MNVista: accurate identification of multi-nucleotide variants from next generation sequencing data

Detecting minimal residual disease (MRD) from cell-free DNA (cfDNA) is challenging because circulating tumor DNA (ctDNA) is scarce and sequencing error rates generate substantial background. In germinal center B-cell derived lymphomas, the somatic hypermutation machinery produces multi-nucleotide variants (MNVs), particularly in super-enhancer regions. As the chance that these patterns arise from sequencing errors is extremely low, MNV analysis strongly increases analytical sensitivity of detecting ctDNA. We developed an open-source tool, called MNVista, that calls MNVs by identifying co-occurring single-nucleotide variants (SNVs) and scoring candidate pairs with a Bayesian model. MNVista accepts standard VCF and indexed BAM file inputs and easily integrates into pipelines with or without matched controls. MNVista detected more patient-specific MNVs even at low VAFs down to a VAF of 0.0025% (25 ppm) in publicly available diluted patient data as compared to Mutect2 (1,783 ppm). MNVista recovered complex MNVs composed of up to nine co-occurring SNVs, implying a theoretical detection limit below the 1ppm threshold cited as a target for clinical MRD detection by the National Comprehensive Cancer Network. The effectiveness of MNVista was additionally supported by synthetic benchmark studies in which we recovered spiked-in MNVs in real sequencing data and compared performance to Mutect2. In contrast to Mutect2, the performance of MNVista remained robust when background SNVs were synthetically introduced in real patient data including somatic variants and sequencing artefacts. MNVista is an open-source MNV caller that can handle standard VCF + BAM inputs using a Bayesian co-occurrence model with read-level metrics. This allows integration into existing pipelines. In publicly available dilution series, MNVista detected patient MNVs down to 25 ppm and synthetic benchmarks showed near-perfect noise-free performance and robust accuracy in patient-derived sequence backgrounds. Together, these findings indicate that the MNVista tool can be used for ultra-sensitive MNV-based ctDNA detection of B-cell lymphoma.