August 20, 2026 | Analyzing genetic changes in cancer cells provides insight into how tumors develop and help with developing treatment through precision oncology. However, identifying cancer-related mutations can be challenging, particularly when they occur at low levels or in complex regions of the genome. While short-read sequencing is commonly used for analyzing DNA, it has limitations when examining complex or repetitive regions of a genome. Long-read sequencing can help, but there is still a higher error rate compared to short-read sequencing.
Researchers at The University of Hong Kong addressed this gap with ClairS, a tool designed to identify small cancer-related genetic mutations. The team trained the model using synthetic data to accurately detect cancer mutations from long-read sequencing data, an approach designed to overcome the limited availability of high-quality somatic mutation data. Led by Professor Ruibang Luo, assistant director (Learning Experience & Student Enrichment) and associate head of the department of AI and data science, School of Computing and Data Science, The University of Hong Kong, the study was published this summer in Nature Methods (DOI: 10.1038/s41592-026-03152-4).
The team already developed deep-learning germline variant callers, so the next step naturally turned toward somatic mutation detection. But high-quality somatic training data is greatly limited. ClairS uses a multistep process to identify and verify somatic mutations. First, it identifies germline variants in the sample and uses phasing to determine which DNA sequences came from the same chromosome. Next, ClairS uses two different neural networks: a pileup-based model, which looks at groups of sequencing reads at a specific location, and a full-alignment model, which examines the detailed alignment of individual reads. Finally, ClairS looks to ancestral haplotype support for more distant genetic variants on the same chromosomes.
What sets ClairS apart from other models is how the team trained it with synthetic tumor and normal samples using real sequencing data. By doing so, the team could treat variants found in one sample but not the other as simulated cancer mutations, thus creating training data that represents different levels of tumor DNA, sequencing coverage, and contamination from normal cells. According to Luo, this training framework has actually been implemented and adapted to multiple long-read platforms, including Oxford Nanopore and PacBio, across different sequencing chemistries and has achieved promising performance.
The team successfully showed that using carefully designed synthetic data, reliable training labels, and phasing can effectively train a deep-learning somatic caller and make reliable predictions of real somatic variants across different sequencing coverages and tumor purities, even without using any real tumor samples for model training. Luo said the team also verified that haplotype information from long reads can improve mutation detection, particularly for low-frequency variants that are more difficult to identify.
The team was also surprised to find that simply increasing tumor sequencing depth did not necessarily lead to increased accuracy. “In our experiments, performance continued to improve beyond around 50-fold tumor coverage, but the gains became relatively limited,” said Luo. This contrasted with how increasing coverage of the matched-normal sample was often more effective because it provided stronger evidence to exclude germline variants and reduce false-positive somatic calls.
The findings suggest that ClairS could eventually provide clinicians with a more comprehensive view of cancer genomes by detecting mutations in repetitive, low-mappability, and complex regions that short-read sequencing may miss, while also improving the detection of low-frequency variants through long-read haplotype information. These could provide useful information for molecular profiling and treatment decisions when working with patients.
The study had a few limitations. Most of the training and benchmarking were based on cancer cell lines, which are cleaner and less heterogeneous than real clinical tumor samples, acknowledged Luo. Even though ClairS was tested on real cancer samples, the scale is still limited. A larger-scale clinical validation will need to be included in next steps.
Furthermore, the tool does not fully solve the challenges of detecting very low-frequency mutations and variants in complex genomic regions, particularly indels. “These regions, sequencing errors, limited coverage, and alignment ambiguity can still make accurate detection difficult,” Luo explained.
The researchers suggest that future work focus on adapting the method for use with a wider range of clinical cancer samples — such as different tumor types, tumor purities, and sample preparation conditions — and determine whether the benefits of long-read sequencing justify the additional costs associated with deeper sequencing. “These studies would help us better understand where and when long-read sequencing provides the greatest benefit for cancer analysis in real clinical settings,” Luo elaborated. He also recommended that future researchers look at mutation burden across different cancer types, as different tumors may have different numbers and somatic mutation patterns.
The vision for ClairS is to develop it into a more general framework for detecting low-frequency mosaic mutations and apply it beyond cancer. ClairS could be used to study how mutations accumulate in healthy tissues as people age, trace how cells develop, and track how tumors change or develop treatment-resistant cells over time. The approach could also be applied to microbial genomics, where detecting low-frequency genetic variants could help researchers distinguish between different strains in a mixed population and track how microbes evolve.
“This could be useful for studying microbial diversity, adaptation, and drug resistance,” said Luo.