sRNAdeep: a novel tool for bacterial sRNA prediction based on DistilBERT encoding mode and deep learning algorithms

• Published in: BMC genomics Vol. 25; no. 1; pp. 1021 - 14
• Main Authors: Qian, Weiye, Sun, Jiawei, Liu, Tianyi, Yang, Zhiyuan, Tsui, Stephen Kwok-Wing
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 31.10.2024; BioMed Central; BMC
• Subjects: Accuracy; Algorithms; Analysis; Anopheles; Bacteria; Bacterial genetics; Bacterial sRNA; Care and treatment; Computational Biology - methods; Computational linguistics; Data mining; Datasets; Deep Learning; Drug metabolism; Drug resistance; E coli; Experimental methods; Gene expression; Genes; Genome analysis; Genome, Bacterial; Genomics; Health aspects; Language processing; Machine learning; Methods; Mycobacterium tuberculosis; Mycobacterium tuberculosis - genetics; Natural language interfaces; Performance evaluation; Physiological aspects; Prediction models; Prevention; Proteins; RNA; RNA, Bacterial - genetics; RNA, Small Untranslated - genetics; Software; Sparsity; Support vector machines; Therapeutic targets; China; Deep learning; Genome analysis; Mycobacterium tuberculosis; Bacterial sRNA
• ISSN: 1471-2164; 1471-2164
• DOI: 10.1186/s12864-024-10951-6

Bacterial small regulatory RNA (sRNA) plays a crucial role in cell metabolism and could be used as a new potential drug target in the treatment of pathogen-induced disease. However, experimental methods for identifying sRNAs still require a large investment of human and material resources. In this study, we propose a novel sRNA prediction model called sRNAdeep based on the DistilBERT feature extraction and TextCNN methods. The sRNA and non-sRNA sequences of bacteria were considered as sentences and then fed into a composite model consisting of deep learning models to evaluate classification performance. By filtering sRNAs from BSRD database, we obtained a validation dataset comprised of 2438 positive and 4730 negative samples. The benchmark experiments showed that sRNAdeep displayed better performance in the various indexes compared to previous sRNA prediction tools. By applying our tool to Mycobacterium tuberculosis (MTB) genome, we have identified 21 sRNAs within the intergenic and intron regions. A set of 272 targeted genes regulated by these sRNAs were also captured in MTB. The coding proteins of two genes (lysX and icd1) are implicated in drug response, with significant active sites related to drug resistance mechanisms of MTB. In conclusion, our newly developed sRNAdeep can help researchers identify bacterial sRNAs more precisely and can be freely available from https://github.com/pyajagod/sRNAdeep.git .