PSP-PJMI: An innovative feature representation algorithm for identifying DNA N4-methylcytosine sites

Abstract

Identifying DNA N4-methylcytosine (4mC) sites is an essential step to study the biological functional mechanism. Feature representation is the primary step to identify 4mC sites due to its influencing the performance of the downstream 4mC site predictive model. Extracting numerical features having strong categorical information from DNA sequences is the key issue to build a 4mC predictive model having good performance. Therefore, a feature representation algorithm referred to as PSP-PJMI is proposed in this paper. It first proposes Pointwise Joint Mutual Information (PJMI), then the bidirectional k-nucleotide Position-Specific Propensities (PSP), so that the PSP-PJMI feature representation algorithm is developed. The parameter ξ is used to indicate the interval from the current nucleotide to the forward or backward dinucleotide in the bidirectional trinucleotide PSP, so that the position information of nucleotides is extracted from a DNA sequence as far as possible. The features corresponding to various ξ are concatenated to comprise the high dimensional feature vector having rich categorical information. The 4mC-BiNP model for identifying DNA 4mC sites is constructed using SVM and the extracted features. The experimental results of 10-fold cross validation test, cross-species validation test, and independent test on 6 species datasets show that the proposed PSP-PJMI algorithm can extract features having richer categorical information than the available feature representation algorithms can do. The 4mC-BiNP model is superior to the state-of-the-art predictive models for identifying DNA 4mC sites. Furthermore, the PSP-PJMI algorithm can be used to extract features for identifying other DNA methylation sites, and also be used for RNA sequences to predict RNA methylation sites.