# A Benchmark of Haplotype Phasing Tools for Long-Read Sequencing: Whatshap, HapCUT2, and MarginPhase Compared

Researchers working with PacBio or Oxford Nanopore long-read data face a practical decision when haplotype phasing is required: which tool to run, at what coverage, and with what expectation of accuracy. This article provides a systematic comparison of three widely used phasing tools, WhatsHap, HapCUT2, and MarginPhase, with emphasis on accuracy, computational cost, and ease of integration into existing bioinformatics workflows. The guidance is intended for biology students, laboratory professionals, and researchers who need to select a phasing strategy for diploid or polyploid datasets and who want to understand the tradeoffs before committing compute time and storage.

## Scope and Reader Context

Haplotype phasing is the computational process of assigning genetic variants to parental chromosomes. For long-read sequencing data, phasing tools use reads that span multiple heterozygous variants to infer which variants travel together on the same homolog. The three tools compared here represent distinct algorithmic approaches: WhatsHap uses a dynamic programming formulation that guarantees optimal solutions under certain conditions, HapCUT2 uses a max-cut heuristic on a graph of variant connections, and MarginPhase applies a probabilistic model that can integrate base quality and alignment uncertainty.

This benchmark is relevant to researchers who have already generated long-read data and need to phase variants for downstream analysis such as allele-specific expression, structural variant characterization, or polyploid genome assembly. The comparison focuses on practical outcomes: which tool produces the fewest switch errors at a given coverage, how long each tool takes to run, and how much hands-on effort is required to install and execute each program.

The scope excludes reference-free assembly-based phasing and Hi-C based phasing, which are covered by other tools such as PhaseGrass for highly heterozygous plant genomes and refLinker for cancer karyotype reconstruction. Those approaches are mentioned where relevant to help readers understand when read-based phasing is appropriate and when it is not.

## At a Glance

The table below summarizes the key characteristics of the three tools compared in this benchmark. The entries reflect typical behavior observed in published evaluations and community usage patterns.

| Tool | Algorithm Basis | Input Requirements | Typical Use Case | Relative Runtime | Ease of Installation |
|------|-----------------|-------------------|------------------|------------------|---------------------|
| WhatsHap | Dynamic programming with optimal phasing of local blocks | BAM/CRAM plus VCF, requires base quality scores | Diploid phasing, integration into pipelines, polyploid support | Moderate | High, available via Bioconda and pip |
| HapCUT2 | Max-cut heuristic on fragment graph | BAM/CRAM plus VCF, can use linked reads | Diploid and polyploid phasing, large datasets | Fast | High, available via Bioconda and GitHub |
| MarginPhase | Probabilistic graphical model with alignment uncertainty | BAM/CRAM plus VCF, uses base quality and alignment features | Diploid phasing, error-prone long reads | Slower | Moderate, requires compilation |

All three tools accept aligned long reads in BAM or CRAM format and a VCF file containing candidate variants. The practical choice depends on whether the user prioritizes switch accuracy, runtime, or the ability to handle polyploid genomes.

## Core Principles of Read-Based Phasing

### How Phasing Information Is Extracted from Long Reads

Long-read sequencing produces molecules that are substantially longer than typical Illumina fragments. A single PacBio or Nanopore read can span multiple heterozygous single-nucleotide polymorphisms, which creates the opportunity to observe which alleles co-occur on the same physical DNA molecule. The phasing tool collects these co-occurrence observations across all reads and attempts to reconstruct the two parental haplotypes.

The fundamental unit of phasing is the fragment, which is the set of variant sites covered by a single read. Each fragment provides evidence that the alleles observed at those sites are linked on the same chromosome copy. When multiple fragments support the same allele combination, the phasing tool can assign variants to haplotypes with increasing confidence.

Base quality scores matter because a sequencing error at a variant site can create a false observation of linkage. Tools that incorporate base quality into their models, such as WhatsHap and MarginPhase, can downweight low-quality observations. Tools that rely primarily on the presence or absence of alleles, such as HapCUT2, may be more sensitive to sequencing errors but can still perform well when coverage is sufficient.

### The Role of Coverage and Read Length

Coverage and read length are the two most important data characteristics that determine phasing accuracy. Higher coverage means more fragments supporting each variant connection, which reduces the chance of spurious phasing decisions. Longer reads mean that each fragment spans more variant sites, which increases the contiguity of the phased blocks.

At low coverage, around 8x, phasing accuracy drops for all tools because the number of fragments connecting distant variants is small. A study of the Duet tool for structural variant calling and phasing using Oxford Nanopore data found that at 8x coverage, dedicated SV phasing tools outperformed general-purpose phasing approaches, but at 20x to 40x coverage, phasing accuracy improved substantially across tools. This finding suggests that researchers working with low-coverage datasets should expect reduced phasing performance and may need to consider coverage-specific tools.

