# Binning Long-Read Metagenomes: How to Recover High-Quality Metagenome-Assembled Genomes (MAGs) from Nanopore and PacBio Data

Long-read sequencing platforms such as Oxford Nanopore and PacBio generate reads that are substantially longer than Illumina short reads, and these longer reads can improve metagenomic assembly contiguity, gene completeness, and the recovery of metagenome-assembled genomes (MAGs). The binning step, which groups assembled contigs into putative genome bins, remains a bottleneck because many established binning tools were designed for short-read assemblies and do not fully exploit the coverage variation, error profiles, and assembly graph structure characteristic of long-read data. This article explains the principles of coverage-based and composition-based binning, introduces binning tools that accommodate long-read assemblies, and provides a practical workflow for refining bins to meet quality thresholds for downstream analysis. The intended reader is a researcher, laboratory professional, or life-science practitioner who has generated or plans to generate long-read metagenomic data and needs concrete decisions for binning, quality assessment, and reporting.

## Why Long-Read Binning Differs from Short-Read Binning

Metagenomic binning assigns assembled contigs to discrete clusters that represent putative microbial genomes. The two primary signals used by binning algorithms are nucleotide composition, typically tetranucleotide frequency patterns, and coverage depth across one or more samples. Short-read assemblers produce many small contigs, and short-read binners rely heavily on coverage signals because composition signals are weak on short sequences. Long-read assemblies produce fewer, longer contigs, which changes the binning problem in several ways.

Long reads can span repetitive elements and capture multiple genes in a single read, which improves assembly contiguity and gene completeness. In a direct comparison of PacBio long reads and Illumina HiSeq short reads from anaerobic digestion microbiome samples, hybrid assemblies that incorporated PacBio long reads produced longer contigs, higher contig N50 values, and more large contigs than short-read-only assemblies. The same study reported that depth-based hybrid assemblies recovered a higher percentage of complete genes compared to short-read-only contigs, because the long reads were long enough to cover repeating short elements and capture multiple genes in a single read. These assembly improvements carry into binning, because longer contigs provide stronger composition signals and reduce the number of fragments that must be correctly assigned to the same genome.

Coverage variation behaves differently in long-read data. Long-read sequencing has lower throughput per flow cell or SMRT cell compared to short-read platforms, and the coverage depth across a metagenome can be uneven. Low-abundance organisms may have very few reads, which makes coverage-based binning unreliable for those genomes. Additionally, long-read error profiles differ by platform and by basecalling model, and these errors can affect composition-based binning if the binner does not account for them. Tools that were optimized for short-read assemblies may not handle these characteristics well, and researchers should expect to use long-read-aware binning strategies or to adjust parameters when using general-purpose binners.

The practical consequence is that binning long-read metagenomes requires a workflow that integrates assembly graph information, coverage depth, and composition features, and that includes a refinement step to evaluate and improve bins before downstream analysis. The remainder of this article describes the core principles, available tools, workflow steps, quality metrics, and common failure patterns that researchers should understand before binning their long-read data.

## At a Glance: Binning Decisions for Long-Read Metagenomes

The table below summarizes the key decisions a researcher must make when binning long-read metagenomic assemblies. These decisions depend on the sequencing platform, the complexity of the microbial community, the availability of multiple samples, and the intended downstream use of the MAGs.

| Decision Point | Short-Read Default | Long-Read Recommendation | Rationale |
| --- | --- | --- | --- |
| Assembly input | Contigs from short-read assembler | Contigs from long-read or hybrid assembler | Long-read assemblies produce longer contigs with stronger composition signals and fewer fragments to bin |
| Binning signal | Coverage depth across samples | Coverage depth plus composition plus assembly graph connectivity | Long-read coverage is uneven, and graph structure helps resolve closely related genomes |
| Binning tool | MetaBAT2, MaxBin, CONCOCT | SemiBin2, GraphMB, LorBin, or MetaBAT2 with long-read options | Long-read-specific binners use graph embeddings or adaptive clustering designed for long-read assemblies |
| Sample number | Multiple samples for differential coverage | Multiple samples if available, but single-sample binning is feasible | Long-read coverage per sample is lower, so multiple samples improve coverage signal |
| Quality assessment | CheckM or CheckM2 completeness and contamination | Same metrics, but interpret with long-read error rates in mind | Long-read assemblies may have different error distributions that affect single-copy gene counts |
| Refinement | Manual curation in Anvi'o or similar | Automated refinement with bin refinement tools, then manual curation | Refinement improves bin purity and completeness before downstream analysis |

## Core Principles of Coverage-Based and Composition-Based Binning

### Coverage Depth as a Binning Signal

Coverage depth, also called abundance or depth of coverage, is the number of sequencing reads that map to a given contig, normalized by contig length and library size. In a metagenome, different microbial species are present at different abundances, and contigs from the same genome tend to have similar coverage depth across samples. If a researcher sequences multiple samples from the same community, for example time series or replicate samples, the coverage profile of each contig across samples becomes a fingerprint that can distinguish genomes even when their composition is similar.

