Kmer-Based DNA Sequence Image Representation for Viral Disease, Translation and Mutated Pattern Prediction

Accurate prediction of viral diseases is crucial for effective public health strategies, as mutations in DNA sequences can lead to various viral infections. The translation rate of these DNA sequences significantly impacts the severity of the disease. DNA sequencing techniques are capable of extracting variable-length sequences associated with these diseases. However, existing computational techniques often struggle to effectively utilize DNA sequence data for predictive modeling. To address this challenge, we propose a generalized Convolutional Neural Networks (CNNs) model trained on DNA sequences for predicting different viral disease classification tasks. In our preprocessing technique, DNA sequences are transformed into image-like structures using 6-mer frequencies. We conducted comprehensive experiments, including realm classification, SARS-CoV2 binary classification, and classification of seven types of coronaviruses (CoVs). Our approach achieved significant improvements in test accuracy: 89.51% for realm (4-class) classification, 99.80% for SARS-CoV2 binary classification, and 90.97% for coronavirus (7-class) classification. Additionally, we identified various mutations and translation rates of different CoVs using CDs. While CNNs demonstrate better performance, they are inherently black boxes. To address this issue, we performed interpretability analyses to extract the relevant features of various CoVs.