Read length interacts with coverage in a nonlinear way. A dataset with 20x coverage of 10 kilobase reads provides more phasing information than a dataset with 20x coverage of 2 kilobase reads, because each read spans more variant sites. Researchers should evaluate their expected read length distribution before selecting a phasing tool, because tools that model alignment uncertainty, such as MarginPhase, may extract more information from error-prone long reads.

### Variant Calling as a Prerequisite

All three tools require a VCF file as input. The variants in this file must be called before phasing can proceed. The quality of the variant calls directly affects phasing accuracy, because a false positive variant creates a spurious site that must be phased, and a false negative variant removes a site that could have connected two phased blocks.

For long-read data, variant calling is typically performed with tools designed for the specific sequencing platform. PacBio data may be called with platform-specific callers, while Nanopore data requires callers that account for the higher error rate of the platform. The phasing tool cannot correct errors in the input VCF, so researchers should invest effort in producing high-quality variant calls before running any phasing tool.

The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide access to reference genomes and variant databases that can be used to validate variant calls and phasing results. Researchers can compare their phased variants against known haplotype structures in public databases to assess the quality of their phasing output.

## Practical Workflow for Phasing Long-Read Data

### Step 1: Align Reads to a Reference Genome

The first step in any read-based phasing workflow is aligning the long reads to a reference genome. The alignment must be performed with a long-read aware aligner that can handle the high error rates and large indels characteristic of PacBio and Nanopore data. The output should be a coordinate-sorted BAM or CRAM file with an index.

Alignment quality directly affects phasing accuracy. Reads that are misaligned create false variant connections, and reads that are clipped or split may lose phasing information. Researchers should inspect alignment statistics, including the fraction of reads mapped, the median read length, and the distribution of mapping quality scores, before proceeding to phasing.

The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible tutorials on long-read alignment and preprocessing that can help researchers establish a reliable alignment workflow. These tutorials cover quality control, alignment, and post-alignment processing steps that are prerequisites for phasing.

### Step 2: Call Variants

Variant calling for long-read data requires a caller that can handle the platform-specific error profiles. The output VCF should contain high-confidence single-nucleotide polymorphisms and small indels. Structural variants are typically handled separately, and tools such as Duet have been developed specifically for SV calling and phasing using Nanopore data.

The variant caller should be run with parameters appropriate for the sequencing platform and coverage. At low coverage, variant calling sensitivity decreases, which means that some true variants will be missing from the VCF. These missing variants create gaps in the phasing graph and reduce the contiguity of phased blocks.

Researchers should filter the VCF to remove low-quality calls before phasing. The filtering criteria depend on the variant caller and the sequencing platform, but common filters include minimum genotype quality, minimum depth, and minimum allele balance. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide guidance on variant calling and filtering for various sequencing platforms.

### Step 3: Run the Phasing Tool

Each phasing tool has its own command-line interface and input requirements. The general workflow is to provide the aligned reads and the VCF file, and the tool outputs a phased VCF where each variant is assigned to a haplotype block.

WhatsHap is typically run with the `whatshap phase` command, which takes the VCF and BAM files as input. The tool can phase diploid and polyploid samples and produces a phased VCF with haplotype block information. WhatsHap can also be used for what is called "whatshap polyphase" for polyploid samples, which uses a different algorithm optimized for multiple haplotypes.

HapCUT2 is run with the `hapcut2` command, which requires the VCF and BAM files. The tool builds a fragment graph and applies a max-cut heuristic to partition variants into haplotypes. HapCUT2 can handle polyploid samples and is often faster than WhatsHap on large datasets.

MarginPhase is run with the `marginPhase` command, which takes the BAM file and a VCF file. The tool uses a probabilistic model that incorporates base quality scores and alignment uncertainty. MarginPhase is designed for error-prone long reads and may produce more accurate results on Nanopore data, but it is computationally more expensive.

### Step 4: Evaluate Phasing Quality

After running a phasing tool, researchers should evaluate the quality of the output before using the phased haplotypes for downstream analysis. The key metrics are the number of phased variants, the number of haplotype blocks, the N50 block length, and the switch error rate if a ground truth is available.

The number of phased variants indicates how much of the input VCF was successfully assigned to haplotypes. Variants that cannot be phased are left unassigned, which reduces the utility of the output. The number of haplotype blocks and the N50 block length indicate the contiguity of the phasing. Longer blocks are more useful for downstream analysis because they span more variants.

Switch error rate can only be calculated when a ground truth haplotype is available, such as from a trio or from a previously validated dataset. In the absence of ground truth, researchers can assess phasing quality by checking for Mendelian consistency in family data or by comparing results across multiple phasing tools.

The [nf-core Documentation](https://nf-co.re/docs) describes community standards for reproducible bioinformatics workflows, including best practices for documenting phasing parameters and evaluating output quality. Following these standards helps ensure that phasing results are reproducible and interpretable.

