DLmeta: a deep learning method for metagenomic identification

Metagenome takes the genome of microbial communities in a specific environment as the research object. It is of great significance in analyzing microbial diversity and exploring the relationship of microbial communities. This paper mainly discusses metagenomic identification, especially the identification of viruses in metagenomes. The identifications of viruses and other sequences are the first step in microbial analysis, and their validity may have implications for downstream work. We develop DLmeta, a deep learning method for accomplishing metagenomic identification. DLmeta obtains domains through gene prediction and protein domain prediction, and uses a model that combines Convolutional Neural Network (CNN) and Transformer to complete the metagenomic identification. We benchmarked DLmeta on data of three species that were viruses, bacteria, and plasmids. The results showed that DLmeta was able to accurately identify species sequences from the metagenomes simultaneously, outperforming other state-of-the-art methods. In the ablation experiment, we demonstrated that our model uses CNN to capture local features and Transformer to capture global features, which can greatly improve the performance of metagenomic identification. DLmeta also can be applied to other metagenomic environments. DLmeta is available at https://github.con xiaozhangzhangl23/DLmeta.