Integrating Chaos Game Representation With Convolutional Neural Networks for Accurate Classification of SARS-CoV-2 Variants
IEEE Access · 2026 · 14 · 125836-125852
DOI: 10.1109/access.2026.3723929 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Gabriel B. M. Camara und 3 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior of Brazil, Conselho Nacional de Desenvolvimento Científico e Tecnológico.
Abstract (Originalsprache)
Since December 2019, the COVID-19 pandemic caused by SARS-CoV-2 has created an urgent need for rapid, scalable tools to identify and track viral variants. Conventional alignment-based classification scales poorly to the millions of genomes now deposited in public repositories, motivating alignment-free alternatives. This study introduces an alignment-free framework that classifies variants of concern (VOCs) and variants of interest (VOIs) of SARS-CoV-2 by encoding complete genome sequences as Frequency Chaos Game Representation (FCGR) images and classifying them with a convolutional neural network (CNN). Each genome of approximately 30,000 nucleotides is mapped to a $64\times 64$ fractal image, which is equivalent to a length-6 ( $k=6$ ) $k$ -mer frequency spectrum, collapsing the sequence into a fixed-size, self-similar representation whose sub-quadrants encode local $k$ -mer frequencies independently of sequence length. Using 47,376 class-balanced genomes from the Global Initiative on Sharing All Influenza Data (GISAID), the CNN achieved 99.97% accuracy on the validation set and 99.92% on the test set across six variants, with per-class precision, recall, F1-score, specificity, and Matthews Correlation Coefficient (MCC) reported to account for any residual imbalance. Beyond in-distribution accuracy, the framework is evaluated under temporal drift (sequences collected after the training window) and geographic shift (training on one continent and testing on others), and is benchmarked against published CGR- and alignment-free methods. The framework classifies the six variants with accuracy above 99.9% and MCC above 0.997 for every class, recovers at least 99.3% of the sequences of each variant when trained on one continent and applied to sequences collected elsewhere, and encodes each genome in a single linear pass over the sequence, in $O(L)$ time and with a memory footprint that is fixed by the grid rather than by genome length, so that encoding and classification together take under 10 ms per genome and the complete set is processed in minutes on a single GPU. Characterising the effect of accumulated mutations over time further delimits the conditions under which this accuracy holds and indicates when the model should be retrained, establishing the integration of CGR with CNNs as an accurate and scalable approach to real-time genomic surveillance.
Förderung
- Coordenação de Aperfeiçoamento de Pessoal de Nível Superior of Brazil
- Conselho Nacional de Desenvolvimento Científico e Tecnológico
Zitieren
APA
Camara, G. B. M., Coutinho, M. G. F., Barbosa, R. D. M., & Fernandes, M. A. C. (2026). Integrating Chaos Game Representation With Convolutional Neural Networks for Accurate Classification of SARS-CoV-2 Variants. IEEE Access, 14, 125836-125852. https://doi.org/10.1109/access.2026.3723929
Vancouver
Camara GBM, Coutinho MGF, Barbosa RDM, Fernandes MAC. Integrating Chaos Game Representation With Convolutional Neural Networks for Accurate Classification of SARS-CoV-2 Variants. IEEE Access. 2026;14:125836-125852. doi:10.1109/access.2026.3723929
BibTeX
@article{Camara2026,
author = {Camara, Gabriel B. M. and Coutinho, Maria G. F. and Barbosa, Raquel De M. and Fernandes, Marcelo A. C.},
title = {Integrating Chaos Game Representation With Convolutional Neural Networks for Accurate Classification of SARS-CoV-2 Variants},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {125836--125852},
doi = {10.1109/access.2026.3723929}
}
EUVIMED