## Options and Tradeoffs Across the Three Tools

### WhatsHap: Optimal Phasing with Higher Computational Cost

WhatsHap implements a dynamic programming algorithm that finds the maximum likelihood phasing of variants within a block. The algorithm guarantees an optimal solution for the given objective function, which is a significant theoretical advantage over heuristic approaches. WhatsHap can phase diploid and polyploid samples, and it incorporates base quality scores into the likelihood calculation.

The main tradeoff is computational cost. The dynamic programming approach becomes expensive as the number of variants in a block increases, and WhatsHap can be slow on high-coverage datasets with long reads. The tool provides options to limit the block size and to use a faster but less accurate mode, which can help manage runtime.

WhatsHap is well integrated into bioinformatics ecosystems. It is available through Bioconda and can be installed with a single command. The [Bioconductor Project](https://bioconductor.org/) provides documentation and workflows that include WhatsHap for phasing in genomic analyses, which can help researchers integrate the tool into existing pipelines.

For polyploid phasing, WhatsHap offers the `polyphase` subcommand, which uses a different algorithm designed for multiple haplotypes. This feature is valuable for researchers working with polyploid crops or other organisms with more than two chromosome copies. The nTChap method for polyploid haplotype reconstruction demonstrates that polyploid phasing remains challenging, and researchers should evaluate whether WhatsHap polyphase meets their accuracy requirements.

### HapCUT2: Speed and Scalability

HapCUT2 uses a max-cut heuristic to partition variants into haplotypes. The algorithm builds a graph where vertices represent variants and edges represent connections supported by reads. The max-cut formulation seeks to partition the graph into two sets that maximize the weight of edges crossing the partition, which corresponds to assigning variants to two haplotypes.

The heuristic approach is faster than the dynamic programming used by WhatsHap, particularly on large datasets. HapCUT2 can handle high-coverage data and long reads without the same computational scaling issues. The tool also supports polyploid phasing, although the accuracy on polyploid data may be lower than on diploid data.

The main tradeoff is that the max-cut heuristic does not guarantee an optimal solution. In practice, HapCUT2 produces accurate phasing on most datasets, but there may be cases where the heuristic converges to a suboptimal partition. Researchers who need guaranteed optimal phasing should consider WhatsHap, while those who prioritize speed and scalability may prefer HapCUT2.

HapCUT2 is available through Bioconda and GitHub, and it can be installed with minimal effort. The tool accepts BAM and VCF input and produces a phased VCF output. The [The Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in command-line tools and shell scripting that can help researchers run HapCUT2 and other phasing tools effectively.

### MarginPhase: Probabilistic Modeling for Error-Prone Reads

MarginPhase uses a probabilistic graphical model that incorporates base quality scores and alignment uncertainty. The model can account for the possibility that a read is misaligned at a variant site, which is particularly relevant for error-prone long-read platforms such as Oxford Nanopore.

The main advantage of MarginPhase is its potential for higher accuracy on error-prone data. By modeling alignment uncertainty, the tool can avoid making phasing decisions based on reads that are likely misaligned. This can reduce switch errors in datasets where alignment errors are common.

The main tradeoff is computational cost. MarginPhase is slower than HapCUT2 and often slower than WhatsHap, particularly on high-coverage datasets. The tool also requires compilation, which can be a barrier for researchers who are not comfortable building software from source.

MarginPhase is less widely used than WhatsHap and HapCUT2, which means there is less community support and fewer tutorials available. Researchers who choose MarginPhase should be prepared to invest time in installation and parameter tuning. The [Galaxy Training Network](https://training.galaxyproject.org/) offers general training on long-read analysis that can help researchers understand the data characteristics that MarginPhase is designed to handle.

## Observations and Measurements from Published Evaluations

### Accuracy at Different Coverage Levels

Published evaluations of phasing tools consistently show that accuracy improves with coverage, but the rate of improvement varies across tools. At low coverage, around 8x, all tools produce higher switch error rates because the number of fragments connecting distant variants is small. A study of the Duet tool for SV phasing using Nanopore data found that at 8x coverage, dedicated SV phasing tools outperformed general-purpose phasing approaches, suggesting that low-coverage data requires specialized methods.

At moderate to high coverage, 20x to 40x, phasing accuracy improves substantially. The same Duet study found that at these coverage levels, the F1-score for SV phasing improved compared to low coverage, and the performance of SV genotyping and SV calling remained higher than other tools. This finding suggests that researchers working with low-coverage data should expect reduced phasing performance and may need to consider coverage-specific tools.

For general variant phasing, the relationship between coverage and accuracy is similar. WhatsHap and HapCUT2 both show improved switch error rates as coverage increases, but the magnitude of improvement depends on the read length distribution and the density of heterozygous variants. Researchers should aim for at least 20x coverage for reliable phasing, and higher coverage may be needed for polyploid samples or for genomes with low heterozygosity.

