Haplotype information is crucial for biomedical and population genetics research. However, current strategies to produce de novo haplotype-resolved assemblies often require either difficult-to-acquire parental data or an intermediate haplotype-collapsed assembly. Here, we present Graphasing, a workflow which synthesizes the global phase signal of Strand-seq with assembly graph topology to produce chromosome-scale de novo haplotypes for diploid genomes. Graphasing readily integrates with any assembly workflow that both outputs an assembly graph and has a haplotype assembly mode. Graphasing performs comparably to trio phasing in contiguity, phasing accuracy, and assembly quality, outperforms Hi-C in phasing accuracy, and generates human assemblies with over 18 chromosome-spanning haplotypes.
Graphasing performs comparably to trio phasing in contiguity, phasing accuracy, and assembly quality
multi-metric agreement across two genomes and two assemblers, with limited breadth for generalization
Graphasing outperforms Hi-C in phasing accuracy
clear for Verkko but not for hifiasm, whose phasing error rates were similar
generates human assemblies with over 18 chromosome-spanning haplotypes
directly reflected in the reported chromosome-scale and telomere-based counts
even a set of 96 low-quality libraries can still produce contiguous assembly with greater than 98% concordance with the HG002 v1.0 assembly for input unitigs longer than 50 kbp
directly tested in the NA24385 library-quality titration with the stated unitig threshold
Derived from the full evaluation — not a separate score.
Strengths
Graphasing is evaluated with two assemblers and two human genomes against native trio and Hi-C modes using contiguity, phasing-error, consensus, structural, and gene-completeness measures. The paper reports outcomes where Graphasing does not lead, including trio's advantage in gene completeness.
↳ Results, Tables 1–2 and Figures 2–7
The Strand-seq titration varies the proportion of high-quality libraries and reports greater than 98% whole-assembly agreement across the tested mixtures. It also separately examines the more difficult acrocentric chromosomes.
↳ Results, Strand-seq Library Titration, Figure 8
The Discussion explains how diploid-only operation, fragmented graphs, difficult Strand-seq production, and propagation of input switch errors constrain use. These boundaries are connected to the intended setting of reference-quality diploid assembly when trio data is unavailable.
↳ Discussion, final two paragraphs
Limitations
The five stages shown in Figure 1 are evaluated as an integrated workflow, so the individual value of clustering, orientation correction, pooling, and graph threading is not empirically isolated.
↳ Results, Graph-Phasing Method, Figure 1; Methods
The principal comparison uses NA24385 and HG00733, both well-characterized human reference samples. This limits evidence for performance across broader populations, non-human organisms, fragmented inputs, and other genome architectures.
↳ Results, Phasing Method Comparison; Discussion
The abstract states an overall phasing-accuracy advantage over Hi-C, but hifiasm error rates are similar across methods and the strongest difference appears in the Verkko comparisons.
↳ Abstract; Results, Phasing Accuracy, Figures 3–4
The contribution score reflects a clear extension of graph-based phasing that is directly compared with trio and Hi-C modes in two current assemblers. Methodological Rigour is moderated by the absence of component-level ablations and by end-to-end comparisons limited to two human genomes, despite the broad metric suite and input-quality titration. Reporting is reduced because the abstract generalizes superiority over Hi-C even though the phasing-error advantage is strongest for Verkko and not evident across hifiasm. Positioning remains a relative strength because the Discussion ties technical limitations to specific use conditions and future development needs.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Sound3.7
Confidence highThe score reflects a distinct single-sample graph-phasing workflow that combines Strand-seq global phase information with assembly-graph topology and is compared against trio and Hi-C approaches. Its demonstrated advance remains bounded by two human genomes and two assembler workflows.
Graphasing performs comparably to trio phasing in contiguity, phasing accuracy, and assembly quality
The evaluation combines two assemblers, two genomes, relevant trio and Hi-C baselines, multiple quality measures, a library-quality titration, and resource profiling. The five-stage workflow is evaluated primarily as a whole, leaving the contribution of its individual components unisolated.
The Graphasing workflow can be summarized in five main steps
The paper follows a coherent progression from method description to comparative evaluation and limitations. The abstract's unqualified claim of better Hi-C phasing accuracy is broader than the configuration-dependent evidence, because hifiasm error rates are similar across methods while the clearest difference occurs with Verkko.
outperforms Hi-C in phasing accuracy
The paper distinguishes reference-based, trio, Hi-C, Strand-seq, and graph-based phasing approaches and identifies their practical trade-offs. Its limitations are connected to use conditions, including diploid-only scope, fragmented graphs, Strand-seq production, and propagation of input switch errors.
any switch errors present in the input assembly will propagate to the final haplotypes
Lower confidence on Methodological Rigour, Reporting — domain match limited.
No concerns4 of 4 checks
The numerical claims in the Abstract, Results, tables, and figures are consistent, and the methods cover the reported evaluation domains. The paper also identifies reference-dependent limitations affecting some comparative estimates.
The workflow is identified as publicly available under an MIT licence, and the study uses existing reference-sample data. No ethics approval or preregistration was required for the described computational analysis.
Flags: 2 declared / 5 total
65 of 65 checkable references verified
86 references in manuscript 21 are books, websites or datasets — counted, but not index-checkable
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The open workflow runs with Verkko and hifiasm and includes reported compute requirements, placing it beyond a conceptual demonstration. Routine deployment is tempered by specialized Strand-seq preparation and validation on only two reference genomes.
Graphasing is publicly available under an MIT license
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Antipov, Rautiainen, Nurk, et al. · 2025 · Computer and information sciences
Together, these changes increase the number of telomere-to-telomere scaffolds by twofold, reduce runtime by fourfold, and improve assembly correctness.
Well-supportedKobashi, Srou, Tsubokura, et al. · 2022 · Public health and health services
Pesticide-protective behaviors significantly correlated with fewer symptoms.
TentativeL. Leuzzi, T. Rizzo · 2024 · Physical Sciences