Millions of animals suffer and die every year in animal testing for new drugs, cosmetics and other chemicals. But this could soon be a thing of the past thanks to new intelligent software tools. Such AI systems make it possible to determine the effects and health risks of chemical substances conveniently on a computer. This could reduce the need for animal testing in the long term. But there are still doubts about the reliability of the tools.
Millions of animals are used for animal testing every year around the world, including rats, mice and rabbits, but also fish, monkeys and cats. Industrial chemical substances and drugs are tested on them, among other things. If they do not cause harm to the animals and also fulfill the purpose for which they were developed, then the substances are probably also safe for humans, that is the logic behind such tests.
AI against animal suffering?
But since these tests are associated with the suffering and death of countless living beings, scientists have long been looking for suitable alternatives. These include, for example, experiments on cell cultures and biochips instead of on entire organisms, but also intelligent software tools. These are intended to make it easy to determine the effects and risks of chemical substances on a computer without having to use a single living being or its cells. This not only prevents animal suffering, but is also cheaper than animal testing and also provides results that can be transferred to humans more directly than some animal data. But there are still doubts about the reliability of the tools.
In order to check software tools for “innate” flaws, Sergey Sosnin from the University of Vienna has developed and tested his own tool called “MolCompass”. It works like this: First, you digitally give the system a chemical compound. Then a machine learning model calculates probabilities from 0 to 100 percent that reflect various properties of the chemical compound. So what is the probability that it is toxic, can accumulate in the human body, or bind to a certain human protein? The sum of the probabilities then provides an overview of the effects and health risks of a substance.
Still to be enjoyed with caution
However, a certain amount of caution is required when classifying the probabilities, as Sosnin has found. Only values close to 0 or 100 percent can be considered a correct prediction. If, however, the tool is not sure and gives a rating of 51 percent, for example, then other test methods should be used instead, explains the researcher. However, it can also happen that the model is completely wrong and declares a substance as probably non-toxic, for example, even though it actually is not. “This is the true nightmare scenario for toxicologists,” says Sosnin.
The only solution: identify in advance those classes of organic compounds where the model has “blind spots” and avoid them. But this requires first checking the predicted results for thousands of chemical compounds one by one. To facilitate this tedious task, Sosnin has developed interactive graphical tools that project chemical compounds onto a 2D plane, similar to geographical maps. “We use colors to highlight the compounds that were predicted incorrectly with a high degree of confidence, allowing users to identify them as clusters of red dots,” explains Sosnin.
Software tools for chemical analysis are not yet perfectly reliable, but when they become more sophisticated and tested in the future, they could save numerous animals from a cruel fate.
Source: University of Vienna; Article: Journal of Cheminformatics, doi: 10.1186/s13321-024-00888-z