### Read Length Effects on Phasing Contiguity

Read length has a direct effect on the contiguity of phased blocks. Longer reads span more variant sites, which creates more connections between variants and allows the phasing tool to build longer blocks. A dataset with 10 kilobase reads will produce longer phased blocks than a dataset with 2 kilobase reads at the same coverage.

The relationship between read length and phasing accuracy is particularly important for Nanopore data, where read length distributions can vary substantially between runs. Researchers should check the read length N50 of their dataset before phasing and should consider whether the read length is sufficient for their phasing goals.

MarginPhase is designed to handle error-prone long reads and may extract more phasing information from Nanopore data than tools that do not model alignment uncertainty. However, the advantage of MarginPhase depends on the specific error profile of the data, and researchers should evaluate the tool on their own data before committing to a workflow.

### Polyploid Phasing Challenges

Polyploid phasing is substantially more challenging than diploid phasing because the phasing algorithm must assign variants to more than two haplotypes. The number of possible haplotype combinations grows rapidly with ploidy, and the computational cost of finding the optimal phasing increases accordingly.

The nTChap method for polyploid haplotype reconstruction addresses some of these challenges by using an iterative procedure involving two-round clustering and consensus construction. The method outperforms current tools in terms of accuracy and error rates on simulated datasets, while maintaining robust performance across high-ploidy genomes and low-coverage data. This finding suggests that polyploid phasing may require specialized tools beyond the three compared in this benchmark.

WhatsHap polyphase and HapCUT2 both support polyploid phasing, but their accuracy on polyploid data may be lower than on diploid data. Researchers working with polyploid genomes should evaluate multiple tools and should consider whether a specialized polyploid phasing method is needed for their application.

## Records and Measurements for Phasing Quality Assessment

### Metrics to Record for Every Phasing Run

Researchers should record the following metrics for every phasing run to enable comparison across tools and datasets:

The number of variants in the input VCF and the number of variants successfully phased. The difference between these two numbers indicates how much phasing information was lost.

The number of haplotype blocks and the N50 block length. These metrics describe the contiguity of the phasing output. Longer blocks are more useful for downstream analysis.

The switch error rate, if a ground truth is available. This metric measures the fraction of adjacent variants that are assigned to the wrong haplotype relative to the true phasing.

The runtime and peak memory usage. These metrics are important for planning computational resources and for comparing the efficiency of different tools.

The phasing tool version and all parameters used. This information is essential for reproducibility and for troubleshooting unexpected results.

### How to Calculate Switch Error Rate

Switch error rate is calculated by comparing the phased haplotypes to a ground truth. The ground truth can come from a trio where the parental haplotypes are known, from a previously validated dataset, or from a computational phasing of a higher-quality dataset.

The calculation involves walking along the phased blocks and counting the number of positions where the phasing switches from the correct haplotype to the incorrect haplotype. A switch error occurs when the relative order of two adjacent variants is flipped between the phased output and the ground truth.

Switch error rate is reported as the number of switch errors divided by the total number of opportunities for a switch, which is the number of adjacent variant pairs in the phased blocks. A lower switch error rate indicates more accurate phasing.

### Using Public Databases for Validation

The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide access to reference genomes, variant databases, and population-level haplotype information that can be used to validate phasing results. Researchers can compare their phased variants against known haplotype structures in public databases to assess the quality of their phasing output.

For human data, the 1000 Genomes Project and other population-scale projects have produced haplotype reference panels that can be used for validation. For non-human species, the availability of reference haplotypes depends on the species and the research community.

Researchers should be cautious when using public databases for validation, because the reference haplotypes may not match the population or the sequencing platform used in their study. The validation should be interpreted as a general check on phasing quality instead of a definitive assessment.

## Common Failure Patterns in Phasing Workflows

### Failure Pattern 1: Low Coverage Produces Fragmented Phasing

The most common failure pattern in phasing workflows is fragmented output caused by low coverage. When coverage is below approximately 15x, the number of reads spanning multiple variant sites is small, and the phasing tool cannot connect distant variants into long blocks. The output contains many short blocks, and a substantial fraction of variants remain unphased.

Researchers who observe fragmented phasing should first check the coverage of their dataset. If coverage is low, they may need to generate additional sequencing data or accept the reduced phasing contiguity. The Duet study found that at 8x coverage, SV phasing accuracy was lower than at higher coverage, which suggests that low-coverage data requires specialized methods or additional data generation.

### Failure Pattern 2: Variant Calling Errors Propagate to Phasing

Errors in the input VCF propagate directly to the phasing output. False positive variants create spurious sites that must be phased, which can introduce switch errors. False negative variants remove sites that could have connected phased blocks, which reduces contiguity.

Researchers who observe poor phasing quality should examine the input VCF for signs of variant calling errors. Common signs include an unusually high number of variants in low-complexity regions, a skewed transition-to-transversion ratio, or a high fraction of variants with low genotype quality.

The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide guidance on variant calling quality control that can help researchers identify and correct variant calling errors before phasing.

