Abstract 1900: Rapid variant detection and annotations from next generation sequencing data using a GPU accelerated framework

Abstract Introduction Next Generation Sequencing (NGS) played a crucial role in revolutionizing the field of human genetics to become an integral part of research and clinical management. Some examples are diagnosis and treatment of cancer patients, identifying syndromes in the NICU (Neonatal Intensive Care Units), insights into neuro-degenerative disease to name a few. NGS integrated with multi-omics data provides greater insight to underlying molecular mechanisms and helps to identify alterations specific to the disease condition. Accurate and robust detection of somatic and germline genomic alterations from whole genome sequencing (WGS) data is complex where the datasets tend to be very large, and they require highly compute intensive operations that often take days to weeks to complete. By leveraging the Graphical Processing Unit (GPU) accelerated framework, it is possible to analyze the WGS data and identify clinically actionable variants with fast turnaround times without compromising on the accuracy. In this study, we implemented a GPU accelerated framework for most used variant callers ie. Mutect2, MuSE, Somatic Sniper and LoFreq with significantly faster run times in comparison to their Central Processing Unit (CPU) counterpart. Methods Commonly used CPU based somatic variant calling algorithms Mutect2, MuSE, Somatic Sniper and LoFreq, were selected and subsequently accelerated to run on GPUs. The goal of this work is to generate results that are equivalent to the CPU version of these algorithms, but at much faster run times. As a control data set, we utilize the SEQC2 consortium WGS dataset consisting of ~50x tumor and normal samples to assess the performance of individual CPU variant callers to their GPU equivalents. Results The GPU accelerated framework overall provides 10-66x faster run times compared to their CPU-based counterparts. Individually, CPU vs GPU comparisons are, Mutect2 achieved a maximum acceleration of 66x, followed by MuSE, LoFreq both 11x and Somatic Sniper 10x acceleration respectively. The results from accelerated MuSE, LoFreq, Somatic sniper are identical in comparison to the CPU-based somatic variant callers, For Mutect2 we achieve 99.9% sensitivity and 99.85% precision. Discussion Recently, machine-learning based variant detection algorithms exhibited better performance compared to the statistics based heuristic approaches. Here, we have presented the results for an ensemble of four variant callers to be used as the starting point for an ongoing effort to develop a deep learning-based approach for somatic variant detection and smart filtering algorithms. The goal is to enable improved variant calling and provide a GPU computing framework for faster, better, and reproducible workflows to analyze large scale NGS datasets. Citation Format: Pankaj Vats, Ankit Sethia, Mehrzad Samadi, Timothy T. Harkins. Rapid variant detection and annotations from next generation sequencing data using a GPU accelerated framework [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1900.