Exploring High-Quality Microbial Genomes by Assembly of Linked-Reads with High Barcode Specificity Using Deep Learning

De novo assembly of metagenomic sequencing data plays an essential role in elucidating the genomes of unculturable microbes. Linked-reads, in which short-reads are linked together by barcodes that mark a long original DNA fragment, are a promising method for cost-effective metagenome assembly. Recently, the original linked-read sequencing platform from 10X genomics was discontinued; however, single-tube Long Fragment Read (stLFR) and Transposase Enzyme-Linked Long-read Sequencing (TELL-Seq) are another two linked-read sequencing platforms, which are designed with high barcode specificity and have the potential to efficiently deconvolve complex microbial communities.

We developed Pangaea, a metagenome assembler that assembles linked-reads with high barcode specificity using deep learning. It adopts a fast binning strategy to group linked-reads using a variational autoencoder, followed by rescue of low-abundance microbes with multi-thresholding reassembly. We sequenced a 20-strain-mixed mock community using 10x, stLFR, and TELL-Seq, and stool samples from two healthy human subjects using stLFR. We compare the performance of Pangaea with Athena, Supernova, and metaSPAdes. For the mock community, we observed that the assemblies from Pangaea on stLFR and TELL-Seq linked-reads achieved substantially better contiguity than the assemblies on 10x linked-reads, indicating that barcode specificity is a critical factor in metagenome assembly. We also observed Pangaea outperformed the other three tools on both stLFR and TELL-Seq linked-reads. For the human gut microbiomes, Pangaea still achieved the highest contiguity and considerably more near-complete metagenome-assembled genomes (NCMAGs) than the other assemblers. For the two human stool samples, Pangaea generated more NCMAGs than metaFlye on PacBio long-reads, as well as two complete and circular NCMAGs, demonstrating its ability to generate high-quality microbial reference genomes.