Somewhere in a sewage sample collected from the other side of the world, a bacterium nobody has ever cataloged might be quietly carrying the genetic toolkit to make people very sick. Figuring out whether an unknown microbe threatens humans has always been slow and painstaking, often requiring researchers to already know what they’re looking for. Now, a team of scientists has built an artificial intelligence system that can scan a bacterium’s entire genetic blueprint and predict whether it can infect humans, even if science has never encountered that particular bug before.
Called PathogenFinder2, the tool works like a language translator, but instead of converting French to English, it reads the “language” of bacterial proteins and determines whether the overall pattern spells trouble for people. Developed by researchers at the Technical University of Denmark, the Technical University of Munich, and Bristol Myers Squibb, the system is a major step forward from earlier methods, especially when confronting bacteria that don’t match anything in existing databases. In a world where climate change, deforestation, and global trade are expected to expose humans to entirely new microbial species, that capability could prove critical.
Researchers turned PathogenFinder2 loose on thousands of bacterial genomes pulled from sewage samples collected around the globe, the kind of messy, real-world surveillance data that public health agencies increasingly rely on to spot emerging threats. Writing in the journal Bioinformatics, the team reported that the AI flagged hundreds of potentially dangerous bacteria, some belonging to groups never before reported as harmful to humans, and pointed directly to the specific proteins that made them suspicious.
How PathogenFinder2 Reads Bacterial Genes
PathogenFinder2 takes a bacterium’s complete set of genetic instructions and predicts all the proteins that set of instructions would produce. It then converts each predicted protein into a numerical representation using a powerful AI system originally designed to understand protein structure and function, called ProtT5. In essence, each protein gets translated into a kind of fingerprint that captures its most important characteristics.
All of those protein fingerprints are stacked together and fed into a deep learning system, a type of AI loosely modeled on how the brain processes information. The system examines the full collection of proteins as a whole, looking for patterns across the entire genome rather than hunting for individual “smoking gun” genes. That distinction matters. A bacterium’s ability to cause disease in humans typically doesn’t come down to a single gene. It emerges from a web of proteins working together: some help the microbe attach to human tissue, some help it dodge the immune system, and some produce toxins.
Older tools match proteins against a pre-existing database of known harmful genes. That approach falls apart when confronted with something truly new, because there’s nothing in the database to match against. PathogenFinder2 sidesteps this problem entirely by learning general patterns in protein structure associated with the capacity to cause human disease.
How Researchers Built a Record-Setting Dataset of Dangerous Bacteria
To train the AI, the research team assembled what they describe as the largest bacterial dataset ever labeled for disease-causing capacity. The final collection included 16,297 genomes from bacteria known to cause infections in humans and 4,882 genomes from bacteria considered harmless. Gathering the harmless bacteria required particular care. Only clearly non-dangerous strains made the cut: bacteria living in healthy human gut communities, species used in fermented foods like cheese and kefir, and organisms from extreme environments like deep-sea vents or superheated springs where no human tissue exists.
To make sure the AI was learning deeper patterns rather than memorizing specific species, the team used a strict splitting method. They ensured that bacteria in the test set shared no species overlap with those in the training set. A system that only performs well on species it has already seen amounts to a fancy lookup table. The real test is whether it can handle the unknown.
When evaluated on bacterial species completely absent from its training data, a set of 4,286 genomes, PathogenFinder2 topped every competing method it was tested against, including earlier tools called BacPacs, WSPC, and DeePac, as well as a retrained version of the original PathogenFinder. The improvement was measured using a statistical score that accounts for both correct identifications and false alarms, a balanced picture of overall accuracy. Its confidence scores were also well-calibrated, especially at the high and low extremes, so researchers got a clearer sense of which predictions warranted the most trust.
One competing tool, DeePac, caught more true positives, but at a steep cost: it produced a very high rate of false alarms that wrongly tagged harmless bacteria as dangerous. In a real surveillance scenario, that kind of noise could waste precious resources and erode trust in the system. Researchers also tested PathogenFinder2 on a separate set of 3,880 bacterial genomes submitted to public databases a full year after the training data was collected, a practical check to see how the tool holds up against the steady stream of newly sequenced bacteria. Performance remained strong.
An AI That Can Show Its Work
Beyond its raw predictive power, PathogenFinder2 offers something most competing tools cannot: explanations. The system includes an attention mechanism that pinpoints which proteins mattered most for any given prediction. When the team looked at the proteins the model singled out in bacteria predicted to be dangerous, they found familiar culprits: toxins, proteins involved in attaching to human cells, molecules that help bacteria form protective coatings, and machinery for injecting harmful substances directly into host cells.
For bacteria predicted to be harmless, the proteins the model singled out told a very different story: genes for breaking down sugars from plants and milk, producing antimicrobial compounds, and functions related to photosynthesis. The AI wasn’t making arbitrary decisions. It was picking up on biologically meaningful signals.
Researchers also built what they call the Bacterial Pathogenic Capacity Landscape, a map that plots each bacterium based on its overall protein profile as interpreted by the AI. Bacteria that infect similar body sites or share similar lifestyles tended to cluster together, sometimes in ways that cut across traditional family trees. For instance, certain bacteria that infect the mouth grouped near foodborne illness bacteria rather than other mouth-dwelling species, suggesting shared strategies for breaking down biological material. A species that colonizes the upper respiratory tract in animals but can infect humans through bites appeared near human respiratory bacteria, a connection rooted in shared habitat rather than shared ancestry.
Source : https://studyfinds.com/new-ai-tool-could-help-spot-dangerous-bacteria-before-the-next-pandemic/
