Language models, such as chemical transformers, are increasingly being used in the natural sciences, for example to predict potential pharmaceutical compounds. A recent study by the University of Bonn examines how these AI algorithms work and concludes that their predictions are not based on a deep understanding of biochemistry.
The research team led by Prof. Dr. Jürgen Bajorath and his doctoral student Jannik Roth investigated how chemical language models—which are trained on text-based molecular representations such as SMILES strings—arrive at their results. By deliberately manipulating the training data in the context of sequence-based molecular design, the scientists found that:
- The models were able to suggest plausible inhibitors for new enzymes, but only if those enzymes resembled a family already used in the training.
- If an enzyme from a completely different family was used, the model produced unusable results.
This suggests that the models do not learn universally applicable chemical or biochemical principles, but rather base their conclusions solely on statistical correlations and similarities in the data. They thus “parrot” what they have previously learned, with slight variations. The models’ results can nevertheless be useful in drug discovery research—for example, in identifying new applications for known drugs—but should not be overinterpreted.