Separating somatic from germline variants and reconstructing somatic haplotypes are central challenges in tumor-only cancer genome analysis. Long-read sequencing provides the linkage information required to phase variants across large genomic spans, but two complications make tumor-only analysis difficult: chromosome-scale loss of heterozygosity (LOH) and an unknown degree of normal-cell admixture (tumor DNA fraction). Both phenomena blur distinctions between somatic and germline haplotypes and complicate downstream variant calling and phasing.
Standard approaches often rely on a matched normal sample or apply somatic calls onto precomputed germline haplotypes. These strategies can fail when LOH spans megabase regions or when tumor purity is low, because phasing and somatic/germline discrimination require joint modeling of allele linkage and allele balance across reads.
The authors present LongPhase-TO, a method designed to reconstruct somatic haplotypes from a tumor sample alone. Instead of first phasing germline haplotypes and then mapping somatic variants onto them, LongPhase-TO co-phases germline and somatic alleles together in a single unified graph representation. This co-phasing uses the long-range linkage carried by long reads to connect variants across large genomic distances and to assign alleles to haplotypes in a manner that incorporates somatic variation directly.
By constructing and solving this unified phasing graph, LongPhase-TO produces haplotype blocks that span larger genomic intervals than germline-only phasing tools applied to the same tumor-only data.
LongPhase-TO does not rely on an external copy-number or ploidy model to address LOH or tumor purity. Instead, the method infers LOH and the tumor DNA fraction internally from two signal patterns in the data: depletion of heterozygosity and imbalance between haplotype-specific read counts. These internal estimates allow the algorithm to distinguish regions where one haplotype is underrepresented due to LOH or admixture and to adjust phasing and variant recalibration accordingly.
This internal handling is especially relevant for tumor-only workflows because it avoids dependence on potentially noisy copy-number calls or separate purity estimates.
The authors evaluated LongPhase-TO across eight datasets derived from six cancer cell lines. When compared to germline phasers applied to the same tumor-only long-read data, LongPhase-TO increased haplotype block N50 by a median of 2.9-fold. This indicates substantially larger contiguous haplotype reconstructions, consistent with the method’s goal of leveraging long-read linkage and co-phasing of somatic and germline alleles.
The reported improvements in haplotype block N50 suggest LongPhase-TO can reconstruct megabase-scale somatic haplotypes in cell-line-derived tumor samples.
LongPhase-TO was tested in combination with existing somatic callers for long reads. The study reports consistent improvements in somatic single-nucleotide variant (SNV) and indel calls when using LongPhase-TO as a preprocessing or phasing-aware recalibration step.
Specifically, mean F1 for somatic SNVs increased from 0.55 to 0.62 when paired with ClairS-TO and to 0.65 when paired with DeepSomatic-TO. For somatic indels, mean F1 increased from 0.19 to 0.23. These gains indicate that improved haplotype resolution and internal handling of LOH and tumor fraction translate into better discrimination of somatic variants in tumor-only long-read data.
The largest gains from LongPhase-TO were observed at low tumor DNA fraction, where the admixture of normal DNA and chromosomal LOH make somatic/germline separation most difficult. By resolving heterozygosity depletion and haplotype imbalance internally, the method enhances both phasing continuity and somatic variant accuracy in these challenging contexts.
Across breast, melanoma and lung cancer cell lines included in the evaluation, LongPhase-TO improved the accuracy of existing somatic callers and reconstructed megabase-scale somatic haplotypes, demonstrating applicability across multiple cancer types modeled by cell lines.
The evaluation used eight datasets from six cancer cell lines, including models from breast, melanoma and lung cancers. The preprint presents aggregated performance metrics (haplotype block N50 improvements and mean F1 changes for SNVs and indels) as described above. Detailed per-dataset values, implementation parameters, and additional validation details are reported in the preprint but are not exhaustively repeated here.
One author (R.L.) receives research funding from Oxford Nanopore Technologies. The other authors declared no competing interests. The work was posted as a bioRxiv preprint; users should consult the preprint for full methodological detail, supplementary material, and links to code or data if provided by the authors.