For long-read data, coverage depth must be calculated carefully. Long reads are longer than short reads, so a single read can cover a large fraction of a small contig, and mapping statistics can be sensitive to read length and mapping quality. Coverage should be calculated from read depth after mapping reads back to the assembled contigs, and researchers should use a mapper that handles long reads appropriately. Low-abundance genomes may have very low coverage, and the binner may not be able to separate them from sequencing noise. In the PacBio versus Illumina comparison, the authors noted that lower single-pass accuracy can make gene discovery and assembly for low-abundance organisms difficult, which implies that coverage signals for rare genomes are weaker in long-read data.

### Nucleotide Composition as a Binning Signal

Nucleotide composition refers to the frequency of short DNA motifs, most commonly tetranucleotides, within a contig. Different microbial genomes have characteristic composition signatures that reflect mutational biases, codon usage, and other genomic features. Composition-based binning groups contigs with similar tetranucleotide frequency patterns.

Composition signals are stronger on longer contigs because the frequency estimates are more stable. Short-read assemblies produce many short contigs, and composition-based binning on those fragments is noisy. Long-read assemblies produce longer contigs, which improves the reliability of composition estimates. However, composition alone cannot always separate closely related strains or species with similar genomic composition, and composition signals can be distorted by sequencing errors. Long-read error profiles, particularly for older basecalling models, can introduce systematic biases in k-mer frequencies. Researchers should be aware that composition-based binning on long-read assemblies may require error-corrected reads or assemblies that have been polished.

### Assembly Graph Connectivity as an Additional Signal

The assembly graph, which records the connections between contigs based on read overlap, contains information that is not captured by coverage or composition alone. Contigs that are connected in the assembly graph are more likely to originate from the same genome, because reads that span the junction between two contigs provide evidence of physical linkage. Short-read binners typically ignore the assembly graph, but long-read assemblies produce graphs with fewer, longer nodes, and graph connectivity can be a powerful binning signal.

Graph-based binning methods use the assembly graph structure to inform clustering. One approach, implemented in the tool GraphMB, uses graph neural networks to integrate the assembly graph with contig features during binning. In tests on long-read datasets of different complexities, GraphMB obtained unique bins on all real datasets and produced on average 17.5% more high-quality genome bins compared to state-of-the-art binners, and 13.7% more when aggregating its results with other binners. These results indicate that the assembly graph contains information that improves binning beyond what coverage and composition provide alone.

## Long-Read-Specific Binning Tools and Their Tradeoffs

### SemiBin and SemiBin2

SemiBin and its successor SemiBin2 are binning tools that were designed with long-read metagenomic assemblies in mind. They use deep learning to combine composition and coverage features, and they can incorporate assembly graph information. SemiBin2 is available through Bioconductor, which provides official package documentation and installation instructions for reproducible genomic analysis workflows. Researchers who use Bioconductor for their analysis pipelines can integrate SemiBin2 into an existing R or Python workflow.

The main advantage of SemiBin2 is that it was explicitly designed for long-read data and does not require the user to adapt short-read parameters. It handles the uneven coverage and error profiles typical of Nanopore and PacBio data. The main limitation is that deep learning models require training data, and the model may not generalize perfectly to unusual microbial communities or extreme GC content. Researchers should evaluate binning results with quality metrics instead of assuming that the tool produces high-quality bins for every dataset.

### GraphMB

GraphMB is a binner that uses graph neural networks to incorporate the assembly graph into the binning process. It was developed specifically to address the limitation that current metagenomic binners do not take full advantage of assembly graphs and are not optimized for long-read assemblies. The tool integrates contig-specific features with graph structure information, and the authors reported that this approach improved the number of high-quality bins obtained from long-read datasets.

The tradeoff with GraphMB is that it requires the assembly graph as input, which means the researcher must retain the graph output from the assembler. Not all assemblers produce graphs in a format that GraphMB can consume, and graph construction can be computationally expensive for very large metagenomes. GraphMB is available from its GitHub repository, and researchers should consult the documentation for input format requirements.

### LorBin

LorBin is a more recent binner that was designed to address two specific challenges in long-read metagenomic binning: identifying unknown species and managing imbalanced species distributions. It uses a two-stage multiscale adaptive clustering approach with DBSCAN and BIRCH clustering, and it uses single-copy genes in evaluation decision models to maximize MAG recovery. In tests on simulated and real microbiomes, including oral, gut, and marine samples, LorBin outperformed six competing binners and generated 15-189% more high-quality MAGs. It also identified 2.4-17 times more novel taxa than state-of-the-art binning methods.

The main advantage of LorBin is its ability to recover genomes from species-rich samples that contain unknown taxa and from communities with imbalanced species distributions, which are common in natural microbiomes. The tradeoff is that LorBin is a newer tool, and its parameter choices and clustering behavior may be less familiar to researchers than those of established binners. Researchers should validate LorBin results with independent quality metrics.

