Complex systems and Biological physics seminar

Beyond statistical mechanics: external information in protein contact prediction

by David Menéndez Hurtado (Stockholm University)

112:028 ()


Proteins are the machines of the cell, performing tasks from pumping water and nutrients to mediate chemical reactions. Their structure regulates their function, so determining the former is very useful for a variety of applications, from drug development, to basic research. Experimental determination is hard and expensive, so we can sometimes create computer models instead. In the absence of reference structures, the most useful tool we have is predicting contacts: the parts of the protein that are close to each other in space. We do this looking for correlated mutations on families of similar proteins: parts that touch each other will mutate together to preserve the function and stability of the protein. The most powerful methods use Potts models to disentangle the direct couplings from transitional interactions. The focus of this talk will be on how we can use machine learning to introduce information orthogonal to our model: the geometry of the protein constrain the possible patterns that can appear in contact maps, so we can use machine learning to filter and amplify the signal that statistical models give us. Finally, I will discuss some of the current problems, and what and what not can the statistical physics community do for us.