A review of computational tools for generating metagenome-assembled genomes from metagenomic sequencing data

Chao Yang; Debajyoti Chowdhury; Zhenmiao Zhang; William K. Cheung; Aiping Lu; Zhao Xiang Bian; Lu Zhang

Source-linked AI summary

A review of computational tools for generating metagenome-assembled genomes from metagenomic sequencing data

Chao Yang, Debajyoti Chowdhury, Zhenmiao Zhang, William K. Cheung, Aiping Lu, Zhao Xiang Bian, Lu Zhang

arXiv:2109.00809v1Sep 2, 2021q-bio.GN

TL;DR

Identifying microbes and their genomes in complex microbial ecosystems remains challenging, motivating culture-independent metagenomics. This review examines tools and techniques for constructing and annotating metagenome-assembled genomes, highlighting the discovery of over 150,000 microbial genomes, more than half from previously undescribed species, and improved mapping ability from 67.76% to 87.51%.

Problem
Identifying microbes and their genomes in complex microbial ecosystems remains challenging, despite metagenomic approaches addressing this challenge.
Method
The paper reviews tools and techniques for constructing and annotating metagenome-assembled genomes, organized into upstream and downstream analyses.
Results
>150,000 microbial genomes were uncovered, with more than 50% of belonging species previously undescribed, while mapping ability increased from 67.76% to 87.51%.
Takeaways & Limitations
The reviewed MAG tools provide a framework for investigating microbial genomes and taxa from complex metagenomic communities.
Takeaways & Limitations
A common limitation of de Bruijn graph assemblers is selecting the k-mer size.

Abstract

from arXiv · show

Microbes are essentially yet convolutedly linked with human lives on the earth. They critically interfere in different physiological processes and thus influence overall health status. Studying microbial species is used to be constrained to those that can be cultured in the lab. But it excluded a huge portion of the microbiome that could not survive on lab conditions. In the past few years, the culture-independent metagenomic sequencing enabled us to explore the complex microbial community coexisting within and on us. Metagenomics has equipped us with new avenues of investigating the microbiome, from studying a single species to a complex community in a dynamic ecosystem. Thus, identifying the involved microbes and their genomes becomes one of the core tasks in metagenomic sequencing. Metagenome-assembled genomes are groups of contigs with similar sequence characteristics from de novo assembly and could represent the microbial genomes from metagenomic sequencing. In this paper, we reviewed a spectrum of tools for producing and annotating metagenome-assembled genomes from metagenomic sequencing data and discussed their technical and biological perspectives.

4 Institute of Brain and Gut Research, School of Chinese Medicine, Hong Kong Baptist

The section lists an affiliation, equal-contribution note, and keywords related to microbiome metagenomics and computational tools.

The affiliation includes the Chinese Medicine Clinical Study Center at Hong Kong Baptist.
The authors’ note states that the listed authors contributed equally to the work.
The keywords cover microbiome, metagenomics, metagenome-assembled genomes, metagenomic sequencing, and computational tools.

1. Introduction

Metagenomics overcomes culture-dependent limitations by enabling culture-independent recovery of microbial genomes, including previously uncharacterized taxa. This review examines tools for assembling, quality-checking, binning, and annotating metagenome-assembled genomes (MAGs), alongside challenges and strategies in metagenomic analysis.

Motivation: Culture-dependent methods miss broad microbial diversity and co-evolution, leaving much gut microbial diversity uncharacterized.Metagenomics addresses this limitation by studying complex microbial communities without species isolation.
Motivation: Reference-based profiling is constrained because approximately 40%-50% of human intestinal microbes lack a reference genome.Incomplete reference sequences make novel genes, species, and strains difficult to annotate.
MAG construction: MAGs represent microbial genomes from metagenome assemblies by binning contigs according to sequence context, abundance, and abundance co-variation.The resulting MAGs undergo quality checks before gene prediction and annotation.
MAG applications: >150,000 microbial genomes were uncovered in a large-scale human-gut metagenomic assembly, with more than 50% of belonging species never described before.These assemblies increased average gut metagenome read-mapping ability from 67.76% to 87.51%.
Review scope: The review covers upstream MAG construction, downstream MAG annotation, and challenges with potential strategies for addressing technical limitations.Upstream topics include assembly, quality control, and contig binning; downstream topics include gene prediction, functional annotation, and taxonomic profiling.

2. Tools for upstream analyses to construct MAGs

Upstream MAG construction combines short-read and long-read assembly, contig quality assessment, binning, and quality-based MAG selection. Short-read graph assemblers address chimeras and memory constraints, while long-read and hybrid methods improve genome connectedness and continuity.

