Chaos Game Representation of DNA and Hypercomplex Neural Networks Classification

We encode DNA strains using modified Chaos Game Representation (CGR) and then use it as an input for neural networks for multiclass classification against the species. After optimization of classical (ANN) and hypercomplex neural networks (HvNN), the results show that comparable classical neural networks have slightly better average F1 scores and smaller standard deviations. However, hypercomplex neural networks have slightly better precision and recall. The number of trainable hypercomplex neural networks is 10⁴ times smaller than for classical models of comparable F1 score. Therefore, we confirmed two hypotheses: 1) CGR encoding of DNA is sufficient to classify DNA concerning the biological species; 2) Simple architectures of hypercomplex neural networks distinguish classes similar to classical neural networks in multilabel classification in CGR-encoded species classification. However, the number of trainable parameters is about a thousand times smaller.