How artificial intelligence can help to design new drugs

Detail of the Nature Methods cover (original image by Laura Persat, EPFL)
Detail of the Nature Methods cover (original image by Laura Persat, EPFL)

Traditional methods for predicting interactions between proteins and other molecules rely on complex supercomputer simulations. Instead, a group of researchers from EPFL and USI developed a new artificial intelligence system that analyzes the 3D structure of protein surfaces. The new method MaSIF (Molecular Surface Interaction Fingerprinting) is a collaboration between the EPFL Protein Design & Immunoengineering lab (headed by Prof. Bruno Correia) and the group of Michael Bronstein , professor at USI and Imperial College of London, and head of research in Graph Learning at Twitter. The work appeared on the cover of the February issue of the prestigious scientific. 

Proteins are among the most important biomolecules in nature, the "building blocks" of life responsible for a vast array of functions in living organisms. Proteins comprise chains of small molecules called aminoacids; these chains are folded into complex three-dimensional structures under the effect of electrostatic forces. The structure of proteins determines their function, namely how they interact with other biomolecules. Understanding these interactions is key for the development of drugs, which are typically designed to bind to a protein target. 

Through the innovative machine learning technique developed by Prof. Bronstein, called  geometric deep learning , the researchers were able to link the geometric and chemical properties of proteins with their ability to interact. "Proteins can bind to other molecules, such as cell receptors, like putting a key in a door lock" explains Bronstein. "MaSIF finds complementary structures that fit together". 

The algorithm is capable of analysing billions of protein surfaces per second, making it orders of magnitude faster than previous methods. "Now we can also study multiple interactions between proteins, comparing a whole bunch of keys with many locks: it’s like a three-dimensional puzzle, which can now be solved up to 10,000 times faster than before", Bronstein adds. 

The system developed by Bronstein and Correia can also help to understand the protein-to-protein interactions in the human body. According to Bronstein, "we have nearly 20,000 proteins encoded in our genome, and the network of their interactions as well as interactions with other molecules, is very complex. It could be the key to understanding many biological processes and diseases". 

The new method will also help to design new anti-cancer drugs, such as "molecules capable of binding and inhibiting the PD-1/PD-L1 protein complex that makes cancer cells "invisible" to the immune system," Bronstein says. But the applications could go far beyond medicine: proteins could also be designed to be used as chemical sensors that detect minute amounts of explosives, or catalysts to speed up reactions in industrial processes. 

(Edited text from the original interview by Elisa Buson, by arrangement with Ticinoscienza ) 

  • Faculty of Biomedical Sciences
  • Faculty of Informatics