Classification on Compressed DNA Sequences: Correlation-Based Feature Extraction Using Self-Supervised Learning

Next-generation sequencing is producing data at an exponential rate, which creates immense computation hurdles for reliably and efficiently storing, transmitting, and analysing the data. Compression plays a vital role in managing the challenges associated with the growing volume of genomic data. Direct analysis over the compressed representations of DNA sequences can facilitate scalability, enabling the analysis of larger datasets within reasonable computational resources. The proposed scheme develops a methodology to extract meaningful features from compressed DNA sequences that maintain a high correlation with features from the original sequences. We employed deep learning techniques to maximize feature correlation between original and compressed sequence representations. This approach ensures that the features derived from compressed sequences are as informative and representative as those from the original sequences. We evaluated the proposed method with various state-of-the-art approaches for genomic sequence analysis using a dataset that includes genome sequences of different viruses. According to the results, the features extracted from compressed sequences are discriminative for the classification task; thus, downstream analysis of genomic sequences is possible without decompression.