The GPhase workflow proceeds through three steps. First, the unitig graph is partitioned into chromosome-scale subgraphs by cutting specific edges, with subgraphs grouped into clusters on the basis of k-mer similarity and then assigned to chromosome bins using global 3C interaction signals. Second, within each chromosome bin, unitigs are separated into haplotypes. Collapsed unitigs—regions where homologous sequences have been merged into a single node—are used to guide phasing: the algorithm first clusters the non-collapsed unitigs directly connected to each collapsed unitig, then validates the clustering by examining the degree of sequence homology within each cluster, and finally assigns the collapsed unitigs themselves according to the strength of 3C signals and sequence homology to the non-collapsed unitigs in each haplotype. Third, within each haplotype, unitigs are linearized using local graph topology, and disjointed sequences are ordered and oriented with global 3C linkage information to yield final chromosome-scale scaffolds.

Benchmarking on simulated polyploid genomes with varying degrees of sequence similarity, GPhase consistently outperformed existing tools, reducing the sequence missing rate by up to 22.5-fold and the sequence error rate by up to 12.5-fold. On real genomes—including potato, rose, and kiwifruit—GPhase produced chromosome-level assemblies with fewer misassemblies and higher contiguity than competing approaches. The algorithm also recovered haplotype-resolved sequences that existing pipelines missed entirely, demonstrating its capacity to resolve the complex homologous relationships that characterize polyploid genomes.

The study, published in Nature Plants, provides both the algorithm and a practical implementation, offering a broadly applicable solution for polyploid genome assembly. By unifying assembly graphs and 3C signals in a single framework, GPhase addresses a long-standing gap in genomics and opens the door to more accurate and complete assemblies of the many crop species whose genomes are polyploid.

The scale of the polyploid assembly problem is easy to underestimate. Roughly 70% of flowering plants carry evidence of polyploid ancestry in their evolutionary history, yet polyploid species make up only about 18% of the high-quality genome assemblies currently available. That gap is not a matter of neglect. It reflects a technical bottleneck: when three, four, or more copies of every chromosome coexist in a single nucleus and share stretches of highly similar sequence, the computational machinery that cleanly separates maternal from paternal chromosomes in diploids struggles to tell homologs apart. Larger genome sizes and higher proportions of repetitive elements compound the difficulty, which is why many economically and scientifically important polyploid crops still lack chromosome-scale, haplotype-resolved reference genomes.

The two-step structure of modern polyploid assembly pipelines helps explain where things go wrong. The first step, in which long reads are assembled into unitigs, is essentially a solved problem: general-purpose assemblers such as hifiasm, Verkko, Flye, and NextDenovo handle it well, and they routinely output assembly graphs as a byproduct. The second step—assigning unitigs to haplotypes and arranging them into chromosome-scale scaffolds—is where polyploid projects stall. This step depends on independent linkage information, and the dominant source has been chromatin conformation capture data, prized for its favorable balance of cost, scalability, and effectiveness. Hi-C remains the most widely used flavor, with newer approaches such as Pore-C and Omni-C gaining adoption because their longer reads or freedom from restriction enzyme cut sites offer theoretical advantages.

But 3C data share a fundamental weakness regardless of flavor: they provide only coarse-grained linkage. Two failure modes follow. Short unitigs that happen to lack restriction cut sites, or that fall in regions of uneven sequencing coverage, produce sparse or absent interaction signals and are often discarded outright, leaving chromosomes incomplete. Meanwhile, homologous regions with high sequence similarity are frequently collapsed into consensus sequences during unitig assembly, and these collapsed unitigs emit strong interaction signals toward multiple haplotypes at once, corrupting the clustering step and producing uneven, mis-phased assemblies. Existing polyploid scaffolding tools, ALLHiC and HapHiC, were designed around Hi-C data and have powered most recent polyploid assemblies, but they cannot escape these data-inherent limitations.

The central insight behind GPhase is that the assembly graph, which is already generated by mainstream unitig assemblers, contains exactly the complementary information that 3C data lack. Because graph edges record overlap relationships between long reads, they carry fine-grained connectivity at the sequence level. A short unitig with no usable Hi-C signal can still be placed correctly through its graph connections to neighbors. A collapsed unitig betrays itself in the graph by forming links to multiple haplotypes simultaneously, allowing the algorithm to detect and handle it rather than letting it mislead the phasing process. The graph topology even hints at the relative ordering of unitigs along a chromosome, and it enables something no existing polyploid scaffolding tool offers: the direct assembly of unitigs into longer contigs, which boosts the contiguity of gap-free sequence.

