Teaching AI to Identify Cancer Mutations with “Synthetic Tumors”: ClairS Fills a Gap in Long-Read Sequencing

A major challenge in cancer genome analysis is identifying somatic mutations that exist only in tumors and are absent from normal cells amid large amounts of sequencing noise. Long-read sequencing can span repetitive or difficult-to-map regions of the genome, but it also introduces error patterns different from those found in short-read data. ClairS, a deep learning tool developed by a University of Hong Kong team, seeks to fill this analytical gap.

ClairS identifies single-nucleotide variants (SNVs) and shorter insertions and deletions (indels) in paired tumor and normal samples. Rather than merely checking whether a particular base is abnormal, it also uses the haplotype information provided by long reads to determine whether a candidate mutation can reasonably be assigned to the paternal or maternal chromosome, helping exclude signals more likely to originate from sequencing errors.

The truly difficult issue is training data. Cancer samples that are both suitable for model training and labeled with high confidence are far less abundant than germline variant data. The team led by Professor Ruibang Luo therefore mixed real sequencing reads from two normal individuals, reassigned germline variants unique to one person as simulated tumor mutations, and adjusted the proportions of reads to create training scenarios with different tumor purities, sequencing depths, variant allele frequencies, and levels of normal-cell contamination.

The study in *Nature Methods* showed that, across six Oxford Nanopore datasets with 50-fold tumor coverage and 25-fold coverage for paired normal samples, models trained solely on synthetic data achieved SNV F1 scores ranging from 87.98% to 98.58%. F1 is a metric that combines precision and recall. However, the quality of the truth sets used for the different cell lines was inconsistent, and some benchmark answers had also incorporated interpretations from multiple algorithms, so these figures cannot be regarded directly as accuracy rates for clinical specimens.

In the triple-negative breast cancer cell line HCC1395 and its paired normal cell line HCC1395BL, which have more stringent reference standards, the synthetic-data model achieved F1 scores of 89.83% for SNVs and 73.38% for indels. After pretraining on synthetic samples and then adding data from real cancer cell lines, the two figures increased to 96.19% and 79.67%, respectively. This indicates that synthetic data can broaden the scope of training but still require calibration against real tumor characteristics; a gap remains particularly in identifying insertions and deletions.

The tool has been released as open source under the BSD 3-Clause License, and Oxford Nanopore’s paired tumor–normal somatic variant workflow has also adopted a modularized version of ClairS. The workflow currently operates only with specific R9 and R10 basecalling models, while short-indel analysis is further restricted to R10 models, reflecting that performance remains affected by sequencing chemistry versions and analytical settings.

There is still a clear boundary between the tool and clinical use: existing validation comes primarily from cancer cell lines rather than prospective patient specimens, and it has not yet been demonstrated that the results can improve diagnostic or treatment decisions. Both the research repository and Oxford Nanopore documentation state that the product is not intended for clinical diagnostic use; the paper also discloses that Ruibang Luo received research funding from the company. The next step will require independent validation across more cancer types, tumor purities, and real-world clinical workflows, as well as clarification of responsibilities related to version updates, quality control, and regulation.