利用NVIDIA Parabricks通过GPU加速基因组分析来加速微小残留病（MRD）检测

Previously we demonstrated that ctDNA present in lymphatic exudate collected via surgical drains (“lymph”) outperformed plasma for detecting MRD in head and neck squamous cell carcinoma (HNSCC) patients through a targeted sequencing (TS) approach 1 . However, detecting ultra low frequency variants requires deep sequencing coverage and computationally intensive workflows, often resulting in long turnaround times. To enhance the performance of the TS approach, we implemented an MRD detection pipeline optimized with NVIDIA Parabricks 2 to accelerate computation and enable faster, scalable molecular analysis without compromising accuracy. 25 unique patients with HPV-independent HNSCC were included in the cohort to demonstrate the performance improvements by porting our MRD workflow from a CPU-based infrastructure to GPU-accelerated Parabricks. Clinical validity was assessed by testing those same 25 patients who each had a minimum of 1 year of clinical follow up data and using the determined mean variant allele fraction (VAF) to classify them as MRD-positive or MRD-negative. MRD classifications were compared to a CLIA / CAP-validated orthogonal tumor-informed whole-genome sequencing-based ctDNA MRD assay. Additionally, as artifacts introduced during library preparation and sequencing remain challenges to sensitive low VAF mutation detection, a base-error model (BEM) that reduces sequencing artifacts and maximizes ctDNA signal was built using a series of high-quality lymph reference samples to quantify the background noise of each tumor variant 3 . A GPU-accelerated pipeline was implemented for this step when generating the baseline. Tumor-derived variants in lymph samples were considered artifacts if the VAF was not greater than BEM cutoff controlled by false discovery rate. We observed significant reduction in processing time and cost at the steps that implemented GPU acceleration. At alignment, we reduced processing time by 30% and computation cost by 15%. At tumor-informed variant calling, we reduced the average processing time from over 2 hours to 30 minutes per sample. For BEM, combined with optimization on VCF filtering, we have reduced processing time from over 5 hours to 45 minutes along with 60% reduction in computation cost. Furthermore, the comparison between our enhanced method and the CLIA / CAP-validated assay showed high concordance with p-val = 0.0005 using Fisher's exact test. We achieved 84% percent agreement, indicating the accuracy of the GPU-accelerated pipeline. We demonstrated that our GPU-accelerated MRD pipeline using NVIDIA Parabricks delivers clinical-grade accuracy with significant performance boost, enabling scalable, rapid molecular insights for adjuvant decision making. Given its advantages, we envision that GPU acceleration will prove useful to be a general strategy for deep sequencing applications.