“Obtaining a 3-D representation of a protein is difficult and very costly, ” said Becker. “It is not actually possible for many proteins, and in the cases where it is possible, we have to use methods like X-ray crystallography or nuclearmagnetic resonance spectroscopy”. This is why scientists try to develop computer-based methods in order to make faster predictions about three-dimensional protein structure, from which they obtain indications about the properties of the proteins that will form. This is the area in which Julien Becker and Louis Wehenkel work. “The Holy Grail in this area is the ability to predict the 3-D structure of proteins,” Becker said.

Toward a tool for predicting 3-D structure

Finding a method to predict the three-dimensional structure of proteins that is computer-based is a tremendous challenge. Researchers are obliged to begin working on the simplest aspects of the problem. “We have been using computers to predict the properties of proteins for 30 years. In the beginning we just looked at their makeup in terms of amino acids and tried to see if it was a matter of soluble or non-soluble residues,” Becker said. “Today we use automatic prediction more and more and we work with thousands of proteins which we try to characterize using mathematical techniques”. Automatic prediction is a technique that allows rules for the formation of protein structure to be produced automatically on the basis of previous data that contains examples, cases that have already been examined and validated. “If you take the example of a database that has pictures of cars, trucks and motorcycles in it, you see that these images are annotated so that you know which image belongs to which category, ” Becker explained. “Then we apply an algorithm to that information that will look for similar forms. Once the model has been learned, we can feed new images into it and it will automatically classify them into the correct category”.
 
With regard to proteins, scientists assign values to various properties that they want to predict such as the accessibility of residues, the secondary structure in the form of a helix or in the form of a sheet, etc. In general, researchers try to generate one predictor per type of property. “My idea was to combine all that into one predictor, and in order to do that I started working on problems a little more complicated than disulfide bridges,” he said. “Here the difficulty is not to predict a property for a single residue but for a pair of residues”.