The most advanced artificial intelligence programs are now capable of significantly accelerating pharmaceutical research by modeling molecular interactions. However, a team from the University of Basel has discovered that these tools rely more on memorization than on a genuine understanding of the underlying physical mechanisms.
In the medical field, proteins represent preferred targets for drugs. These biological macromolecules, composed of amino acid chains, adopt three-dimensional structures that determine their function. Deciphering these molecular architectures constitutes a fundamental step in designing innovative treatments. In recent years, the emergence of algorithms like AlphaFold has revolutionized this approach by enabling the prediction of protein shapes from their genetic sequence.
The most recent versions of these models go even further by simulating how proteins interact with different molecules, particularly pharmaceutical active ingredients. However, Professor Markus Lill and his team noticed that the announced success rates seemed abnormally high. This observation led them to suspect that artificial intelligences might be operating through pattern recognition rather than through in-depth physical analysis of molecular interactions.
To verify this hypothesis, the scientists artificially modified hundreds of proteins by specifically altering their binding sites. They created amino acid sequences with radically different electrical charge distributions, and even completely blocked interaction zones. Despite these significant transformations, the AI models continued to predict the same structures, as if the modifications didn't exist. Similar tests on ligands confirmed this tendency.
The researchers found that in more than half of the cases, predictions remained unchanged despite the introduced alterations. This cognitive rigidity becomes particularly problematic when the proteins being studied show little similarity to those used for algorithm training. Yet, according to the research team, it's precisely these original structures that could pave the way for truly innovative drugs.
Faced with these limitations, scientists recommend a cautious approach systematically integrating experimental validations. They also advocate for the development of new generations of algorithms explicitly incorporating the laws of chemistry and physics. Such hybrid models could offer more reliable predictions for poorly understood protein structures, potentially carrying new therapeutic approaches.