Ayyuce Demirbas

“There are two main approaches to predicting protein structures. The first is called homology modeling. Protein sequences and structures are the product of billions of years of evolution. If two proteins are near relatives (the technical term is “homologs”) that only recently diverged from each other, they probably have similar structures. To predict a protein’s structure by homology modeling, you first look for a homolog whose structure is already known, then try to adjust it based on differences between the sequences of the two proteins. Homology modeling works reasonably well for determining the overall shape of a protein, but it often gets details wrong. And of course, it requires that you already know the structure of a homologous protein. The other main approach is physical modeling. Using knowledge of the laws of physics, you try to explore many different conformations the protein might take on and predict which one will be most stable. This method requires enormous amounts of computing time. Until about a decade ago, it simply was impossible. Even today it is only practical for small, fast-folding proteins. Furthermore, it requires physical approximations to speed up the calculation, and those reduce the accuracy of the result. Physical modeling will often predict the right structure, but not always.”
― Bharath Ramsundar, Deep Learning for the Life Sciences: Applying Deep Learning to Genomics, Microscopy, Drug Discovery, and More