### MetaBAT2 with Long-Read Options

MetaBAT2 is a widely used binner that was originally designed for short-read assemblies, but it can be applied to long-read assemblies with appropriate parameter adjustments. MetaBAT2 uses coverage and composition features and can accept multiple samples for differential coverage binning. When applied to long-read assemblies, MetaBAT2 may require lower coverage thresholds and different distance metrics to account for the lower coverage depth and longer contig lengths.

The advantage of MetaBAT2 is its familiarity and extensive documentation. The limitation is that it does not use assembly graph information, and it may not perform as well as long-read-specific binners on complex communities. Researchers who use MetaBAT2 on long-read data should compare results with a long-read-specific binner and should expect to refine bins manually.

## Practical Workflow for Binning Long-Read Metagenomes

### Step 1: Assemble with a Long-Read or Hybrid Assembler

The binning step depends entirely on the quality of the assembly. Long-read metagenomic assemblies should be generated with an assembler that is designed for long reads or with a hybrid assembler that combines long and short reads. The choice of assembler affects contig length, assembly graph structure, and error rates, all of which influence binning.

Researchers should evaluate assembly statistics before binning, including contig N50, number of contigs, and the fraction of reads that map back to the assembly. The PacBio versus Illumina comparison demonstrated that hybrid assemblies using PacBio long reads and HiSeq contigs produced improvements in assembly statistics, including increases in average contig length, contig N50, and the number of large contigs. These improvements translate directly into better binning inputs.

### Step 2: Map Reads Back to the Assembly to Calculate Coverage

Coverage depth must be calculated for each contig by mapping the sequencing reads back to the assembled contigs. For long-read data, the mapper must be able to handle long reads and the specific error profile of the platform. The resulting BAM or SAM file is used to calculate per-contig coverage depth.

If multiple samples are available, coverage should be calculated for each sample separately, and the coverage profiles should be combined into a matrix where each row is a contig and each column is a sample. This differential coverage matrix provides a stronger signal for binning than single-sample coverage, because contigs from the same genome will have correlated coverage patterns across samples.

### Step 3: Run One or More Binning Tools

Researchers should run at least one long-read-specific binner and one general-purpose binner, then compare the results. For example, a researcher might run SemiBin2 and MetaBAT2 on the same assembly and compare the number and quality of bins produced by each tool. Running multiple binners is a form of ensemble binning, and the results can be aggregated to produce a consensus set of bins.

The choice of binner depends on the research question. If the goal is to recover novel taxa from a species-rich sample, LorBin may be the best choice. If the goal is to integrate assembly graph information, GraphMB is appropriate. If the researcher wants a familiar tool with extensive documentation, MetaBAT2 with adjusted parameters is a reasonable starting point.

### Step 4: Assess Bin Quality with Completeness and Contamination Metrics

Bin quality is assessed using single-copy marker genes. Completeness is the fraction of single-copy marker genes that are present in the bin, and contamination is the fraction of marker genes that are present in multiple copies, which indicates that the bin contains sequences from more than one genome. Tools such as CheckM or CheckM2 calculate these metrics.

For long-read assemblies, researchers should interpret these metrics with caution. Long-read error rates can cause marker genes to be missed or duplicated, and the quality metrics may not perfectly reflect the true completeness and contamination of the bin. Researchers should also consider the expected genome size and GC content of the bin and compare these to known relatives.

### Step 5: Refine Bins to Improve Quality

Refinement is the process of improving bin quality by removing contaminating contigs and adding missing contigs. This can be done automatically with bin refinement tools or manually with visualization tools. The goal is to maximize completeness while minimizing contamination, and the acceptable thresholds depend on the downstream application.

For downstream applications such as taxonomic assignment, functional annotation, or comparative genomics, researchers should aim for bins that meet the minimum quality thresholds for high-quality MAGs. The specific thresholds depend on the standards used by the research community and by journals, and researchers should consult the relevant standards for their field.

### Step 6: Validate Bins with Independent Evidence

Bin quality metrics are necessary but not sufficient for validating a MAG. Researchers should validate bins with independent evidence, such as the presence of a complete or near-complete set of core genes, the absence of marker genes from unrelated taxa, and the consistency of coverage and composition within the bin. For novel taxa, researchers should also verify that the bin is not a chimera of sequences from multiple organisms.

Validation is particularly important for long-read bins because the lower coverage depth and higher error rates can produce artifacts that are not captured by standard quality metrics. Researchers should examine the coverage distribution within each bin and should check for contigs with anomalous coverage or composition.

## Records and Measurements for Binning Reproducibility

Reproducibility is a core requirement for metagenomic analysis, and binning is no exception. Researchers should record the exact versions of all software used, including the assembler, mapper, binner, and quality assessment tool. They should also record all parameters and thresholds, because binning results can be sensitive to parameter choices.

The nf-core documentation provides standards for community pipelines that emphasize reproducibility, configuration, and usage. Researchers who use nf-core pipelines for metagenomic analysis can benefit from the standardized parameter recording and pipeline configuration that these pipelines provide. Similarly, the Galaxy Training Network offers accessible workflow training that emphasizes reproducibility in analysis workflows, and researchers can use these resources to design reproducible binning workflows.

