ClairS-TO: a deep-learning method for long-read tumor-only somatic small variant calling

Accurate detection of somatic variants in tumors is of critical importance and remains challenging. Current methods typically require matched normal samples for reliable detection, which are often unavailable in real-world research and clinical scenarios. Without a matched normal sample, more proficient algorithms are required to distinguish true somatic variants from germline variants and technical artifacts. However, existing tumor-only somatic variant callers that were designed for short-read sequencing data are not able to work well with long-read data. To fill the gap, we present ClairS-TO, a deep-learning-based method for long-read tumor-only somatic variant calling. ClairS-TO uses an ensemble of two disparate neural networks trained from the same samples but for opposite tasks—how likely/not likely a candidate is a somatic variant. Benchmarks using COLO829 and HCC1395 cancer cell lines show that ClairS-TO outperforms DeepSomatic and smrest in ONT and PacBio long-read data. ClairS-TO is also applicable to short-read data and outperforms Mutect2, Octopus, Pisces, and DeepSomatic. Extensive experiments across various sequencing coverages, variant allelic fractions, and tumor purities support that ClairS-TO is a reliable tool for somatic variant discovery. ClairS-TO is open-source, available at https://github.com/HKU-BAL/ClairS-TO . The accurate detection of somatic variants in cancer without matched normal controls remains challenging, particularly for long-read sequencing data. Here, the authors develop ClairS-TO, a deep learning method for long-read tumour-only somatic variant calling that outperforms similar algorithms and can also work with short-read sequencing data.