Using a mixture of probabilistic decision trees for direct prediction of protein function