A deep siamese neural network improves metagenome-assembled genomes in microbiome datasets across different environments

Abstract

Metagenomic binning is the step in building metagenome-assembled genomes (MAGs) when sequences predicted to originate from the same genome are automatically grouped together. The most widely-used methods for binning are reference-independent, operating de novo and enable the recovery of genomes from previously unsampled clades. However, they do not leverage the knowledge in existing databases. Here, we introduce SemiBin, an open source tool that uses deep siamese neural networks to implement a semi-supervised approach, i.e. SemiBin exploits the information in reference genomes, while retaining the capability of reconstructing high-quality bins that are outside the reference dataset. Using simulated and real microbiome datasets from several different habitats from GMGCv1 (Global Microbial Gene Catalog), including the human gut, non-human guts, and environmental habitats (ocean and soil), we show that SemiBin outperforms existing state-of-the-art binning methods. In particular, compared to other methods, SemiBin returns more high-quality bins with larger taxonomic diversity, including more distinct genera and species.

binny: an automated binning algorithm to recover high-quality genomes from complex metagenomic datasets

“…SemiBin (1.0.2) [37] was run using the mode with and default parameters otherwise. For the single sample binning the global model was used, except for the CAMI 2 Gastrointestinal tract samples, for which was used and the CAMI 2 Oral samples, for which was used.…”

“…It evaluates clusters based on the presence of lineage-specific single copy marker genes [23]. We benchmarked binny against six CAMI [17,18] data sets and compared the results with the most popular binning methods MetaBAT2 [15], MaxBin2 [14], CONCOCT [13], and the recently developed VAMB [36], SemiBin [37], and MetaDecoder [38]. We evaluated the contribution of binny to automatic MAG refinement using MetaWRAP [27] and DAS Tool [28].…”

Recovery of Metagenome-Assembled Genomes from a Human Fecal Sample with Pacific Biosciences High-Fidelity Sequencing

“…N 50 and L 50 values were 164,736 bp and 628, respectively. Contig binning was performed with MetaBAT v2.12.1 ( 10 ) and SemiBin v0.5.0 ( 11 ) (parameters: ‐‐environment human_gut). Results from both tools were combined with the bin_refinement module implemented in metaWRAP v1.3.2 ( 12 ).…”

KPop: Accurate and scalable comparative analysis of microbial genomes by sequence embeddings

“…Our strategy is conceptually different from most other comparable methods based on k -mers, which can be broadly divided into the following categories: Methods using specific, possibly very long k -mers to detect the presence or absence of specific sequences, as in Mykrobe [25]. Methods using all k -mers with a low value of k = 4 or 5 and relying on a dimensional reduction in order to cluster and visualise sequences in an abstract small-dimensional space (“binning”, see for instance [26, 27]). Methods using simplified signatures based on some dataset-independent choices (for instance, k -mer “minimizers” as in Mash [28]) to keep signatures (“sketches”) small. Methods such as PopPUNK [29] that use distances, possibly derived from sketches, to summarise the relationships between sequences. …”

“…Methods using all k -mers with a low value of k = 4 or 5 and relying on a dimensional reduction in order to cluster and visualise sequences in an abstract small-dimensional space (“binning”, see for instance [26, 27]).…”