Transformer-based embedding applied to classify bacterial species using sequencing reads

With the emergence of next-generation sequencing and metagenomic approaches, the necessity for read-level taxonomy classifiers has increased. Although the 16S rRNA gene sequence has been widely employed as a taxonomic marker, recent studies have revealed that 16S rRNA is not sufficient to assign species. Therefore, an accurate classifier is required to classify whole-genome sequencing reads into species. With the advancement of deep learning methods and natural language processing technologies, several studies attempted to apply these methods to genomic data and successfully achieved state-of-the-art performance. In this study, we applied transformer- based embedding into bacterial genomes to accurately classify species using sequencing reads. As a case study, we classified Staphylococcus species using sequencing reads. Our model achieved ROC-AVC values of over 0.98 and 0.99 for 151bp and 251bp paired-end reads, respectively. Compared with a cutting-edge method Kraken2, our model classified significantly more S. aureus reads while maintaining comparable precision.