The Carpentries lessons provide foundational training in computing, data, shell, Git, and programming, which are relevant for researchers who need to manage the computational aspects of binning workflows. Version control with Git is particularly important for tracking changes to analysis scripts and parameters.

Researchers should maintain a record of the following for each binning run:

- Sequencing platform and basecalling model
- Assembler version and parameters
- Mapper version and parameters
- Binner version and parameters
- Quality assessment tool version and parameters
- Input assembly statistics, including contig N50 and number of contigs
- Output bin statistics, including number of bins, completeness, and contamination

These records enable the researcher to reproduce the analysis, to compare results across datasets, and to report the methods accurately in publications.

## Common Failure Patterns in Long-Read Binning

### Failure Pattern 1: Low Coverage for Rare Genomes

Low-abundance organisms may have too few reads to produce a reliable coverage signal, and the binner may fail to recover their genomes or may merge them with other low-coverage contigs. This failure is more common in long-read data because long-read platforms have lower throughput than short-read platforms. Researchers can mitigate this by sequencing more deeply, by using multiple samples to increase the coverage signal, or by using a binner that is designed to handle imbalanced species distributions, such as LorBin.

### Failure Pattern 2: Chimeric Bins from Closely Related Genomes

Closely related species or strains may have similar composition and coverage, and the binner may merge them into a single bin. This produces a bin with high contamination, which is detected by the presence of multiple copies of single-copy marker genes. Assembly graph information can help resolve closely related genomes, because reads that span the junction between two genomes provide evidence that the contigs are not from the same genome. Researchers should use a binner that incorporates graph information, such as GraphMB, or should manually curate bins to separate closely related genomes.

### Failure Pattern 3: Fragmented Bins from Incomplete Assembly

If the assembly is fragmented, a single genome may be represented by many small contigs, and the binner may assign these contigs to different bins. This produces bins with low completeness. Long-read assemblies are less fragmented than short-read assemblies, but fragmentation can still occur in regions with extreme GC content or repetitive elements. Researchers should evaluate assembly quality before binning and should consider hybrid assembly to improve contiguity.

### Failure Pattern 4: Parameter Mismatch from Short-Read Defaults

General-purpose binners that were designed for short-read assemblies may have default parameters that are inappropriate for long-read data. For example, coverage thresholds may be too high for long-read data, or the distance metric may not account for the error profile of long reads. Researchers who use MetaBAT2 or other general-purpose binners on long-read assemblies should adjust parameters and should compare results with long-read-specific binners.

### Failure Pattern 5: Overlooked Novel Taxa

Standard binning methods may fail to identify novel taxa because they rely on known marker genes or because the novel taxa have unusual composition or coverage. LorBin was specifically designed to address this limitation, and it identified 2.4-17 times more novel taxa than state-of-the-art binning methods in its evaluation. Researchers who are interested in discovering novel taxa should use a binner that is designed for this purpose and should validate novel bins with additional evidence.

## Limitations and Interpretation of Binning Results

Binning is a heuristic process, and the resulting bins are hypotheses about the genomic content of microbial populations. A bin is not a confirmed genome until it has been validated with additional evidence, and even high-quality bins may contain errors. Researchers should interpret binning results with the following limitations in mind.

First, binning cannot resolve strain-level variation. If a sample contains multiple strains of the same species, the binner may produce a single bin that represents a consensus of the strains, or it may produce multiple bins that are difficult to distinguish. Long-read assemblies can improve strain resolution because longer contigs can capture strain-specific variation, but the binning step may still merge strains with similar composition and coverage.

Second, binning quality metrics are estimates. Completeness and contamination are calculated from the presence of single-copy marker genes, and these estimates depend on the completeness of the marker gene set and the accuracy of the assembly. Long-read error rates can cause marker genes to be missed or duplicated, which biases the quality estimates. Researchers should use multiple quality metrics and should validate bins with independent evidence.

Third, binning results are sensitive to the input assembly. Different assemblers produce different assemblies, and the binning results will differ accordingly. Researchers should not expect to obtain identical bins from different assembly pipelines, and they should report the assembly and binning parameters in their methods.

Fourth, the choice of binner affects the results. Different binners use different algorithms and features, and they may produce different bins from the same assembly. Ensemble binning, which aggregates results from multiple binners, can improve the robustness of the results, but it also adds complexity to the workflow.

## Safety and Regulatory Context for Metagenomic Data

Metagenomic data may contain sequences from pathogenic organisms, and researchers should follow institutional biosafety and data security policies when handling such data. The NCBI provides data resources and search systems for sequence data, and researchers should deposit their metagenomic data in appropriate public databases in accordance with journal and funding agency requirements. The NCBI data resources include databases for raw sequencing reads, assembled genomes, and metagenomic samples, and researchers should use these resources for data deposition and retrieval.

