Evaluating the Impact of Encoding Schemes on Deep Auto-Encoders for DNA Annotation

Survey on encoding schemes for genomic data representation and feature learning—from signal processing to machine learning

“…Thus, this encoding scheme can vary in different DNA genomes. For human genome, the below frequencies of dinucleotide were used as an encoding scheme [5] : CG: 0.01, GC: 0.043, CC: 0.047, GT: 0.049, GG: 0.050, AC:0.054, TC: 0.057, GA: 0.061, TA: 0.067, AG: 0.070, CT: 0.071, TG: 0.074, CA: 0.074, AT: 0.081, AA: 0.097, TT: 0.097.…”

“…Such symmetric and complementary properties are beneficial to data training and feature learning. For example, the code book fC: 1, T: 2; A: 2, G: 1g has showed its importance in supervised deep learning networks in recent studies [5] .…”

“…Several common encoding schemes, including EIIP, Galois field GF(4), Binary, Enthalpy, Entropy, and so forth, were assessed and their impact on DNA annotation for the deep auto-encoder network was evaluated [5] . The standard benchmarks were adopted on human gene splicing sites [112] , which contained real and fake splice sites and were divided into training and validation sets.…”

“…It was shown that Complementary scheme (C D 1, T D 2, A D 2, G D 1) performs more effectively in a data set with a large number of features [5] , and schemes such as Binary and DAX perform more effectively in a data set with fewer features. This also demonstrates that the performance of encoding schemes is basically application-specific.…”

“…The first step in GSP is to convert character data or text in DNA genomes into numerical sequences, which is identical for both DSP and machine learning. This conversion can be called numeric representation [4] , encoding scheme [5] , or symbolic-to-digital mapping [6] , which are rather equivalent.…”