FRANKFURT. In early phases of drug development, new active substances are tested in animals –alongside numerous other experimental methods. Researchers face a dilemma: on the one hand, for ethical reasons, they aim to keep the number of animals used in an experiment as low as possible. On the other hand, animal experiments must include enough animals to produce reliable and representative results, for example to determine whether a new drug candidate produces a specific effect.

Professor Jörn Lötsch, data scientist and clinical pharmacologist at Goethe University, in cooperation with computer scientist Professor Alfred Ultsch from Philipps University Marburg—neither of whom conducts animal experiments himself—has developed a generative artificial intelligence called genESOM. genESOM is based on a network of thousands of artificial neurons that “learns” the internal structure of a dataset. This allows it to expand the volume of experimentally obtained data and simulate a larger number of animals in the experiment than were actually used.

Integrated Error Monitoring

To train the AI, the scientists used existing data from a previously published mouse study conducted at Fraunhofer ITMP. The research team achieved two key innovations: first, training the AI to generate new data points based on the study data that integrate into the learned data structure as if they had been obtained in real experiments.

The second innovation was integrating error monitoring directly into the process of generating new data points. Generative AI methods generally risk amplifying not only the relevant signal but also noise and random variation. This problem is known as error inflation and can lead to variables that are actually insignificant being incorrectly identified as treatment-relevant (so-called false-positive variables).

By deliberately separating the learning phase from the synthesis phase, it becomes possible to introduce an artificial error signal into the process and precisely measure its propagation. This results in a data-driven stopping criterion that halts data generation before scientific validity is compromised.

AI Training with Published Study Data

genESOM passed a practical test using data from a preclinical study on a multiple sclerosis model. In the original study, 26 mice were divided into three treatment groups to investigate the effects of an experimental drug. Lötsch and Ultsch reduced the dataset to 18 animals (six per group) to simulate a smaller experiment. When they analyzed this reduced dataset, all previously detected treatment effects disappeared completely: statistical tests showed no significance, and machine learning methods could not distinguish between the treatment groups. After augmenting the reduced dataset with additional data points using genESOM, all effects of the full experiment reappeared at the original level of significance – without introducing relevant false-positive findings. Alternative AI methods, including complex deep-learning neural networks tested by the researchers, failed in this case.

Lötsch explains: “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, the data scientist emphasizes that genESOM can only learn from data obtained in real animal experiments. Nor can the number of laboratory animals be reduced arbitrarily: “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.” Nevertheless, Lötsch is convinced: “With genESOM, we can make an important contribution to reducing the number of animal experiments in large areas of preclinical research.”

The project was funded by the German Research Foundation (DFG) under the title “Generative artificial intelligence-based algorithm to increase the predictivity of preclinical studies while keeping sample sizes small.”