AI system reduces animal testing in drug discovery by 50 percent

Researchers at the Goethe University Frankfurt have developed a generative artificial intelligence system that could significantly reduce the number of animals needed in early-stage drug testing.

Scientists involved in the project say the technology, known as genESOM, may help address one of the major ethical and scientific dilemmas in preclinical research: balancing the need to minimise animal testing with the requirement for robust and meaningful experimental results. 

In the early phases of drug development, potential new medicines are tested in animals alongside various other experimental methods. Researchers aim to keep the number of animals used as low as possible for ethical reasons. However, studies must still include enough animals to ensure the findings are statistically reliable and capable of identifying whether a treatment has a genuine effect.

AI designed to simulate larger studies

The new system was developed by Jörn Lötsch, a data scientist and clinical pharmacologist at Goethe University Frankfurt, in collaboration with computer scientist Alfred Ultsch from Philipps University Marburg.

Neither researcher conducts animal experiments themselves. Instead, their work focuses on using artificial intelligence and data analysis to improve scientific research methods.

genESOM is based on a network of thousands of artificial neurons that can learn the internal structure of a dataset. Once trained, the AI can generate additional data points that behave as though they originated from real experiments, effectively simulating a larger group of animals than was actually used in the study.

genESOM is based on a network of thousands of artificial neurons that can learn the internal structure of a dataset

To train the system, the researchers used data from a previously published mouse study carried out at Fraunhofer Institute for Translational Medicine and Pharmacology.

Built-in monitoring to prevent false results

One of the main challenges with generative AI in science is the risk of amplifying random noise or insignificant variations within the data. This issue, known as error inflation, can result in false-positive findings where variables appear important despite having no genuine scientific relevance.

The researchers said genESOM addresses this problem by separating the learning phase from the data synthesis phase. This allows scientists to introduce an artificial error signal and measure how it spreads through the generated data.

Using this approach, the system can stop generating synthetic data before scientific validity is compromised.

Successful test in multiple sclerosis research

The AI system was tested using data from a preclinical study involving a mouse model of multiple sclerosis. In the original experiment, 26 mice were divided into three treatment groups to investigate the effects of an experimental drug.

Lötsch and Ultsch then reduced the dataset to 18 animals, with six mice in each group, to simulate a smaller experiment. When the reduced dataset was analysed, all previously identified treatment effects disappeared and the results no longer showed statistical significance.

However, after the reduced dataset was expanded using genESOM-generated data points, the treatment effects reappeared at the same level of significance as the original study, without introducing meaningful false-positive findings.

According to the researchers, alternative AI systems, including more complex deep-learning neural networks, failed to achieve the same result.

Lötsch said: “We have now tested a number of datasets in a similar way and can say today: with genESOM, the number of animals used in exploratory research can be reduced by 30 to 50 percent while maintaining scientific validity.”

We have now tested a number of datasets in a similar way and can say today: with genESOM, the number of animals used in exploratory research can be reduced by 30 to 50 percent while maintaining scientific validity

However, he cautioned that the system could not replace real experimental data entirely.

“If too few animals are included in an experiment and the number is then simply supplemented using generative AI, the experiment could quickly become scientifically worthless due to the amplification of random findings,” he said.

Nevertheless, Lötsch believes the technology could have a major impact on future preclinical research.

“With genESOM, we can make an important contribution to reducing the number of animal experiments in large areas of preclinical research.”