An improved sequence based prediction protocol for DNA-binding proteins using SVM and comprehensive feature analysis

Abstract

Background: DNA-binding proteins (DNA-BPs) play a pivotal role in both eukaryotic and prokaryotic proteomes. There have been several computational methods proposed in the literature to deal with the DNA-BPs, many informative features and properties were used and proved to have significant impact on this problem. However the ultimate goal of Bioinformatics is to be able to predict the DNA-BPs directly from primary sequence.

Results: In this work, the focus is how to transform these informative features into uniform numeric representation appropriately and improve the prediction accuracy of our SVM-based classifier for DNA-BPs. A systematic representation of some selected features known to perform well is investigated here. Firstly, four kinds of protein properties are obtained and used to describe the protein sequence. Secondly, three different feature transformation methods (OCTD, AC and SAA) are adopted to obtain numeric feature vectors from three main levels: Global, Nonlocal and Local of protein sequence and their performances are exhaustively investigated. At last, the mRMR-IFS feature selection method and ensemble learning approach are utilized to determine the best prediction model. Besides, the optimal features selected by mRMR-IFS are illustrated based on the observed results which may provide useful insights for revealing the mechanisms of protein-DNA interactions. For five-fold cross-validation over the DNAdset and DNAaset, we obtained an overall accuracy of 0.940 and 0.811, MCC of 0.881 and 0.614 respectively.

Conclusions: The good results suggest that it can efficiently develop an entirely sequence-based protocol that transforms and integrates informative features from different scales used by SVM to predict DNA-BPs accurately. Moreover, a novel systematic framework for sequence descriptor-based protein function prediction is proposed here.

Figure 1

The overall workflow of the present method. Firstly, the input amino acid sequence is represented numerically by four kinds of features. Secondly, these feature values are transformed to feature descriptor matrices from three different levels. Thirdly, the first round of the evaluation is adopted based on the original descriptor pool and individual SVM models obtained. At last, mRMR-IFS feature selection method and ensemble learning approach are applied as the final evaluation of the optimal SVM model.

Figure 2

The count of three kinds of Dipeptide composition D0, D1, D2.

Figure 3

Definitions of the N-terminal, middle, and C-terminal parts depending on sequence length L for SAA method.

Figure 4

The performance of different AC features with various LG values over DNAdset and DNAaset.

Figure 5

The IFS curves of DNAdset, DNArset and DNAaset.

Figure 6

Distribution of the number of each type of features (a total 12 types) in the optimal feature set.