Summary
A bioRxiv preprint presents LongPhase-TO, a long-read method for reconstructing somatic haplotypes from tumour samples without matched normal DNA. In tests on eight datasets from six cancer cell lines, the method lengthened haplotype blocks and improved somatic variant-calling scores.
A preprint posted on September 19, 2026, describes LongPhase-TO, a computational method that reconstructs somatic haplotypes from long-read sequencing of a tumour sample alone. The authors report that it improved haplotype-block length and somatic variant calling across eight datasets from six cancer cell lines, with the largest gains at low tumour-DNA fractions.
Contents
- Why tumour-only sequencing is difficult
- What LongPhase-TO changes
- Results across cancer cell lines
- How to interpret the preprint's evidence
Why tumour-only sequencing is difficult
Cancer genome analysis needs to distinguish somatic variants, which arise in tumour cells, from germline variants, which are inherited and can also be present in the tumour. It also needs to determine which variants lie on the same physical copy of a chromosome. That arrangement is called a haplotype.
Long sequencing reads can span multiple variants, providing the physical linkage needed to reconstruct haplotypes. Tumour samples, however, can contain a mixture of cancer cells and normal cells. Cancer genomes can also lose one copy of a chromosome region, a change known as loss of heterozygosity, or LOH. These effects can make it difficult to decide whether a sequence pattern reflects a somatic change, an inherited variant or the proportions of different cell populations.
The preprint identifies separation of somatic and germline variants and reconstruction of somatic haplotypes as the two central problems in tumour-only analysis.
What LongPhase-TO changes
The method takes a different approach from workflows that first construct germline haplotypes and then place somatic variants onto them. LongPhase-TO instead puts germline and somatic alleles into a single graph and phases them together.
According to the authors, the method estimates LOH and the fraction of tumour DNA internally. It uses depletion of heterozygosity and imbalance between haplotypes rather than relying on a separate copy-number and ploidy model. In this context, phasing means assigning variants to the chromosome copies on which they occur; a longer phased block connects more variants across a continuous stretch of DNA.
The method is designed to work with existing somatic variant callers, including ClairS-TO and DeepSomatic-TO. Its role is to improve the haplotype information available to those callers and to recalibrate their variant calls.
Results across cancer cell lines
The authors evaluated LongPhase-TO on eight datasets from six cancer cell lines covering breast, melanoma and lung cancer. They report that the median haplotype-block N50 increased 2.9-fold compared with germline phasing methods. N50 is a length statistic for assembled or phased blocks: a higher value generally indicates that more sequence has been connected into longer blocks.
Somatic variant-calling performance also improved. For single-nucleotide variants, the reported mean F1 scores rose from 0.55 to 0.62 and 0.65. For insertions and deletions, or indels, mean F1 increased from 0.19 to 0.23. F1 combines precision and recall into a single score, so an increase can indicate a better balance between identifying true variants and limiting incorrect calls.
The largest improvements occurred when the tumour DNA fraction was low. That is an important operating condition because a smaller proportion of tumour DNA leaves fewer tumour-specific molecules for distinguishing cancer variants from background sequence.
How to interpret the preprint's evidence
The findings support LongPhase-TO as a method for improving tumour-only long-read analysis in the tested cell-line datasets. Longer somatic haplotypes can connect variants across larger genomic regions, while improved variant calls may make it easier to study how cancer mutations are arranged on chromosome copies rather than treating each mutation as an isolated event.
The evidence is from a bioRxiv preprint and its evaluation covers eight datasets from six cancer cell lines. Patient-sample performance and clinical applications therefore remain subsequent validation steps rather than findings established by this study. The preprint also states that Ruibang Luo receives research funding from Oxford Nanopore Technologies; the other authors declare no competing interests.