Assessing chemical toxicity across Eukaryota using multimodal transformers

The natural world is a vast, interconnected web of life, from the tiniest single-celled organisms to the largest mammals. For decades, scientists have known that human-made chemicals threaten this web, causing harm to insects, birds, and entire ecosystems. To protect nature, regulators need to know which chemicals are dangerous to which living things. However, a massive gap exists in our knowledge. While thousands of new chemicals are discovered every year and millions of species inhabit the planet, we have tested only a tiny fraction of them. Traditional testing is slow, expensive, and often limited to a few standard animals like fish, rats, and algae. This leaves us guessing about the safety of most chemicals for most species, making it difficult to stop pollution before it causes irreversible damage.

To bridge this gap, researchers have turned to artificial intelligence. Instead of testing every chemical on every animal in a lab, they are teaching computers to recognize patterns in the data we already have. By learning how chemical structures interact with the biology of different species, these computer models can predict toxicity for situations where no experiment has ever been run. This approach promises to speed up safety assessments and reveal hidden dangers to the countless species that have never been studied.

In a new study, a team of scientists has developed a powerful new tool called TRIDENT-2 to tackle this challenge. This system is designed to predict how toxic a chemical will be to any eukaryotic organism—a group that includes almost all complex life forms, from plants and fungi to animals. The researchers trained their model on a massive collection of existing data, pulling together more than 560,000 toxicity tests. These tests covered over 82,000 different chemicals and nearly 6,800 species, representing about one-third of all known groups of eukaryotic life. The data included a wide variety of scenarios, such as how long an organism was exposed to a chemical, how it was exposed, and what specific effects were observed, like death, changes in growth, or behavioral shifts.

The core of TRIDENT-2 is a type of artificial intelligence known as a multimodal transformer. In simple terms, this means the computer does not just look at the chemical structure or the animal separately; it learns to understand how they fit together along with the conditions of the test. It processes the chemical's molecular shape, the type of organism, the route of exposure, and the duration of the experiment all at once. By learning from this diverse mix of information, the model can spot deep connections between different species and chemicals. For instance, it can use what it knows about how a chemical affects a mouse to make an educated guess about how that same chemical might affect a related rodent or even a distant relative in a different part of the animal kingdom.

When the team tested TRIDENT-2, they found it could predict toxicity with remarkable accuracy, even for species and chemicals it had never seen before. The model was able to estimate the concentration of a chemical that would cause a specific effect, such as killing half of a population, with a level of precision that rivals traditional methods. Crucially, the model performed well across the entire tree of life, from insects and plants to fungi and marine creatures. It showed that by learning from a broad range of data, it could fill in the blanks for species that are rarely tested. For example, while standard tests often focus on fish and rodents, TRIDENT-2 successfully predicted toxicity for groups like worms, fungi, and insects, which are often overlooked but vital to ecosystem health.

The researchers also compared TRIDENT-2 against several existing tools used by regulators and scientists. The new model consistently outperformed these older methods, especially when predicting effects for species or chemicals outside the narrow range of traditional testing. While other tools were limited to a handful of test animals and specific types of chemicals, TRIDENT-2 could handle thousands of species and a vast array of substances. The study demonstrated that the model could maintain its accuracy even when asked to predict the effects of chemicals that were structurally very different from those in its training data, or for species that were evolutionarily distant from the ones it had studied.

To show how this works in a real-world scenario, the team applied TRIDENT-2 to thousands of pesticides. They mapped out how these chemicals would likely affect different species across the eukaryotic tree. The results revealed that while some pesticides were highly toxic to their intended targets, like insects or weeds, they also posed risks to non-target species, such as certain worms, fungi, and plants. The model was able to detect these subtle, species-specific sensitivities that would have been missed if the data were only looked at in broad groups. This level of detail is essential for understanding the true impact of pollution on biodiversity.

The study concludes that artificial intelligence can overcome the long-standing data limitations that have hindered ecological risk assessment. By using the vast amount of existing, albeit uneven, data to learn general principles of toxicity, TRIDENT-2 allows scientists to assess the safety of chemicals for the millions of species that have never been tested. This capability does not replace the need for physical experiments but offers a way to prioritize which tests are most needed and to identify potential threats before they cause widespread harm. The tool is now available for public use, offering a new path toward protecting the planet's biodiversity from the growing tide of chemical pollution.