MULTI-SCALE FEATURE LEARNING WITH SELF-ATTENTION LSTM FOR ROBUST CLASSIFICATION OF COVID-19 GENOMIC VARIANTS

Abstract

Accurate and scalable classification of SARS-CoV-2 variants based on their genomics sequences is
critical for efficient surveillance and public health management. However, conventional deep learning
models suffer from issues such as modeling the complexity of high-dimensional genomics sequences,
learning long-range dependencies, and ensuring generalization capabilities for newly emerging variants.
In this study, a new deep learning model called Three-Level Hierarchical Deep Learning with SA-LSTM
is presented for COVID-19 variants classification. Genome sequences are represented in the form of
overlapping K-mers with different levels of granularity and then embedded through Continuous Bag of
Words (CBOW). The designed three-level feature extraction framework utilizing convolutional and
recurrent layers at low-, mid-, and high-level representations allows for the discrimination of
discriminatory genomic patterns through multi-scale learning. Self-attention is then applied to refine the
extracted features in order to highlight the relevant biological motifs and long-term dependencies.
Extensive experimental results on SARS-CoV-2 publicly available genome data reveal that our proposed
system outperforms all existing deep learning approaches, obtaining an accuracy of 99.56%, a high F1-
Score, MCC, and lower inference time. These findings underscore the importance of hierarchical
representation learning along with attention-guided temporal modeling for genomic sequence
classification, thus providing a scalable and understandable solution for COVID-19 variant surveillance
and other future applications of genomic epidemiology.