Clair-Mosaic: A deep-learning method for long-read mosaic small variant calling

Mosaic variants, defined as postzygotic mutations occurring during an organism’s development from zygote to adult, play critical roles in developmental biology, aging, and diseases such as cancer and neurological disorders. However, their accurate detection remains challenging due to low abundance in the genome and low variant allelic fractions (VAF). While current mosaic variant callers are primarily designed for short-read sequencing, no method is available for long-read sequencing, which can generate tens of kilobase-long reads to cover complex genomic regions inaccessible to short reads. To fill the gap, we present Clair-Mosaic, a deep-learning-based method for detecting mosaic small variants from long-read data. Clair-Mosaic was trained on hundreds of millions synthetic variants encompassing diverse read coverages and allelic fractions, enabling it to detect low-VAF mosaic variants with high sensitivity in paired-sample and single-sample modes. In addition to neural network prediction, Clair-Mosaic distinguishes genuine mosaic variants from sequencing artifacts by leveraging their inherent haplotype relationship in phased long reads. Furthermore, a Bayesian mosaic-germline discriminator is introduced to distinguish mosaic variants from germline variants. It also employs multiple post-calling filters, including a mosaic variant database and multiple germline population resources, to tag common germline and mosaic variants. Comprehensive benchmarking on synthetic datasets and real samples demonstrated Clair-Mosaic’s outstanding performance in ONT and PacBio. Clair-Mosaic is also applicable to short-read data and outperforms methods like MosaicHunter, MosaicForecast, DeepMosaic, and DeepSomatic. Clair-Mosaic is open-source and available at https://github.com/HKU-BAL/Clair-Mosaic.