The 3D structure of a protein is made up of chains of amino acids that fold into shapes determining the protein’s function. There are millions of different proteins with different shapes of which only a fraction has been known to scientists. The protein folding problem has been a major challenge in biological research for about 50 years. This week, DeepMind announced that their AI system 'AlphaFold' has been recognized as a solution to this problem by the organizers of the biennial Critical Assessment of protein Structure Prediction (CASP).
Prof. Dr. Stefan Harmeling, professor of machine learning at HHU and member of the Manchot research group, classes this breakthrough 'on a similar scale as AlphaGo'.
One of the leaders of the Use Case Health in the Manchot research group is Prof. Dr. Markus Kollmann, who is a professor of mathematical modelling of biological systems at HHU. One of his research topics comprises deep generative models to predict RNA/protein folding. His assessment of the situation is as follows: 'The unexpectedly accurate prediction of 3D protein structures from sequence information using neural networks is indeed a major breakthrough. Not only is it now possible to solve problems of this complexity algorithmically to a very good approximation, but also because of the enormous practical importance. The structure of a protein essentially determines its function and is therefore central to the development of new drugs. Interestingly, this breakthrough did not come about through the development of completely new concepts, but is based on the smart integration of all information about protein sequences and protein structures freely available on the internet into a single prediction model. However, the computing power required for the development of the prediction model is so huge that a single university cannot keep up.'