Predicting functions from protein sequences--where are the bottlenecks?

Predicting functions from protein sequences--where are the bottlenecks?