Fast Target Set Reduction for Large-Scale Protein Function Prediction: A Multi-class Multi-label Machine Learning Approach

Large-scale sequencing projects have led to a vast amount of protein sequences, which have to be assigned to functional categories. Currently, profile hidden markov models and kernel-based machine learning methods provide the most accurate results for protein classification. However, the prediction of new sequences with these approaches is computationally expensive. We present an approach for fast scoring of protein sequences by means of feature-based protein sequence representation and multi-class multi-label machine learning techniques. Using the Pfam database, we show that our method provides high computational efficiency and that the approach is well-suitable for pre-filtering of large sequence sets.