How AI Models Track Zoonotic Diseases to Help Prevent the Next Pandemic

Bwindi Impenetrable National Park in southwestern Uganda is renowned for its endangered mountain gorillas, which live in the lush, mist-covered highlands. The park is also home to around 120 mammal species, 350 bird species and hundreds of plant species, alongside people and livestock living within or near its boundaries. This close contact between humans, domestic animals and wildlife creates ideal conditions for what epidemiologists call spillover—the transmission of pathogens between species.

In 2003, the Ugandan non-profit organization Conservation Through Public Health (CTPH), based in Entebbe, began working to protect mountain gorillas (Gorilla beringei beringei) from diseases such as scabies, which can be transmitted by nearby humans. CTPH uses a holistic One Health approach that considers the health of people, animals and the ecosystems they share. Sali Ronald Ogwal, a CTPH public-health expert, says disease monitoring has traditionally been reactive, with teams responding after outbreaks occur. Artificial intelligence (AI), however, is helping researchers move toward more proactive disease prevention.

As part of a three-year collaboration with the public-health initiative NESTLER, a joint project involving the European Union and African countries, Ogwal collected samples from cattle and poultry around Bwindi. The samples were sent to a laboratory in Entebbe and tested for diseases that can infect humans, including brucellosis and Rift Valley fever. After sharing the results with local communities, the team combined the findings with CTPH’s routine gorilla-monitoring records and shared the data with NESTLER colleagues. These combined datasets are being used to train a predictive AI model that can act as an early-warning system for diseases affecting people, livestock and gorillas.

Many infectious diseases in humans are zoonotic, meaning they originate in animals. Examples include SARS-CoV-2, the virus responsible for the COVID-19 pandemic, and Ebola virus, which is thought to have originated in fruit bats belonging to the pteropod family. People can contract Ebola through direct contact with infected bats’ bodily fluids or through animals that have come into contact with infected bats.

Changes in land use, including agriculture and deforestation, are bringing people and wildlife into closer contact, increasing the risk of zoonotic spillover. Climate change can further alter the distribution of animals, insects and pathogens. Globalization and geopolitics can also accelerate the spread of infectious diseases among people. As the COVID-19 pandemic demonstrated, spillover events can have far-reaching consequences for physical and mental health, livelihoods, education and economies worldwide.

For researchers such as Ogwal, AI—combined with practical measures such as regulating the wildlife trade—could help manage zoonotic diseases, reduce spillover risk and potentially prevent outbreaks. AI tools remain limited and are not yet widely used in infectious-disease epidemiology, but advances in machine learning are beginning to change that.

“Rather than just waiting for an outbreak to occur, these AI machines will support deep analysis of large amounts of data to identify patterns,” Ogwal said. Prediction can happen “before something escalates.”

AI as a virus detective

Rapid and efficient data processing is essential for detecting infectious diseases and stopping outbreaks before they grow.

AI technologies such as machine-learning models are designed to “skillfully classify large amounts of data, follow rules and find patterns,” says Edward Holmes, a virologist at the University of Sydney in Australia who studies metagenomics in environmental samples.

Stop the next influenza pandemic

Holmes is developing machine-learning models to identify previously unknown zoonotic viruses in samples that were not necessarily collected for that purpose. His team uses genetic sequence data from public resources such as GenBank and Pathoplexus, an open-source database of viral pathogens. The model is trained on sequences from known human pathogens, allowing it to recognize characteristics linked to the emergence of disease in people. “The receptors on the cell, the mode of transmission, all those things influence these AI algorithms,” Holmes explains. The model could help identify potentially dangerous viruses for further field investigation and support the design of preventive vaccines.

One such tool is LucaProt, a deep-learning algorithm developed by Holmes and colleagues to discover RNA viruses using protein structures derived from genetic sequence data. In 2024, researchers applied LucaProt to 10,487 publicly available metatranscriptome datasets and identified 161,979 RNA viruses. Of these, 70,458 had never previously been identified—the largest virus-discovery effort recorded in a single study.¹ Although only a small proportion may threaten human health, researchers say understanding emerging infectious diseases requires a better picture of viral diversity and the ways viruses evolve and move between species.

Other researchers are also using AI to detect potential zoonotic viruses. In 2021, scientists at the University of Glasgow in the United Kingdom described a machine-learning method that uses known zoonotic viruses to predict which newly discovered viruses are most likely to infect humans.² Earlier this year, US researchers reported a predictive model that identifies a virus’s likely host and recommends when disease surveillance and sample collection may be most productive.³

Although these models are currently used mainly for research rather than routine disease surveillance, Holmes suggests establishing regular antibody testing along “fault lines”—locations where spillover is more likely, such as live-animal markets, poultry farms and settlements near bat roosts. “We want our local staff to be trained in data generation and analysis,” Holmes says. “And that’s all sent to a central location where you get a global radar map of what’s going on, such as air traffic control.” One potential hub, he suggests, could be the World Health Organization’s pandemic and outbreak information hub in Berlin.

