CLMB: Deep Contrastive Learning for Robust Metagenomic Binning

The reconstruction of microbial genomes from large metagenomic datasets is a critical procedure for finding uncultivated populations and defining their functional roles. To achieve that, we need to perform binning, clustering the assembled contigs into draft genomes. Despite existing computational tools, most them neglect one important property data, that is, noise. further improve binning step reconstruct better metagenomes, propose deep Contrastive Learning framework Metagenome Binning (CLMB), which can efficiently eliminate disturbance noise produce more stable robust results. Essentially, instead denoising data explicitly, add simulated training force learning model similar representations both noise-free distorted data. Consequently, trained will be handle it implicitly during usage. CLMB outperforms previous state-of-the-art methods significantly, recovering near-complete on almost all benchmarking (up 17% reconstructed compared second-best method). It also improves performance bin refinement, reconstructing 8–22 high-quality 15–32 middle-quality than result. Impressively, in addition being compatible with refiner, single even recovers average 15 HQ refiner VAMB Maxbin datasets. On real mother-infant microbiome dataset 110 samples, scalable practical recover 365 (including 21 new ones), providing insights transmission. open-source available at https://github.com/zpf0117b/CLMB/ .