Researchers should also be aware of the ethical and regulatory considerations associated with metagenomic data from human-associated microbiomes. Human microbiome data may contain identifiable information, and researchers should follow institutional review board requirements and data sharing policies. The EMBL-EBI training resources provide guidance on bioinformatics data management and analysis, and researchers can use these resources to understand data sharing and ethical requirements.

For environmental metagenomic data, researchers should follow the access and benefit-sharing regulations that apply to genetic resources in their jurisdiction. These regulations vary by country and by the source of the samples, and researchers should consult the relevant authorities before conducting research that involves genetic resources from other countries.

## Professional Escalation Criteria for Binning Problems

Researchers should escalate binning problems to a bioinformatics specialist or collaborator when the standard workflow does not produce acceptable results. The following situations warrant escalation:

- The assembly produces very few contigs or very low contig N50, which indicates a problem with the sequencing or assembly that cannot be fixed by binning parameter adjustments.
- The binner produces no high-quality bins, or the completeness and contamination metrics are consistently poor across multiple binners and parameter settings.
- The binning results are inconsistent across replicates or across different binners, which indicates that the binning signal is weak or that the community is too complex for standard methods.
- The researcher needs to resolve strain-level variation or to recover genomes from very low-abundance organisms, which may require specialized methods or additional sequencing.
- The researcher is working with a novel or unusual microbial community for which standard binning tools have not been validated.

When escalating, the researcher should provide the assembly statistics, the binning parameters, the quality metrics, and the specific research question. This information enables the specialist to diagnose the problem and to recommend appropriate methods.

## A Decision Framework for Selecting Binning Tools and Parameters

Choosing a binning tool for long-read metagenomic data is not a one-time decision that applies to every dataset. The optimal choice depends on measurable properties of the assembly, the biological question, and the computational resources available. This section provides a structured decision framework that researchers can apply before running any binner, along with a record system for documenting binning decisions and a troubleshooting method for diagnosing poor results.

### Step 1: Characterize the Assembly Before Binning

Before selecting a binner, measure the following assembly properties and record them in a laboratory notebook or electronic log. These measurements determine which binning approach is appropriate.

**Contig N50 and contig count.** Compute the contig N50 and the total number of contigs longer than 1,000 base pairs. Long-read assemblies typically produce contig N50 values in the tens to hundreds of kilobases, whereas short-read assemblies often produce N50 values below 10 kilobases. If the long-read assembly has a contig N50 below 5 kilobases, the assembly may be fragmented, and binning will produce many incomplete bins regardless of the tool chosen. In this case, revisit the assembly step before binning.

**Coverage distribution.** Map reads back to the assembly and calculate per-contig coverage. Examine the distribution of coverage values across all contigs. A healthy long-read metagenome assembly shows a range of coverage values spanning several orders of magnitude, reflecting the natural abundance distribution of the community. If all contigs have similar coverage, the sample may contain a dominant organism, or the sequencing depth may be insufficient to resolve low-abundance members. Record the median coverage and the fraction of contigs with coverage below 5x, because these contigs will be difficult to bin reliably.

**Number of single-copy marker genes in the assembly.** Run a quick marker gene scan on the full assembly to estimate how many genomes are represented. This measurement provides an upper bound on the number of bins that can be recovered. If the assembly contains fewer marker gene sets than expected from the community composition, the assembly may have missed low-abundance organisms, and binning cannot recover what was not assembled.

**Assembly graph availability.** Check whether the assembler produced a graph file in a format that graph-based binners can consume. GraphMB requires the assembly graph as input, and not all assemblers export graphs in compatible formats. If the graph is unavailable, the researcher must either rerun the assembler with graph output enabled or restrict the choice to binners that do not require graph input.

### Step 2: Match the Binner to the Assembly Characteristics

Use the following decision rules to select a binner based on the measurements from Step 1. These rules are heuristics derived from the design goals of the tools described earlier, not guaranteed performance guarantees.

**Rule A: Species-rich samples with unknown taxa.** If the sample is expected to contain many species, including novel or uncharacterized taxa, and the assembly contains a large number of marker gene sets, use LorBin. LorBin was designed specifically to address the challenge of identifying unknown species and managing imbalanced species distributions in natural microbiomes. Its evaluation on oral, gut, and marine samples demonstrated that it generated 15-189% more high-quality MAGs and identified 2.4-17 times more novel taxa than state-of-the-art binning methods. This tool is the strongest choice when the research goal includes discovering previously uncharacterized microbial diversity.

**Rule B: Assembly graph available and closely related genomes suspected.** If the assembly graph is available and the community may contain closely related species or strains that are difficult to separate by composition or coverage alone, use GraphMB. GraphMB integrates the assembly graph into the binning process using graph neural networks, and its evaluation showed that it produced on average 17.5% more high-quality bins compared to state-of-the-art binners. The assembly graph provides evidence of physical linkage between contigs, which helps resolve genomes that have similar composition and coverage.