Holmes says this approach could be both feasible and cost-effective. The World Bank estimated that One Health-based prevention would cost up to US$11.5 billion annually in 2022—around one-third of the cost of managing a pandemic.⁴ “This is a matter of politics and people,” Holmes says, noting that short-term political thinking can overshadow the potential long-term financial savings of disease prevention.

Human expertise will remain essential, Holmes and his colleagues emphasize. “Although the combination of AI and metagenomic sequencing is valuable, it alone cannot resolve the fundamental uncertainties surrounding pathogen emergence,” wrote virologists Nader Ebrahimi and Amir Ghami of the Pasteur Institute of Iran in Tehran in The Lancet Infectious Diseases in January.⁵

AI-powered disease surveillance

Beyond identifying viruses, AI can help predict and monitor the spread of infectious diseases. BlueDot, a Toronto-based company that assesses infectious-disease risks, provides these services to clients including the Gulf Disease Control Center in Riyadh and the city of Chicago, Illinois. The company uses AI to collect and filter thousands of news articles and official reports from public-health agencies, translate information into 65 languages and extract details about diseases and symptoms. It then combines these data with additional sources, including airline-ticket sales, to help clients understand location-specific risks. “Based on how people travel, these are the risks most relevant to your location,” explains Andrea Thomas, BlueDot’s vice-president of epidemiology and data science.

In 2015, BlueDot researchers identified Miami, Florida, as a potential location for the spread of Zika virus across North and South America. Their assessment combined ecological information, data on Aedes mosquitoes that transmit Zika, temperature patterns and flight routes from the Brazilian epicenter.⁶ In the years that followed, the prediction proved prescient, with 1,471 cases recorded in Florida.⁷

Monitoring fruit bat colonies could provide early warning of Nipah virus outbreaks

In another BlueDot project, researchers built boosted-regression-tree models—a form of AI that combines predictions from multiple models—to identify areas at future risk of dengue, chikungunya, Zika and other mosquito-borne diseases. The team combined information about mosquito species and habitats with climate-change projections to estimate how mosquito distributions could change. Looking ahead to 2036, the researchers identified potential disease-risk patterns across Europe, the United States and southern Canada (see go.nature.com/4xpyryl).

Although AI has become more advanced over the past decade, its use in global health surveillance is not new. Around 20 years ago, epidemiologist John Brownstein and colleagues at Boston Children’s Hospital in Massachusetts created HealthMap, a Google Maps-based tool for tracking disease outbreaks worldwide (see go.nature.com/4wlarvg). The system uses specialized data dictionaries to match terms, including the different names used for diseases in different countries. The team has since moved toward large language models (LLMs), significantly changing how HealthMap collects and organizes information, Brownstein says.

HealthMap gathers information from news websites, government health departments and social media, then filters out irrelevant material using Fisher–Robinson Bayesian filtering, an AI method also used to detect spam emails. In late 2019, the system issued one of the world’s first alerts about the disease later identified as COVID-19. “Informal sources of infectious-disease surveillance can provide an early window into what’s going on in a population,” Brownstein explains.

Since 2025, the team has expanded its work through projects including BEACON—the Biological Threat Emerging, Analysis and Communications Network—based at Boston University’s Center for Emerging Infectious Diseases. BEACON uses LLMs to combine HealthMap alerts with information from other sources and create dashboards that help public-health professionals target advice about diseases such as Ebola, which has occurred in the Democratic Republic of the Congo and Uganda. “This is a combination of both AI and human experts,” Brownstein says. “So this is a more detailed and contextualized risk assessment.” More than 227,000 users, including public-health professionals and clinicians from 233 countries and territories, have accessed the platform since its launch.

At BlueDot, Thomas says large language models are simplifying how the company shares information and coordinates with clients. “New technologies are making it easier to collect and extract all kinds of information,” she says. As modeling tools become more accessible, non-experts can increasingly combine their own datasets without having to share sensitive information externally.

Using AI responsibly

As with AI in other fields, using artificial intelligence to detect and prevent zoonotic diseases raises important ethical questions.

One major concern is how local communities—particularly those in low- and middle-income countries—are included in disease surveillance and AI development. Researchers, including epidemiologists and social anthropologists, warn that AI systems can reproduce bias when they fail to incorporate residents’ lived experiences, local knowledge and on-the-ground realities.