Machine learning for novel feature detection in proteins

Sometimes, we'll have a list of protein sequences that have come from some analysis or experiment that are in some way biologically related—for example, they may all bind the same target—and we will want to determine the parts of those proteins that are responsible for the action. Domain and motif finding, as we've done in the preceding recipes, can be helpful, but only if we've seen the domains before or the sequence is particularly well conserved or statistically over-represented. A different approach is to try machine learning in which we build a model that can classify our proteins of interest accurately and then use the properties of the model to show us which parts of the proteins result in the classification. We'll take that approach in this recipe; specifically, we'll train...