. The main constraints are inherent to the label: pathogenic capacity is a binary distilled from database metadata, non-pathogenic genomes are roughly three times rarer than pathogenic ones in training, and the model needs an assembled genome, so it cannot score raw reads or non-coding material.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.