### Failure Pattern 3: Misaligned Reads Create False Connections

Misaligned reads create false connections between variants that are not actually linked on the same chromosome copy. This can happen when a read is aligned to the wrong genomic location or when a read contains a large indel that causes misalignment at variant sites.

Tools that model alignment uncertainty, such as MarginPhase, are designed to be robust to misalignment, but no tool can fully correct for systematic alignment errors. Researchers should inspect alignment quality metrics, including the fraction of reads with high mapping quality and the distribution of soft-clipped bases, before running phasing tools.

### Failure Pattern 4: Polyploid Data Confuses Diploid Tools

Running a diploid phasing tool on polyploid data produces incorrect results because the tool attempts to assign variants to two haplotypes when there are more than two. The output may contain switch errors at positions where the true phasing involves more than two haplotypes.

Researchers working with polyploid genomes should use a phasing tool that explicitly supports polyploid phasing, such as WhatsHap polyphase or HapCUT2. They should also consider whether a specialized polyploid phasing method, such as nTChap, is needed for their application.

## Limitations of Read-Based Phasing

### Inability to Phase Variants in Low-Complexity Regions

Read-based phasing cannot phase variants in regions where reads cannot be uniquely aligned. These regions include repetitive sequences, segmental duplications, and other low-complexity genomic features. Reads from these regions may align to multiple locations, and the phasing tool cannot determine which location is correct.

Researchers should be aware that phased blocks will have gaps in low-complexity regions. These gaps are not errors in the phasing tool but rather limitations of the read-based approach. Hi-C based phasing methods, such as refLinker for cancer karyotype reconstruction, can provide chromosome-length continuity that read-based phasing cannot achieve.

### Sensitivity to Sequencing Platform Error Profiles

The accuracy of read-based phasing depends on the error profile of the sequencing platform. PacBio HiFi data has a lower error rate than Oxford Nanopore data, which means that phasing tools can make more confident decisions on PacBio data. Nanopore data requires tools that can handle higher error rates, and the accuracy of phasing may be lower even with error-aware tools.

Researchers should consider the error profile of their sequencing platform when selecting a phasing tool and when interpreting phasing results. MarginPhase is designed for error-prone reads, but its advantage over other tools depends on the specific error profile of the data.

### Computational Cost on Large Datasets

Read-based phasing can be computationally expensive on large datasets, particularly when coverage is high and reads are long. The dynamic programming approach used by WhatsHap becomes expensive as the number of variants in a block increases, and the probabilistic model used by MarginPhase requires substantial computation.

Researchers should estimate the computational cost of phasing before running a tool on a large dataset. The [nf-core Documentation](https://nf-co.re/docs) provides guidance on configuring computational resources for bioinformatics workflows, which can help researchers plan for the resource requirements of phasing.

## Safety and Reproducibility Context

### Reproducibility Standards for Phasing Workflows

Reproducibility is a critical concern in bioinformatics, and phasing workflows are no exception. Researchers should document the exact version of each tool used, the parameters applied, and the input data versions. This documentation enables other researchers to reproduce the phasing results and to compare results across studies.

The [nf-core Documentation](https://nf-co.re/docs) describes community standards for reproducible bioinformatics workflows, including version pinning, parameter documentation, and containerization. Following these standards helps ensure that phasing results are reproducible and interpretable.

The [The Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in reproducible research practices, including version control with Git and documentation of computational workflows. These skills are essential for researchers who want to produce reproducible phasing results.

### Data Management for Phasing Projects

Phasing projects generate large intermediate files, including aligned reads, variant calls, and phased output. Researchers should establish a data management plan that includes storage for these files, backup procedures, and documentation of file versions.

The [Bioconductor Project](https://bioconductor.org/) provides documentation on reproducible genomic-analysis workflows, including data management practices that can be applied to phasing projects. The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible tutorials on data management for bioinformatics workflows.

### Professional Escalation Criteria

Researchers should escalate phasing issues to a bioinformatics specialist or a colleague with phasing expertise when they encounter any of the following situations:

The phasing output contains an unexpectedly high switch error rate, and the cause cannot be identified from the input data or the phasing parameters.

The phasing tool crashes or produces an error that cannot be resolved by consulting the tool documentation or community forums.

The phasing results are inconsistent across multiple tools, and the discrepancies cannot be explained by differences in the algorithms or parameters.

The phasing results will be used for clinical or regulatory decisions, and the accuracy of the phasing has not been validated against a ground truth.

The phasing results will be published, and the reproducibility of the workflow has not been established.

## Decision Framework for Tool Selection Based on Data Characteristics

### A Structured Approach to Choosing a Phasing Tool

The choice between WhatsHap, HapCUT2, and MarginPhase should follow a structured decision process based on measurable data characteristics instead of habit or convenience. This framework translates the benchmark observations into concrete selection criteria that researchers can apply before committing computational resources.

