
ï“

Drug development: AI models do not understand protein interactions

by | Oct 29, 2025 | Health, Research

Modern AI programs support drug development by predicting protein interactions with small molecules. Researchers at the University of Basel show that these models merely memorize patterns instead of capturing physical relationships, and fail at new proteins that are relevant for innovative therapies.

Proteins serve as active ingredients or targets for drugs. Deciphering their three-dimensional structure is crucial. AI models such as AlphaFold and RosettaFold calculate folds and interactions with ligands, which is valuable for drug development. The developers received the Nobel Prize in Chemistry in 2024.

Symbolic image (Credits: pixabay)
Symbolic image (Credits: pixabay)

The high success rates raised doubts, as training data is limited to only 100,000 structures. Tests showed that models predict the same structures when altered amino acid sequences block binding sites or change charges. The same applies to modified ligands.

In more than half of the cases, models ignored interference. They fail especially with unknown proteins. For drug development, predictions must be validated by experiments or physicochemical analyses.

Future models should integrate physicochemical laws to enable more realistic predictions for new structures and open up therapeutic approaches.

Original Paper:

Investigating whether deep learning models for co-folding learn the physics of protein-ligand interactions | Nature Communications


Editor: X-Press Journalistenbüro GbR

Gender Notice. The personal designations used in this text always refer equally to female, male and diverse persons. Double/triple naming and gendered designations are used for better readability. ected.