Moment Vector Encoding of Protein Sequences for Supervised Classification

摘要

Automated prediction of biological attributes of protein sequences with machine learning methods depends on a well-suited protein representation. A central challenge is to represent variable-length sequences as fixed-length feature vectors. In this paper we introduce a new approach for representing the protein sequences as a fixed length vector based on statistical moments applied directly to the values of physicochemical properties of amino acids. The results show that this approach of encoding gives higher prediction accuracy on four benchmarks compared to the previous approaches that applied moments of complex descriptors extracted from the physicochemical properties, and even better than the PseAAC encoding method. The best results are achieved by removing highly correlated features with principal component analysis.