The first decision point is ploidy. For diploid samples, all three tools are viable candidates. For polyploid samples, the selection narrows to WhatsHap polyphase or HapCUT2, with the understanding that specialized methods such as nTChap may produce better results for high-ploidy genomes. The nTChap method demonstrates that polyploid phasing remains an active area of method development, and researchers should not assume that diploid-optimized tools will perform adequately on polyploid data.

The second decision point is the error profile of the sequencing platform. PacBio HiFi data has a lower error rate than Oxford Nanopore data, which means that simpler models may suffice. For PacBio HiFi data, HapCUT2 offers a good balance of speed and accuracy. For Oxford Nanopore data, MarginPhase's probabilistic model that incorporates alignment uncertainty may provide better accuracy, particularly when the read error rate is high or when alignment quality is variable.

The third decision point is coverage. At coverage below 15x, all tools will produce fragmented phasing, and the choice of tool matters less than the quality of the input data. At coverage between 15x and 30x, WhatsHap and HapCUT2 produce reliable results for most applications. At coverage above 30x, the computational cost of WhatsHap and MarginPhase increases substantially, and HapCUT2 may be the only practical choice for large genomes.

The fourth decision point is the downstream application. For allele-specific expression analysis, switch accuracy is critical because a single switch error can misassign multiple downstream variants. For structural variant characterization, the contiguity of phased blocks matters more than the switch error rate, because the goal is to determine which variants travel together on the same chromosome copy. The Duet study on SV phasing using Nanopore data found that dedicated SV phasing tools can outperform general-purpose phasing approaches at low coverage, which suggests that researchers with SV-focused questions should evaluate specialized tools.

### A Scoring Matrix for Tool Selection

The following scoring matrix provides a practical method for comparing the three tools against specific data characteristics. Researchers can assign scores from 1 to 5 for each criterion based on their specific dataset and application, then sum the scores to identify the most suitable tool.

| Criterion | WhatsHap | HapCUT2 | MarginPhase |
|-----------|----------|---------|-------------|
| Diploid accuracy at 20x coverage | 5 | 4 | 4 |
| Polyploid support | 4 | 4 | 1 |
| Speed on high-coverage data | 2 | 5 | 1 |
| Robustness to Nanopore errors | 3 | 3 | 5 |
| Ease of installation | 5 | 5 | 2 |
| Community support and documentation | 5 | 4 | 2 |
| Integration with existing pipelines | 5 | 4 | 2 |

This matrix reflects the general characteristics of each tool as observed in published evaluations and community usage. Researchers should adjust the scores based on their specific data and requirements. For example, a researcher working with PacBio HiFi data at 30x coverage who needs fast results may score HapCUT2 higher on accuracy than the matrix suggests, because the lower error rate of HiFi data reduces the advantage of MarginPhase's error modeling.

### Decision Rules for Common Scenarios

The following decision rules translate the scoring matrix into concrete recommendations for common research scenarios.

For a diploid human genome sequenced with PacBio HiFi at 30x coverage, HapCUT2 is the recommended choice. The low error rate of HiFi data means that the probabilistic modeling of MarginPhase provides little advantage, and the speed of HapCUT2 allows rapid iteration on parameter settings. WhatsHap is a reasonable alternative if the researcher prioritizes the optimal phasing guarantee, but the runtime may be substantially longer.

For a diploid plant genome sequenced with Oxford Nanopore at 20x coverage, MarginPhase is the recommended choice if the read error rate is high. The probabilistic model can account for alignment uncertainty and may reduce switch errors compared to tools that do not model this uncertainty. However, researchers should be prepared for longer runtimes and should evaluate whether the accuracy improvement justifies the computational cost.

For a polyploid crop genome sequenced with PacBio HiFi at 25x coverage, WhatsHap polyphase is the recommended choice among the three tools. HapCUT2 also supports polyploid phasing, but the dynamic programming approach in WhatsHap polyphase may produce more accurate results. Researchers should also evaluate specialized polyploid phasing methods such as nTChap, which has shown improved accuracy on simulated high-ploidy datasets.

For a low-coverage dataset below 10x, none of the three tools will produce reliable phasing. Researchers should either generate additional sequencing data or consider whether the phasing question can be answered with the available data. The Duet study found that at 8x coverage, dedicated SV phasing tools outperformed general-purpose phasing approaches, which suggests that specialized tools may be needed for low-coverage applications.

### Implementing the Decision Framework

The decision framework should be implemented as a documented process that includes the following steps.

First, characterize the input data by measuring coverage, read length N50, and the error profile of the sequencing platform. These measurements should be recorded before any phasing tool is run.

Second, apply the decision rules to select a primary tool and a secondary tool for comparison. Running two tools on the same dataset provides a cross-check on phasing quality and can reveal systematic errors in either tool.

Third, run the selected tools and record the metrics described in the records section, including the number of phased variants, the number of haplotype blocks, the N50 block length, and the runtime.