The benchmarking design deserves attention because it separates what an algorithm claims from what it can prove. The evaluation set included two semi-simulated polyploid genomes with high-quality ground-truth assemblies, constructed by merging four and three homologous haploid rice genomes respectively—a clever trick that provides perfect known answers against which phasing accuracy can be measured. Four purely in-silico-simulated autopolyploid genomes with varying degrees of sequence similarity probed how performance degrades as homologs become harder to distinguish. Five real polyploid genomes, including potato, wax apple, and rose, tested performance under messy real-world conditions, and an additional four polyploid genomes served as an independent validation set. Across Hi-C, Pore-C, and Omni-C data types, GPhase reduced the BUSCO-based sequence missing rate by up to tens of times, cut the k-mer-based sequence error rate by up to an order of magnitude, and improved the N50 contiguity of gap-free sequences by several-fold relative to existing tools.

These metrics map onto distinct biological consequences. A reduced missing rate means fewer genes and regulatory elements are absent from the final assembly, which matters enormously for gene annotation, comparative genomics, and any analysis that depends on knowing what is present. A lower error rate means the assembled sequence more faithfully represents the actual chromosomes, reducing false variants and spurious structural differences that would otherwise propagate into downstream studies. Improved contiguity of gap-free sequence means fewer unresolvable breaks, which preserves synteny information and makes it easier to trace how genes and genome regions have been rearranged, duplicated, and retained or lost across evolutionary history.

The distinction between autopolyploids and allopolyploids, though not the focus of the algorithm itself, underlies why sequence similarity among homologs is such a decisive variable. Autopolyploids arise when genome copies derive from the same species, so their homologous chromosomes can be nearly identical at the sequence level, making discrimination hardest. Allopolyploids combine subgenomes from different ancestral species, giving each subgenome a distinct sequence identity that partially eases phasing. By testing on simulated genomes with graded levels of similarity, the benchmark captured this spectrum directly, and the consistent advantage of GPhase across that spectrum suggests the assembly-graph signal remains informative even when 3C signals alone would be hopelessly confounded.

It is also notable that diploid-aware tools that already exploit assembly graphs, such as hifiasm, do not simply extend to polyploid genomes. This indicates that the polyploid problem is not merely a diploid problem with more copies; the combinatorics of assigning unitigs among three or more haplotypes, the prevalence of collapsed regions, and the ambiguity of multi-haplotype 3C signals create qualitatively new challenges that require purpose-built algorithmic treatment. GPhase’s preprocessing normalizes and unifies 3C input, its partitioning step cuts the unitig graph into chromosome-scale subgraphs, its phasing step uses collapsed unitigs as anchors to guide local clustering with validation by homology degree, and its scaffolding step linearizes unitigs within each haplotype using local graph topology before applying global 3C linkage to order and orient disjointed sequences.

For the plant genomics community, the practical implications are broad. Many staple and specialty crops—potato among the most prominent—are polyploid or have polyploid ancestors, and their improvement depends on understanding trait architecture, subgenome dominance, and heterosis, all of which require knowing which alleles sit on which homolog. Haplotype-resolved, chromosome-scale assemblies are the substrate for that knowledge. By making such assemblies more accurate, more complete, and more contiguous, and by doing so with data types that laboratories already generate, an approach like this lowers the barrier for species that have waited longest for reference-quality genomes. The open availability of both the algorithm and its implementation further suggests that the gap between the polyploid share of flowering plants and the polyploid share of high-quality assemblies may finally begin to close.

Subject of Research: Genomics

Article Title: Assembly-graph-assisted haplotype-resolved scaffolding of polyploid genomes

Article References: Du, W., Li, Y., Zeng, X., Yang, J., Wang, X., Zhang, X., Shang, L., Fu, X., Zhang, X., Wei, X.-Q., Cheng, T., Li, D., Wu, Z., & Pan, W. (2026). Assembly-graph-assisted haplotype-resolved scaffolding of polyploid genomes. Nature Plants. https://doi.org/10.1038/s41477-026-02422-5

Image Credits: AI Generated

DOI: 10.1038/s41477-026-02422-5

Keywords: GPhase, polyploid genomes, assembly graphs, 3C signals, haplotype-resolved assembly, chromosome-level scaffolding, Nature Plants, Assembly-graph-assisted, haplotype-resolved, scaffolding, polyploid, genomes