**Rule C: Balanced community with moderate complexity.** If the community is expected to have a relatively balanced species distribution and the researcher wants a tool with established documentation and community support, use SemiBin2. SemiBin2 is available through Bioconductor, which provides official package documentation and installation instructions for reproducible genomic analysis workflows. It was designed with long-read assemblies in mind and handles the coverage variation and error profiles typical of Nanopore and PacBio data.

**Rule D: Familiarity and parameter control.** If the researcher prefers a widely used tool with extensive documentation and is willing to adjust parameters manually, MetaBAT2 can be applied to long-read assemblies. However, the researcher must lower coverage thresholds and adjust distance metrics to account for the lower coverage depth and longer contig lengths of long-read data. MetaBAT2 does not use assembly graph information, so it may not perform as well as long-read-specific binners on complex communities.

**Rule E: Ensemble approach for critical results.** If the binning results will be used for a publication, a genome announcement, or a comparative genomics study, run at least two binners from different categories and compare the results. For example, run LorBin and GraphMB, or SemiBin2 and MetaBAT2. The aggregation of results from multiple binners can improve the robustness of the final bin set, as demonstrated by the GraphMB evaluation, which reported 13.7% more high-quality bins when aggregating its results with other binners.

### Step 3: Set Parameters Based on Measured Coverage

Parameter selection should follow directly from the coverage distribution measured in Step 1, not from default values intended for short-read data.

**Minimum contig length threshold.** Set the minimum contig length for binning based on the contig N50. For long-read assemblies with N50 above 20 kilobases, a minimum contig length of 1,000 to 2,500 base pairs is reasonable. For assemblies with lower N50, raising the minimum contig length to 2,500 or 5,000 base pairs reduces noise from short fragments that have unreliable composition and coverage signals. The tradeoff is that raising the threshold may exclude contigs from small genomes or from regions with extreme GC content that assembled poorly.

**Coverage filtering.** Identify contigs with coverage below a threshold that the binner cannot reliably cluster. This threshold depends on the sequencing depth and the binner. A common approach is to flag contigs with coverage below 2x or 3x as low-confidence and to examine whether the binner assigns them to bins or leaves them unclustered. If the binner assigns many low-coverage contigs to a single bin, the bin may be a chimera of unrelated fragments that happen to have similar low coverage.

**Composition window size.** Some binners allow the user to set the k-mer size or composition window. For long-read assemblies, the default k-mer size may be appropriate, but researchers should verify that the binner accounts for the error profile of the sequencing platform. Older basecalling models produce higher error rates that can distort k-mer frequencies, and researchers should consider whether the assembly was polished before binning.

### Step 4: Record Binning Decisions in a Structured Log

Maintain a structured record for each binning run that includes the following fields. This record supports reproducibility and enables troubleshooting when results are poor.

| Field | Example Entry |
| --- | --- |
| Sample identifier | AD_biogas_01 |
| Sequencing platform | PacBio Sequel II |
| Basecalling model | CCS or HiFi |
| Assembler and version | Flye 2.9 |
| Assembly contig N50 | 45,200 bp |
| Number of contigs over 1 kb | 12,400 |
| Median coverage | 18x |
| Fraction of contigs below 5x coverage | 0.12 |
| Mapper and version | minimap2 2.24 |
| Binner and version | LorBin 1.0 |
| Binner parameters | min_contig_length 2500, coverage_threshold 3 |
| Number of bins produced | 87 |
| Number of high-quality bins | 34 |
| Quality assessment tool and version | CheckM2 1.0 |
| Date and operator | 2025-06-15, J. Smith |

The nf-core documentation provides standards for community pipelines that emphasize reproducibility, configuration, and usage. Researchers who use nf-core pipelines can benefit from the standardized parameter recording that these pipelines provide. The Galaxy Training Network offers accessible workflow training that emphasizes reproducibility in analysis workflows, and researchers can use these resources to design reproducible binning workflows. The Carpentries lessons provide foundational training in computing, data, shell, Git, and programming, which are relevant for managing the computational aspects of binning workflows and for tracking changes to analysis scripts with version control.

### Troubleshooting Method for Poor Binning Results

When a binning run produces poor results, defined as few high-quality bins, high contamination across many bins, or inconsistent results across replicates, use the following systematic troubleshooting method. Do not change parameters randomly. Instead, isolate the cause by testing one variable at a time.

**Troubleshooting Step 1: Verify the assembly input.** Check whether the assembly statistics match the expectations for the sequencing platform and sample type. If the contig N50 is far below the platform expectation, the assembly may have failed, and no binner will produce good results. The PacBio versus Illumina comparison demonstrated that hybrid assemblies using PacBio long reads and HiSeq contigs produced improvements in assembly statistics, including increases in average contig length, contig N50, and the number of large contigs. If the assembly is fragmented, consider hybrid assembly to improve contiguity before binning.

**Troubleshooting Step 2: Examine the coverage distribution.** Plot the coverage distribution across all contigs. If the distribution is narrow, with most contigs at similar coverage, the binner will have difficulty separating genomes by coverage. This situation can arise when sequencing depth is too low to sample the community adequately. The same PacBio study noted that lower single-pass accuracy can make gene discovery and assembly for low-abundance organisms difficult, which implies that coverage signals for rare genomes are weak in long-read data. If the coverage distribution is narrow, consider deeper sequencing or the use of multiple samples to create a differential coverage matrix.

