In drug development, experimental methods are often used to determine the three-dimensional structures of target proteins and understand how molecules bind to them. This information is needed to efficiently design drug molecules. However, the process of determining structures can be demanding, meaning that this strategy cannot always be used.

Thanks to the development of AI methods, the structures of proteins can now be predicted more accurately than before.

In the study, researchers from Uppsala University used AI to model the unknown three-dimensional structure of a receptor. In this case, the TAAR1 receptor, an interesting target protein for the development of drugs for mental disorders. Drug molecules that activate TAAR1 have shown promising results in the treatment of schizophrenia and depression.

Using supercomputers, the researchers then searched through chemical libraries containing millions of molecules to find those that best fit the model. Molecules that were predicted to bind to the receptor were then tested in experiments by research colleagues at Karolinska Institutet. An unexpectedly large number of the molecules activated TAAR1, and one of the most potent also showed promising effects in animal experiments.

In the final phase of the research, experimental structures for TAAR1 suddenly became available and the researchers were able to compare them with the AI ​​models.

“The accuracy of the structures generated with AI was astonishing — I couldn’t believe it. The results also show that AI modeling is significantly better than traditional methods. We can now use the same strategy for receptors that we could only dream of before,” explains Jens Carlsson, who led the Uppsala University part of the study.