The General Regression Neural Network to Classify Barcode and mini-barcode DNA

In the identification of living species through the analysis of their DNA sequences, the mitochondrial “cytochrome c oxidase subunit 1” (COI) gene has proved to be a good DNA barcode. Nevertheless, the quality of the full length barcode sequences often can not be guaranteed because of the DNA degradation in biological samples, so that only short sequences (mini-barcode) are available. In this paper, a prototype-based classification approach for the analysis of DNA barcode, exploiting a spectral representation of DNA sequences and a memory-based neural network, is proposed. The neural network is a modified version of General Regression Neural Network (GRNN) used as a classification tool. Furthermore, the relationship between the characteristics of different species and their spectral distribution is investigated. Namely, a subset of the whole spectrum of a DNA sequence, composed by very high frequency DNA

k

-mers, is considered providing a robust system for the classification of barcode sequences. The proposed approach is compared with standard classification algorithms, like Support Vector Machine (SVM), obtaining better results specially when applied to mini-barcode sequences.