**Troubleshooting Step 3: Run a second binner from a different category.** If the first binner produced poor results, run a binner that uses a different algorithmic approach. For example, if a composition-based binner produced chimeric bins, run a graph-based binner such as GraphMB. If a coverage-based binner failed to recover low-abundance genomes, run LorBin, which was designed to handle imbalanced species distributions. Comparing results across binners reveals whether the problem is specific to one tool or inherent to the dataset.

**Troubleshooting Step 4: Inspect the worst bins manually.** Select the bins with the highest contamination and the bins with the lowest completeness, and examine their contig composition. Check whether the contaminating contigs have anomalous coverage or composition compared to the rest of the bin. Check whether the incomplete bins contain contigs that were assigned to other bins but have similar coverage and composition. This manual inspection often reveals whether the binner is merging closely related genomes or splitting a single genome across multiple bins.

**Troubleshooting Step 5: Adjust one parameter at a time.** If the manual inspection suggests a specific cause, adjust the corresponding parameter. For example, if closely related genomes are merged, increase the sensitivity of the clustering or switch to a graph-based binner. If low-coverage contigs are causing chimeric bins, raise the minimum contig length or the coverage threshold. After each parameter change, rerun the binner and record the results. Do not change multiple parameters simultaneously, because the effect of each change cannot be isolated.

**Troubleshooting Step 6: Escalate with a complete record.** If the troubleshooting steps do not resolve the problem, escalate to a bioinformatics specialist or collaborator. Provide the complete record from Step 4, including the assembly statistics, the binner parameters, the quality metrics, and the results of the manual inspection. The specialist needs this information to diagnose the problem and to recommend appropriate methods. Situations that warrant escalation include assemblies with very low contig N50, consistently poor quality metrics across multiple binners and parameter settings, inconsistent results across replicates, and the need to resolve strain-level variation or to recover genomes from very low-abundance organisms.

### Common Failure Patterns and Their Diagnostic Signatures

The following failure patterns have characteristic signatures that can be identified from the records and measurements described above.

**Low completeness across many bins.** This pattern indicates that the assembly is fragmented or that the binner is splitting genomes across multiple bins. Check the contig N50 and the number of contigs per bin. If bins contain many small contigs, the assembly is likely fragmented, and hybrid assembly should be considered. If bins contain few contigs but completeness is low, the binner may be too conservative and may be leaving contigs unclustered.

**High contamination in a few bins.** This pattern indicates that the binner is merging closely related genomes or that low-coverage contigs from unrelated organisms are being assigned to the same bin. Check whether the contaminating contigs have similar coverage to the rest of the bin. If they do, the genomes may be closely related and require graph-based binning to resolve. If the contaminating contigs have very low coverage, the binner may be clustering noise.

**No high-quality bins at all.** This pattern indicates a fundamental problem with the assembly or the binning approach. Verify that the assembly contains a reasonable number of single-copy marker genes. If the marker gene count is very low, the assembly may have missed most of the community. If the marker gene count is reasonable but no bins pass quality thresholds, the binner parameters may be inappropriate for the data, or the community may be too complex for standard methods.

**Inconsistent results across replicates.** This pattern indicates that the binning signal is weak or that the binner is sensitive to small changes in the input. Check whether the coverage distribution is narrow, which would weaken the coverage signal. Check whether the assembly is fragmented, which would weaken the composition signal. If the signal is weak, consider deeper sequencing or the use of multiple samples for differential coverage binning.

### Integration with Reproducible Workflow Standards

The decision framework and record system described in this section align with the reproducibility standards promoted by community pipeline projects. The nf-core documentation provides standards for community pipelines that emphasize reproducibility, configuration, and usage, and researchers who use nf-core pipelines can integrate the binning record into the pipeline configuration files. The Galaxy Training Network offers accessible workflow training that emphasizes reproducibility in analysis workflows, and researchers can use these resources to design reproducible binning workflows that include the decision framework as a documented step.

The EMBL-EBI training resources provide guidance on bioinformatics data management and analysis, and researchers can use these resources to understand how to structure the binning record for data sharing and publication. The NCBI provides data resources and search systems for sequence data, and researchers should deposit their metagenomic data and the associated binning records in appropriate public databases in accordance with journal and funding agency requirements.

The decision framework presented here is not a substitute for domain expertise or for the judgment of an experienced bioinformatician. It is a structured approach that helps researchers make consistent, documented decisions and that provides a clear path for troubleshooting when results are poor. By characterizing the assembly before binning, matching the binner to the assembly characteristics, setting parameters based on measured coverage, and recording all decisions in a structured log, researchers can improve the reproducibility of their binning results and reduce the time spent on trial-and-error parameter adjustment.

