Droplet Biosciences, a diagnostics company using lymph-based liquid biopsy testing, is reducing genomic analysis time with NVIDIA AI infrastructure, improving test turnaround times.
Unlike blood tests, which are widely used for minimal residual disease (MRD) monitoring, lymphatic fluid allows for earlier collection and increased sensitivity. Droplet has found that postsurgical lymphatic fluid has 130 times more tumor-derived DNA molecules than time-matched blood plasma. This is because the fluid is local to the tumor site rather than being diluted in liters of blood.
Lymphatic fluid as a liquid biopsy source
A paper in Head & Neck journal by researchers at the University of Pittsburgh and Mayo Clinic, including a Droplet co-founder, found that proximal liquid biopsies, particularly surgical lymphatic fluid (SLF), show promise for postoperative disease monitoring, but “standardization and validation remain challenges,” and that “clinical implementation remains in the early stages.”
Droplet Biosciences’ first clinical test is a Clinical Laboratory Improvement Amendments (CLIA) validated clinical test for human papillomavirus (HPV)-negative head and neck cancer. Additionally, they have late-stage programs for lung and bladder cancer that are currently going through proof of concept. Collecting lymphatic fluid is applicable across solid tumors and can be applied to a wide variety of procedures and cancers in the future.
Droplet’s LymphDetect test uses deep sequencing to interrogate the presence of residual cancer after a tumor has been removed. NVIDIA published a new case study about Droplet’s use of NVIDIA Parabricks, a GPU-accelerated software suite designed to drastically speed up genomic data analysis for DNA sequencing. Using Parabricks, Droplet has achieved an order-of-magnitude reduction in the analysis time required across critical bioinformatics steps.
Accelerating genomic analysis from days to hours
Droplet focuses on detecting residual cancer after tumor removal by analyzing lymphatic fluid collected just 24 hours after surgery, allowing clinicians to assess residual disease earlier than other blood-based testing options, which are typically conducted four to six weeks after surgery, when enough time has passed for residual disease to become detectable in blood.
This is particularly relevant to rare residual cancer cells. As Wendy Winckler, chief scientific officer at Droplet Biosciences, explains, “finding them is like looking for a needle in a haystack. If your sample contains many more DNA fragments—and a higher concentration of tumor-derived DNA—you’re simply more likely to capture that ‘needle’ when you test. That’s translating to about double the sensitivity of time-matched blood.”
To detect rare variants, lymphatic fluid samples require significant next-generation sequencing (NGS) coverage, resulting in 5,000x coverage after deduplication, with raw coverage typically more than 10,000x. Without GPU acceleration, sequencing to this coverage depth can result in significant computational challenges.
By moving CPU-only pipelines to GPU-accelerated workflows, the company said it reduced key analysis steps from days to hours with NVIDIA Parabricks. This method accelerated variant calling time from approximately 10 hours to about one hour, reduced sequence alignment from up to 36 hours to less than three hours, and could compress overall analysis timelines from 10 days to two days.
This was driven by NVIDIA’s accelerated computing architecture, including the NVIDIA L4 Tensor Core GPU and NVIDIA L40S, which enable parallelization for genomics workloads.
To analyze unique molecular identifiers (UMIs), the team is working to implement fgbio, developed by Fulcrum Genomics, for UMI processing. This could help shorten UMI processing time from more than 15 hours to less than two hours, said Zhuosheng Gu, senior director of informatics, R&D at Droplet.
In addition to performance gains, Droplet Biosciences has realized operational benefits. Despite higher hourly costs for GPU compute, the reduced runtime results in a lower overall cost per sample.
