Early drug development always pits ethics against solid science. You want fewer animals tested, yet need enough for trustworthy outcomes on whether a new substance actually works.
Two professors just cracked this with genESOM, a generative AI that expands small datasets to mimic larger experiments.
Smart Data Expansion
Professor Jörn Lötsch from Goethe University and computer scientist Professor Alfred Ultsch from Philipps University Marburg built genESOM on thousands of artificial neurons. It learns the hidden structure of real study data, then creates extra data points that behave exactly like results from additional animals.
They trained it on published mouse data from Fraunhofer ITMP and added built-in error monitoring. This stops the common generative AI problem of error inflation, where noise gets mistaken for real treatment effects.
Proof It Works
In a multiple sclerosis model, the original study used 26 mice. When the team cut it to 18, all treatment signals vanished. After feeding the reduced set through genESOM, the full effects returned at original significance levels, with no false positives. Other deep-learning tools failed here.
The studies appear in Pharmacological Research, iScience, and Briefings in Bioinformatics.
Lötsch sums it up: “With genESOM, the number of animals used in exploratory research can be reduced by 30 to 50% while maintaining scientific validity.”
It only works on real experimental data and can’t replace animals entirely. Still, this tool offers a practical way to ease animal testing in preclinical work.