Short-read assembly: Short-read assemblers use OLC or De Bruijn graphs, but OLC tools struggle with large reads and chimeric contigs, whereas De Bruijn methods reduce memory costs.De Bruijn approaches remain sensitive to k-mer-size selection, which affects repetitive-sequence resolution and uneven node coverage.
Short-read assembly: MetaVelvet partitions De Bruijn graphs using coverage peaks, while MetaVelvet-SL and MetaVelvet-DL apply support-vector-machine and CNN/LSTM models to resolve chimeric contigs.MetaVelvet-DL was reported as more powerful than MetaVelvet-SL for deciphering chimeric contigs.
Short-read assembly: IDBA-UD iteratively prunes graphs and increases k-mer sizes, MEGAHIT couples k-mer selection with succinct De Bruijn graphs, and metaSPAdes uses multiple k-mer sizes to identify chimeric contigs.Ray Meta generates local coverage distributions and supports distributed metagenome assembly on computer clusters without large-memory machines.
Long-read and hybrid assembly: Long-read and linked-read tools improve contig connectivity by using barcode sharing, scaffold graphs, hybrid assembly, or long-read error correction.Athena recruits co-barcoded reads between contig pairs, cloudSPAdes connects graph edges with similar barcode sets, and hybrid tools combine short-read accuracy with long-range connectivity.

3. Tools for downstream analyses to annotate MAGs · 3.1. Gene prediction tools

The section presents gene identification and annotation as essential but challenging steps after MAG selection, especially because short reads are difficult to assemble. It reviews model-based, dynamic-programming, and deep-learning tools for predicting biologically functional sequences in MAGs.

3.1. Gene prediction tools: Gene identification and annotation follow MAG selection and aim to identify biologically functional DNA sequences within genomes.Short reads can be difficult to assemble, making gene prediction challenging.
Model-based gene prediction tools: Hidden Markov model tools are prevalent for model-based gene prediction, including MetaGeneMark, Glimmer-MG, and FragGeneScan.Their approaches train probabilistic models from sequence composition, likely shared origins, or sequencing-error patterns.
Model-based gene prediction tools: FragGeneScan incorporates sequencing-error models into six-periodic inhomogeneous Markov models to identify genes with frameshifts.This specifically addresses frameshift-containing genes in sequencing data.
Model-based gene prediction tools: Dynamic-programming tools predict bacterial and archaeal genes by optimizing gene models from frame-bias and hexamer coding scores.Prodigal applies these scores to produce a final list of protein-encoding genes, while MetaGene combines intrinsic and extrinsic ORF scores.
Model-based gene prediction tools: MetaGene evaluates intrinsic and extrinsic characteristics of possible ORFs, combines the scores, and estimates an optimal ORF path.MetaGeneAnnotator subsequently improves MetaGene’s prediction ability.
Deep learning-based gene prediction tools: Deep-learning tools have gained attention for gene prediction, including Meta-MFDL, CNN-MGP, and Balrog.These tools classify ORFs or predict genes using fused sequence features, convolutional neural networks, and GC content.
Deep learning-based gene prediction tools: Balrog constructs a universal temporal CNN-based predictor from amino acid sequences in high-quality prokaryotic genomes and newly assembled genomes.Meta-MFDL fuses monocodon usage, mono amino acid usage, ORF length coverage, and Z-curve features for deep stacking classification.

3.2. Gene functional annotation tools

Gene functional annotation tools are reviewed mainly as broad-scope approaches, including accelerated homology-based methods and gene-context methods for novel or incompletely assembled MAG proteins. Gene-context annotation addresses sequences lacking known homology, while combining motif-based and homology-based analyses is recommended for MAGs with substantial novelty.

The review focuses on tools providing broad functional overviews rather than annotations restricted to one or a few biological processes.
Partial, incomplete, or misassembled MAG proteins can impair homology-based annotation, so motif-based and homology-based approaches are recommended together.InterProScan searches protein domains, active sites, and potential functional annotations using patterns and motifs from InterPro-related databases.
Gene-context tools address novel genes without known homology by using gene neighborhoods, orthologs, domains, or flanking-gene conservation.GeConT illustrates genome context and COG orthologs; FunGeCo records Pfam domain locations; FlaGs annotates and clusters flanking genes before phylogenetic visualization.
76% of sequences received specific functional assignments and 83% received non-specific assignments with a combined homology- and gene-neighborhood approach, outperforming standard BLAST methods.

