Preclinical drug research often forces researchers to strike a difficult balance. Scientists want to use as few animals as possible for ethical reasons.
But if too few mice are studied, the statistics can start to fall apart. The real therapeutic effect disappears in the noise.
Now, German researchers say they may have found a way around this problem. They built an AI system that generates realistic synthetic data designed to reduce animal testing.
This model is designed to recover lost signals in small-scale experiments while controlling for false positives.
laboratory dilemma
For ethical reasons, preclinical studies require the use of as few animals as possible while still producing reliable results.
The problem is that small samples often don’t provide useful answers. Variations from mouse to mouse mask what the drug actually does.
Jörn Rötsch, a data scientist and clinical pharmacologist at Goethe University Frankfurt, worked with Alfred Urtsch, a computer scientist at Philipps University Marburg, to approach the problem from a mathematical perspective.
Neither researcher is conducting any animal experiments themselves. Both have spent years grappling with the fundamental problem of extracting clean signals from datasets too small to support firm conclusions.
Teach AI mouse patterns
The two built genESOM and trained it on real mouse data from the Fraunhofer Institute for Translational Medicine and Pharmacology.
This system is easy to explain, but extremely difficult to implement in practice.
It’s a type of generative AI built around thousands of artificial neurons that arrange themselves to reflect the structure of a dataset, an approach known as self-organizing maps.
Once the system learns that structure, it can generate new data points that fit that structure.
Most generation methods handle training and generation in a single sweep. Retsch and Urtsch intentionally separate the two processes: first learning and then synthesizing.
Maintain confidence in your results
Generation techniques involve known risks. In other words, it amplifies everything it sees. Random blips can come back doubling or tripling, loud enough for statistical tests to read an actual treatment effect.
This process is called error inflation, and it produces false positives that look exactly like the real results. genESOM includes internal protection against it.
Before generating new points, the team intentionally injects an artificial error signal. As additional points accumulate, researchers will track how that planted signal grows.
When inflation exceeds a threshold, the model stops. Stopping points are determined based on data, not researchers’ guesses about when to stop.
Comparable results can be obtained with fewer animals
The research team tested the method in a preclinical study of multiple sclerosis. The original study divided 26 mice into three treatment groups to test an experimental drug.
Lötsch and Ultsch reduced their dataset to 18 animals (six mice per group) to mimic small-scale experiments. In that reduced set, all previously detected treatment effects disappeared.
Statistical tests revealed nothing. Machine learning tools were also unable to differentiate between the three treatment groups. The signal was buried in noise.
The researchers then ran genESOM on a smaller dataset. All effects from the original 26 animal study returned with the same strength.
No new false positives occurred. The signal remained in place.
Deep learning wasn’t enough
Other AI approaches have also failed the same test. Complex deep learning networks of the kind that dominate the generative AI headlines are unable to reproduce the original treatment effect from smaller datasets.
These models produced plausible data points. However, the constructs they created are no longer able to convey the original therapeutic signal through statistical testing.
Self-organizing maps are older than today’s deep learning networks. Project data into a structured grid while preserving relationships between data points.
For small, messy datasets, this old approach works well when fancy methods stall.
AI can’t solve everything
This method cannot save an experiment that started too thin. If a study starts with 3 mice per group, there is nothing structurally to learn in genESOM.
If you set the number too low, the expansion will often amplify random fluctuations, turning a small noisy experiment into a larger, equally noisy experiment.
“If an experiment involves too few animals and that number is simply supplemented using generative AI, the experiment can quickly become scientifically worthless as random results are amplified,” Retsch said.
Although the fix is practical, there are limitations. Across the multiple datasets the team worked on, consistent reductions ranged from 30% to 50%.
This 30-50% reduction applies to exploratory animal studies. It does not apply to confirmatory trials or human studies, where different standards dictate what counts as appropriate data.
Additionally, this method has only been validated using a limited number of animal datasets, and researchers do not yet know how consistently it holds across different research fields.
The future of animal testing
In the case of drug development, one-third to half of the mice used in early-stage studies could be replaced with calculated data points that behave similarly to real measurements.
“By using genESOM, we can make an important contribution to reducing the number of animal experiments in a wide range of areas of preclinical research,” said Lötsch.
Until this study, no one had shown that generative models could compensate for missing mice without increasing false positives. The genESOM system seems to do just that, and knows when to stop.
What changes is the mathematics behind exploratory experiments.
Labs may be able to design smaller, early-stage studies to recover meaningful patterns from limited data and reduce the number of animals needed to reach useful conclusions.
The research will be published in a journal Pharmacahlogical research.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check us out on EarthSnap, the free app from Eric Ralls and Earth.com.
—–
