An encoding scheme capturing generic priors and properties of amino acids improves protein classification

Abstract

Feature engineering aims at representing non-numeric data with numeric features that keep the essential information of the underlying problem, and it is a non-trivial process in building a predictive model. In bioinformatics, there is a profound scale of DNA and protein sequences available, but far from being fully utilized. Computational models can facilitate the analyses of large-scale data. However, most computational models require a numeric representation as input. Expert knowledge can help design features to cast the raw symbolic data effectively. But generally, the features vary from case to case and have to be redesigned for a problem. Automated feature engineering, i.e., an encoding scheme automating the construction of features, saves the redesigning process and allows the researchers to try different representations with minimal effort. This is more in line with the explosion of data and the goal of building an intelligent system. In this paper, we introduce an encoding scheme for protein sequences, which encodes the representative sequence dataset into a numeric matrix that can be fed into a downstream learning model. The method, Context-Free EncodingScheme (CFreeEnS), was proposed for a dataset with labels for pairwise sequences. Here, we improve the method by making it applicable to a batch of protein sequences, requiring no sequence alignment beforehand. The improved method is applied to protein classification at the functional level, including identifying antimicrobial peptides, screening tumor homing peptides, and detecting hemolytic peptides and phage virion proteins. Compared with the traditional methods using task-specific designed features, CFreeEnS improves the predicting accuracy, with an increase ranging from 5.54% to 14.14%. The results indicate that the improved CFreeEnS, free from dependence on carefully designed features, is promising in capturing generic priors and essential properties of amino acids, thereby serving as an automated feature engineering method for protein sequences.