Suppose we've collected a big (hundreds of thousands) library of different protein sequences with certain features. Then we want to use this data base to train a classifier. And for several statistical reasons we want different functional domains to have somewhat equal frequencies in the data base to reduce domain-specific bias. The straightforward way to achieve this is to use structural information from UniProt and the like, but many proteins don't have any verified structures and de-novo structure prediction might take ages to compute. Alternatively we can perform sequence clustering and pick an even number of sequences out of each cluster. We can either apply a local-alignment based clustering algorithm with ~35% identity threshold or use some sort of hidden Markov model to cluster by profile. What do you think? How biologically relevant this might be? Can this simple clustering based approach help normalise the library to some extent?

  • 1
    $\begingroup$ Biologically relevant for what? If you told us what are the "statistical reasons" or why would you want that we might be able to help. I understand the question but I have no idea what would you want to get rid of some data just to have a "non-biased database". $\endgroup$ – Athe Aug 21 '15 at 14:22
  • $\begingroup$ @Athe We are working on a prediction tool that can get biased due to overrepresentation of specific functional domains in the data set. I'm sorry, but that is as much as I can tell due to information policies in my lab. The sole question can be put in a single sentence: "Is sequence clustering sufficient to normalise the library domain-wise?" $\endgroup$ – Eli Korvigo Aug 21 '15 at 14:28
  • $\begingroup$ Any bias you will get will be the result of a deeper knowledge on certain structures or domains, rather than because of your database. Anyway, if you had a biased database that wouldn't mean you identified wrongly the proteins, but rather that your ability to identify proteins would greater when dealing with certain stuff. $\endgroup$ – Athe Aug 21 '15 at 14:33
  • $\begingroup$ @Athe basically, we want to compare the performance of biased and unbiased classifiers. It's likely that an unbiased classifier will fare better with new data on poorly studied proteins. $\endgroup$ – Eli Korvigo Aug 21 '15 at 14:38
  • 1
    $\begingroup$ @James good to hear from you. I've been working on this ever since. Regretfully, clustering based on sequence alignment is of little use for complicated/modular objects such as proteins. $\endgroup$ – Eli Korvigo Nov 30 '15 at 12:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.