Detecting the mutations that distinguish a tumour from its host tissue is a core computational task in cancer genomics, and most established methods were designed for short-read sequencing. They struggle in structurally complex regions of the genome, precisely where cancer rearrangements often occur. A team at the School of Computing and Data Science at HKU has built ClairS, a deep learning method for calling somatic variants from long-read tumour and normal pairs.
The more interesting contribution is how the model is trained. Real labelled cancer sequencing data are scarce and expensive, so the team generates training data by mixing sequencing data from normal human samples to create synthetic tumour and normal pairs spanning varied tumour purities, sequencing depths and mutation levels. This sidesteps the data bottleneck that constrains most clinical machine learning, and does so without requiring access to patient cohorts.
ClairS was evaluated on breast, lung and melanoma cell-line datasets, with pancreatic cancer also examined, and the work was published in Nature Methods. It has since been integrated into Oxford Nanopore Technologies' official somatic variant-calling workflow and released as open source, which is a more meaningful indicator of practical utility than benchmark performance alone, since a commercial sequencing platform adopting an academic caller implies it survived contact with real production data.
The synthetic training strategy is likely to travel further than the specific tool. Medical machine learning is routinely limited by the scarcity of annotated clinical data, and constructing realistic training distributions from abundant non-clinical material is a general answer to that problem. Whether it generalises beyond sequencing, where the noise model is unusually well understood, remains an open and worthwhile question.