Fourth, compare the results across tools and assess whether the differences are consistent with the expected behavior of each tool. If the results are inconsistent in unexpected ways, investigate the input data for errors before accepting either result.

Fifth, document the decision process, including the data characteristics, the tools selected, the parameters used, and the rationale for the final choice. This documentation supports reproducibility and provides a basis for future decisions on similar datasets.

### When to Escalate to Specialized Tools

The three tools compared in this benchmark are general-purpose phasing tools that work well for many applications. However, there are situations where specialized tools may be more appropriate.

For structural variant phasing using Nanopore data, the Duet tool was specifically designed for SV calling and phasing and outperformed general-purpose tools at low coverage. Researchers with SV-focused questions should evaluate Duet alongside the general-purpose tools.

For polyploid haplotype reconstruction, the nTChap method uses an iterative procedure involving two-round clustering and consensus construction, and it outperformed current tools on simulated high-ploidy datasets. Researchers working with polyploid genomes should evaluate nTChap alongside WhatsHap polyphase and HapCUT2.

For cancer genomes with complex karyotype changes, the refLinker method uses Hi-C data to determine chromosomal haplotypes and is insensitive to large-scale DNA deletions, duplications, and high-level amplification. Researchers working with cancer genomes should consider whether Hi-C based phasing is needed to achieve chromosome-length continuity.

For highly heterozygous plant genomes, the PhaseGrass workflow combines reference-based phasing with haplotype-specific k-mers to partition reads to haplotypes. The method binned 20 percent more reads to haplotypes than WhatsHap and generated balanced haplomes compared to Hifiasm. Researchers working with highly heterozygous genomes should evaluate PhaseGrass for assembly-based phasing.

For long RNA-seq data, the longcallR tool performs joint single-nucleotide polymorphism calling, haplotype phasing, and allele-specific analysis. Researchers working with RNA-seq data should consider whether a dedicated RNA-seq phasing tool is needed, because the three tools compared in this benchmark are designed for DNA sequencing data.

### Recording the Decision Process

The decision process should be recorded in a structured format that includes the following elements.

The data characteristics measured before tool selection, including coverage, read length N50, and sequencing platform.

The criteria used to select the primary and secondary tools, including the scores from the scoring matrix and the decision rules applied.

The parameters used for each tool, including any filtering thresholds applied to the input VCF.

The metrics recorded for each tool run, including the number of phased variants, the number of haplotype blocks, the N50 block length, and the runtime.

The comparison of results across tools and the rationale for the final tool choice.

This record serves multiple purposes. It supports reproducibility by documenting the exact conditions of the phasing run. It provides a basis for troubleshooting if the phasing results are later found to be incorrect. It enables comparison across datasets and projects, which can reveal systematic patterns in tool performance. And it provides a reference for future decisions on similar data.

The [nf-core Documentation](https://nf-co.re/docs) describes community standards for reproducible bioinformatics workflows, including version pinning and parameter documentation. Following these standards for the decision process ensures that the phasing results are interpretable and reproducible.

### Common Mistakes in Tool Selection

The most common mistake in tool selection is choosing a tool based on familiarity instead of data characteristics. Researchers who have used WhatsHap for one project may default to WhatsHap for all projects, even when the data characteristics suggest that another tool would perform better. The decision framework described here provides a structured alternative to this habit-based selection.

Another common mistake is failing to evaluate the input data before selecting a tool. Researchers who run a phasing tool without measuring coverage, read length, and error profile may select a tool that is poorly suited to their data. The decision framework requires these measurements as the first step, which prevents this mistake.

A third common mistake is running only one tool and accepting the result without cross-validation. Running a second tool provides a check on phasing quality and can reveal systematic errors. The decision framework recommends running two tools as a standard practice.

A fourth common mistake is ignoring the computational cost of phasing. Researchers who select WhatsHap for a high-coverage dataset may find that the runtime is prohibitive, and they may need to switch to HapCUT2 after investing substantial compute time. The decision framework includes speed as a selection criterion, which helps researchers anticipate computational costs before running the tool.

A fifth common mistake is applying diploid tools to polyploid data without considering the implications. The decision framework addresses ploidy as the first decision point, which prevents this mistake.

### Professional Escalation Criteria for Tool Selection

Researchers should escalate the tool selection decision to a bioinformatics specialist or a colleague with phasing expertise when they encounter any of the following situations.

The data characteristics fall outside the range covered by the decision rules, such as coverage below 10x or ploidy above four.

The scoring matrix produces a tie between two tools, and the researcher cannot determine which tool is more appropriate for the specific application.

The researcher is uncertain about the error profile of their sequencing platform and cannot determine whether MarginPhase's error modeling would provide a meaningful advantage.

The phasing results will be used for clinical or regulatory decisions, and the tool selection must be justified with published evidence.

The phasing results will be published, and the tool selection must be defensible to reviewers who may question the choice of one tool over another.

