Predicting functions from protein sequences - where are the bottlenecks? Article uri icon