Evaluating DNA Sequence Encoding Methods for Taxonomic Classification Using Deep Learning Techniques: One-Hot Encoding, k-mer, and Embedding Schemes

Abstract

The selection of the numerical representation of DNA sequences plays an essential role in taxonomy classification using deep learn-ing. However, systematic comparisons of different ways to encode DNA sequences are still missing. In our work, we compare three basic representation types: one-hot encodings, k-mer frequency vectors, and learned representations (DNA2Vec). By employing a unified CNN-BiLSTM architecture and three benchmark data sets (H3, H4, species classification) and evaluating accuracy, computation time, sparsity of the representations, and generalization on other datasets, we show that k-mer frequency vec-tors (k = 6) yield the best single-method accuracy (94.2%), whereas the combination of k-mer features and DNA2Vec leads to 96.1% ac-curacy, which exceeds the recently published state-of-the-art accuracy obtained using DNABERT and Nucleotide Transformer on the same data-sets. Statistical testing (paired t-test, p < 0.05) shows that the proposed hybrid approach is significantly better than all individual representa-tions. We provide confidence intervals and guidelines for reproducing our results. Although the results look promising, they require further ex-ternal validation to demonstrate generalization.