In these situations, consulting a specialist can prevent costly mistakes and ensure that the tool selection is appropriate for the research question.

## Frequently Asked Questions

### What is the difference between phasing and assembly?

Phasing assigns genetic variants to parental chromosomes using reads that span multiple variants. Assembly reconstructs the full genome sequence from reads, which may include haplotype resolution as part of the assembly process. Read-based phasing tools such as WhatsHap, HapCUT2, and MarginPhase operate on aligned reads and a variant call set, while assembly-based methods such as PhaseGrass generate haplotype-resolved assemblies directly from reads. The choice between phasing and assembly depends on whether the research question requires the full haplotype sequence or only the assignment of variants to haplotypes.

### Which phasing tool is best for Oxford Nanopore data?

MarginPhase is designed for error-prone long reads and may produce more accurate phasing on Nanopore data than tools that do not model alignment uncertainty. However, the advantage of MarginPhase depends on the specific error profile of the data, and researchers should evaluate the tool on their own data before committing to a workflow. WhatsHap and HapCUT2 can also phase Nanopore data, and their accuracy may be sufficient for many applications. The Duet study found that at low coverage, dedicated SV phasing tools outperformed general-purpose phasing approaches, which suggests that the choice of tool depends on the specific application.

### How much coverage is needed for reliable phasing?

Reliable phasing typically requires at least 20x coverage, and higher coverage may be needed for polyploid samples or for genomes with low heterozygosity. At 8x coverage, phasing accuracy is reduced, and the Duet study found that SV phasing accuracy was lower at this coverage than at higher coverage. Researchers working with low-coverage data should expect reduced phasing performance and may need to consider coverage-specific tools or generate additional sequencing data.

### Can these tools phase polyploid genomes?

WhatsHap polyphase and HapCUT2 both support polyploid phasing, but their accuracy on polyploid data may be lower than on diploid data. The nTChap method for polyploid haplotype reconstruction outperforms current tools in terms of accuracy and error rates on simulated datasets, which suggests that specialized polyploid phasing methods may be needed for high-ploidy genomes. Researchers working with polyploid genomes should evaluate multiple tools and should consider whether a specialized method is needed for their application.

### What is the switch error rate and why does it matter?

Switch error rate measures the fraction of adjacent variants that are assigned to the wrong haplotype relative to a ground truth. A lower switch error rate indicates more accurate phasing. Switch errors can affect downstream analyses such as allele-specific expression and structural variant characterization, because these analyses depend on the correct assignment of variants to haplotypes. Researchers should calculate switch error rate whenever a ground truth is available and should report this metric in publications.

### How do I choose between WhatsHap and HapCUT2?

The choice between WhatsHap and HapCUT2 depends on the priority of accuracy versus speed. WhatsHap uses a dynamic programming algorithm that guarantees optimal phasing for the given objective function, but it can be slow on high-coverage datasets. HapCUT2 uses a max-cut heuristic that is faster but does not guarantee optimal phasing. Researchers who need guaranteed optimal phasing should consider WhatsHap, while those who prioritize speed and scalability may prefer HapCUT2.

### What should I do if my phasing results are fragmented?

Fragmented phasing results are often caused by low coverage or by variant calling errors. Researchers should first check the coverage of their dataset and the quality of their variant calls. If coverage is low, they may need to generate additional sequencing data or accept the reduced phasing contiguity. If variant calls contain errors, they should improve the variant calling workflow before re-running the phasing tool.

### Can I use these tools for RNA-seq data?

Long-read RNA-seq data can be used for phasing, and tools such as longcallR have been developed for joint single-nucleotide polymorphism calling, haplotype phasing, and allele-specific analysis with long RNA-seq reads. The three tools compared in this benchmark are designed for DNA sequencing data, and their performance on RNA-seq data may differ. Researchers working with RNA-seq data should consider whether a dedicated RNA-seq phasing tool is needed for their application.

## Related Bioinformatics Guides

- [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)
- [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)
- [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)
- [Long-Read Sequencing Cost and Market: What to Expect](/knowledge/bioinformatics/long-read-sequencing-cost-and-market-what-to-expect)

## 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.
- [Duet: SNP-assisted structural variant calling and phasing using Oxford nanopore sequencing.](https://pubmed.ncbi.nlm.nih.gov/36344913). BMC bioinformatics, 2022.
- [nTChap: an accurate method for polyploid haplotype reconstruction.](https://doi.org/10.1186/s12864-025-12440-w). 2026.
- [Haplotype-resolved reconstruction and functional interrogation of cancer karyotypes.](https://doi.org/10.1016/j.cels.2025.101477). 2026.
- [Chromosome-level haplotype-resolved assembly of highly heterozygous grass genomes with PhaseGrass.](https://doi.org/10.1038/s41467-025-66377-5). 2025.
- [SNP calling, haplotype phasing and allele-specific analysis with long RNA-seq reads.](https://doi.org/10.1038/s41592-026-03045-6). 2026.

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