Protein coding region prediction based on the adaptive representation method

This article proposes a new protein-coding-region prediction technique. The technique maps DNA sequences to numerical strings using an adaptive representation scheme and then uses signal processing to identify coding regions. We learn a mapping from symbols to numerical sequences by computing the distribution variance of each nucleotide in a DNA sequence, and then use the period-3 spectrum to distinguish coding and non-coding regions. Compared to other spectral methods, our method boosts the period-3 spectrum peaks in putative protein-coding regions and attenuates the extraneous peaks in putative non-coding regions by learning to weight the signal by the C-G to A-T ratios. Our adaptive representation method outperforms all other state-of-the-art spectral methods on every benchmark dataset available according to 3 different performance measures.