3.3. MAG taxonomic profilers

MAG taxonomic profilers characterize the identities and relative abundances of organisms represented by MAGs without assigning individual sequences. They use protein-, k-mer-, marker-gene-, or SNP-based strategies, each supporting distinct profiling functions and resolutions.

Motivation and profiling scope: MAG taxonomic profilers estimate microbial taxonomic identities and relative abundances, providing a community snapshot rather than assigning individual sequences.Alignment-based approaches are slow for thousands of MAGs, and reference genomes are often incomplete.
Marker-gene and phylogenetic profilers: Single-copy marker-gene methods improve taxonomic resolution over 16S-based methods, which provide limited resolution and poorly represent whole bacterial genomes.GTDB-Tk places query MAGs using marker genes, reference-tree position, relative evolutionary divergence, and average nucleotide identity.
Profiler categories: Available profilers span protein-based Kaiju, k-mer-based Kraken, Kraken2, Bracken, CLARK, and k-SLAM, marker-gene-based MetaPhlAn2 and IGGsearch, and SNP-based strain tools.K-mer methods quantify specific MAG sequences, whereas marker-gene methods report taxonomic abundance; the tools perform distinct profiling functions.
Protein-based profilers: Kaiju translates reads into amino acids and classifies them against protein references, increasing sensitivity relative to nucleotide-based methods.It uses Burrows-Wheeler transformation and FM-indexing to reduce computational time and memory cost, and can report a best-match identifier or lowest common ancestor.
K-mer-based profilers: Kraken2 maintains Kraken’s accuracy while reducing memory and computational requirements, enabling larger reference databases.Bracken estimates species abundance with a Bayesian model, CLARK uses discriminative k-mers with confidence scores, and k-SLAM uses local alignments and pseudo-assembly for more specific assignments.
Marker-gene and SNP-based profilers: Marker-gene and SNP-based tools extend profiling from species abundance and gene presence/absence to strain-level genotypes and mixed-strain identification.MetaPhlAn2 estimates species abundance from marker-gene counts, StrainPhlAn infers strain genotypes, PanPhlAn reports gene-level presence/absence, and ConStrain identifies strain mixtures from universal-gene SNP profiles.

4. Outlook, potential challenges, and strategies to address them

MAG construction is advancing but remains constrained by low-abundance recovery, assembly and binning limitations, incomplete gene and strain characterization, and sequencing errors. Emerging long-read, graph-based, deep-learning, single-cell, and large-scale reference approaches address these gaps while leaving important areas for further development.

Assembly: Assemblers often miss low-abundance microbes, produce poorly contiguous incomplete genomes, and struggle with contamination; long reads may enable complete genomes but introduce sequencing-error and circularization challenges.Efficient algorithms are especially needed for complete and plasmid genome assembly.
Contig binning: Graph-based binning improves performance by linking long and short contigs, while GraphBin2 permits contigs to belong to multiple bins using connectivity and read depth.Most binning tools require contigs longer than 1 Kb, yet many contigs fall below this threshold.
Contig binning: VAMB uses variational autoencoders to reduce high-dimensional TNF noise, represent contigs in low dimensions, and make clustering more accurate.More sophisticated deep-learning models are expected to emerge for this challenge.
Gene prediction: Short and irregular gene families are commonly neglected despite potential roles in host-microbe interaction, phage defense, and bacterial adaptation; deep-learning predictors may improve efficiency and quality.These models avoid requiring manual feature selection.
Strain characterization: Strain-level characterization lacks a ubiquitously accepted tool because populations include low-abundance and extremely closely related strains; long-read phasing and single-cell sequencing may improve distinction and MAG quality.Single-cell sequencing isolates and directly sequences microbial cells, helping deconvolve species mixtures.
Future directions: 92,143 representative genomes reconstructed from human gut assemblies categorized 73% of underlying read data, demonstrating the value of improved MAG references for classifying complex communities.Large-scale assembly and binning can recover genomes from previously undetermined microbiome species.

Figure legends

The figure legends outline contrasts between culture-independent and culture-based genomics and between assembly-based and reference-based approaches. Tables catalog tools for assembly, quality control, binning, gene prediction and annotation, and MAG taxonomic classification and profiling.

Conceptual approaches: The schematic contrasts culture-independent metagenomics with culture-based genomics and assembly-based approaches with reference-based approaches.The legends describe sequencing-data generation for the two strategies and stepwise processes for assembly-based and reference-based approaches.