AI news ·
Generative AI tool cuts animal experiment numbers by up to 50% in preclinical drug research
Goethe University Frankfurt researchers built an AI system that cuts animal use in drug studies by 30-50% without weakening results. The tool generates synthetic data points to expand small datasets while keeping findings statistically valid.
AI System Reduces Animal Testing While Maintaining Research Validity
Researchers at Goethe University Frankfurt have developed a generative AI system that can reduce the number of animals needed in preclinical drug studies by 30 to 50 percent without compromising scientific results.
The system, called genESOM, learns the internal structure of experimental datasets and generates synthetic data points that integrate seamlessly with real observations. This allows researchers to expand datasets from smaller animal groups while preserving the statistical validity of their findings.
The work addresses a persistent tension in preclinical research: ethical pressure to minimize animal use conflicts with the need for sufficient sample sizes to produce reliable results. genESOM offers a practical middle ground.
How the System Works
genESOM uses a network of artificial neurons trained on existing experimental data. During the training phase, the system learns the patterns and structure of a dataset. In a separate synthesis phase, it generates new data points that behave as if they came from actual experiments.
The key innovation is built-in error monitoring. Most generative AI systems risk amplifying noise alongside real signals-a problem called error inflation that can produce false-positive findings. genESOM addresses this by deliberately introducing an artificial error signal during data generation and measuring how it propagates through the system. This creates a stopping point before data synthesis compromises scientific validity.
Testing Against Real Data
Researchers tested genESOM using data from a multiple sclerosis study in mice. The original experiment used 26 animals divided into three treatment groups. The team reduced this to 18 animals and analyzed the smaller dataset alone-the treatment effects disappeared entirely, with no statistical significance detected.
After augmenting the reduced dataset with genESOM-generated data points, all original treatment effects reappeared at their original significance levels without introducing false positives. Alternative AI methods, including deep-learning neural networks, failed to replicate this result.
Jörn Lötsch, a data scientist and clinical pharmacologist at Goethe University who led the work, said the system has been tested across multiple datasets with consistent results.
Limits and Constraints
Lötsch emphasized that genESOM cannot generate data from scratch. It can only work with data obtained from real animal experiments. The system also has practical boundaries-reducing animal numbers arbitrarily before using AI augmentation risks producing scientifically worthless results by amplifying random variations.
The method applies primarily to exploratory preclinical research, not confirmatory studies where larger sample sizes remain necessary.
The findings appear in three peer-reviewed journals: Pharmacological Research, iScience, and Briefings in Bioinformatics.