## Frequently Asked Questions

### What is the difference between coverage-based and composition-based binning?

Coverage-based binning groups contigs by their depth of coverage across one or more samples, under the assumption that contigs from the same genome have similar abundance. Composition-based binning groups contigs by their nucleotide composition, typically tetranucleotide frequency, under the assumption that different genomes have characteristic composition signatures. Long-read binning tools often combine both signals, and some also use assembly graph connectivity.

### Why do short-read binning tools fail on long-read assemblies?

Short-read binning tools were designed for the characteristics of short-read assemblies, which have many small contigs and rely heavily on coverage signals. Long-read assemblies have fewer, longer contigs, and the coverage depth is lower and more uneven. Short-read binners may use coverage thresholds that are too high for long-read data, and they may not account for the error profile of long reads or for the assembly graph structure.

### Which binning tool should I use for Nanopore or PacBio data?

The choice of tool depends on the research question and the characteristics of the data. SemiBin2 is a long-read-specific binner that is available through Bioconductor. GraphMB uses assembly graph information and is appropriate when the assembly graph is available. LorBin is designed for species-rich samples with unknown taxa and imbalanced species distributions. MetaBAT2 can be used with adjusted parameters, but researchers should compare its results with a long-read-specific binner.

### How do I calculate coverage depth for long-read binning?

Coverage depth is calculated by mapping the sequencing reads back to the assembled contigs and counting the number of reads that map to each contig, normalized by contig length. The mapper must be able to handle long reads and the error profile of the sequencing platform. If multiple samples are available, coverage should be calculated for each sample separately and combined into a differential coverage matrix.

### What completeness and contamination thresholds should I use for high-quality MAGs?

The acceptable thresholds depend on the standards used by the research community and by journals. Researchers should consult the relevant standards for their field and should aim for bins that meet the minimum thresholds for high-quality MAGs. Researchers should also validate bins with independent evidence, because quality metrics are estimates that can be biased by long-read error rates.

### How can I improve bin quality through refinement?

Refinement involves removing contaminating contigs and adding missing contigs to improve completeness and reduce contamination. This can be done automatically with bin refinement tools or manually with visualization tools. Researchers should examine the coverage distribution and composition of contigs within each bin and should use assembly graph information to identify contigs that are incorrectly assigned.

### What should I do if my binning results are poor?

If the standard workflow does not produce acceptable results, researchers should first evaluate the assembly quality, because binning cannot recover genomes from a poor assembly. Researchers should then try multiple binners and parameter settings, and should consider ensemble binning to aggregate results. If the results remain poor, the researcher should escalate to a bioinformatics specialist with the assembly statistics, binning parameters, and quality metrics.

### How do I report binning methods in a publication?

Researchers should report the exact versions of all software used, including the assembler, mapper, binner, and quality assessment tool, and all parameters and thresholds. They should report the assembly statistics and the bin quality metrics, and they should describe the refinement steps. The nf-core documentation and the Galaxy Training Network provide standards for reproducible workflows that can guide method reporting.

## Related Bioinformatics Guides

- [Evaluating Metagenomic Assembly Tools: A Benchmarking Framework for Short-Read and Long-Read Data](/knowledge/bioinformatics/evaluating-metagenomic-assembly-tools-a-benchmarking-framework-for-short-read-and-long-read-data)
- [Long-Read Metagenome Assembly: Overcoming Challenges with Nanopore and PacBio Data](/knowledge/bioinformatics/long-read-metagenome-assembly-overcoming-challenges-with-nanopore-and-pacbio-data)
- [How to Choose a Long-Read Sequencing Platform: PacBio vs Oxford Nanopore](/knowledge/bioinformatics/how-to-choose-a-long-read-sequencing-platform-pacbio-vs-oxford-nanopore)
- [Metagenomic Binning Tools Benchmark: How to Evaluate and Choose](/knowledge/bioinformatics/metagenomic-binning-tools-benchmark-how-to-evaluate-and-choose)
- [Genomic Data Analysis Tools: A Comparative Guide for Researchers](/knowledge/bioinformatics/genomic-data-analysis-tools-a-comparative-guide-for-researchers)

## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [Metagenomic tools in microbial ecology research.](https://pubmed.ncbi.nlm.nih.gov/33592536). Current opinion in biotechnology, 2021.
- [Long-Read Metagenomics and CAZyme Discovery.](https://pubmed.ncbi.nlm.nih.gov/37149537). Methods in molecular biology (Clifton, N.J.), 2023.
- [PacBio Long Reads Improve Metagenomic Assemblies, Gene Catalogs, and Genome Binning.](https://pubmed.ncbi.nlm.nih.gov/33101371). Frontiers in genetics, 2020.
- [Metagenomic binning with assembly graph embeddings.](https://pubmed.ncbi.nlm.nih.gov/35972375). Bioinformatics (Oxford, England), 2022.
- [LorBin: efficient binning of long-read metagenomes by multiscale adaptive clustering and evaluation.](https://pubmed.ncbi.nlm.nih.gov/41136439). Nature communications, 2025.

> This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.