Machine learning for discovering missing or wrong protein function annotations: A comparison using updated benchmark datasets