DCVBin: a novel binning method for single-sample metagenomes based on DNA language model and variational autoencoder

DNA contigs binning is necessary to reconstruct metagenome-assembled genomes. Current metagenomic DNA contigs binning methods often leverage coverage profiles across multiple related metagenomes and have demonstrated strong performance on co-assembled contigs. However, in single-sample scenarios where coverage information is rare, their performance drops significantly, limiting the in-depth development of metagenomics at the individual sample level. To address this issue, we propose DCVBin, a novel single-sample metagenomic contigs binning method that incorporates semantic features extracted from a DNA language model. Specifically, our approach continues pretraining on a DNA language model to capture more domain-specific semantic representations, which are then integrated with 4-mer frequencies using a variational autoencoder. Clustering is subsequently performed using the k-means algorithm, in which the number of clusters is determined by single copy genes. Experimental results on six publicly available datasets demonstrate that DCVBin achieves high-accuracy single-sample metagenomic binning and outperforms other state-of-the-art methods. Furthermore, DCVBin is included into a disease diagnostic framework that is evaluated on a cohort of gut metagenomes from people with colorectal cancer and healthy people. The framework is shown to be accurate in predicting colorectal cancer using gut metagenomes and has identified a list of potential microbial biomarkers.