GPU-enabled Bioinformatics
why bioinformatics will move to accelerated hardware // benchmarks for nf-core/methylseq on GPUs // collaborate on our assay roadmap
The bioinformatics stack will get rewritten for GPUs over the next decade. Latch is disseminating accelerated code on a usage basis to provide large time and cost savings to the industry for popular analysis tasks.
In this essay, we outline why new molecular assays, managed data infrastructure and tailwinds from the “AI” boom, create the perfect storm for the regular use of accelerated hardware. We then release benchmarks for accelerated nf-core/methylseq for epigenetic analysis and invite biotech teams to collaborate on our assay roadmap.
Why bioinformatics code will run on GPUs
The volume of data generation in biotech continues to increase with the routine use of sequencing-based experiments that produce hundreds of GBs of data. We regularly see bioinformatics workflows take days to weeks to run. In some end-to-end laboratory workflows this is a concrete example of computation being a rate limiting step in drug discovery.
Most bioinformatics algorithms are incredibly parallelizable and benefit from the time and cost savings from implementations on accelerated hardware. As a new generation of larger molecular assay techniques come online, the impact of these GPU implementations move from irrelevant (~30MB bulk RNA-seq data) to desirable (~300GB methyl-seq data) to necessary (10M+ single cell atlases). Bench top kits now sequence millions of single cells and spatial tissue maps capture epigenetic state at 10 micrometer resolution. Molecular throughput is only increasing. These new techniques seem to benefit from improvements that double resolution, and therefore data, every few years.
The rise of managed cloud platforms also removes traditional barriers to the widespread use of GPU-based bioinformatics. Installing hardware, managing drivers, building and configuring CUDA programs, optimizing parameters all require expertise and upfront time investment. Usage-based workflow infrastructure allows routine processing jobs, and even much smaller workloads, to benefit from time and cost savings previously only accessible to large volume research projects with dedicated teams.
These platforms also create the concentration of demand, and favorable pay, to attract and retain the sparse talent needed to write and maintain GPU programs. The previous options to develop fit-for-purpose GPU bioinformatics code was academia, but talented systems engineers had no desire to make $23K a year only to have the great privilege of wrestling with the ivory tower.
The recent AI craze is also commoditizing GPUs, improving the developer ecosystem and creating new infrastructure to use them at scale. Data intensive bioinformatics workloads will benefit from the new libraries, improvements in systems software and increasing GPU memory induced by the demand for training large models.
Accelerated nf-core/methylseq for epigenetics
As a first case-study, we are releasing an accelerated version of nf-core/methylseq that uses the GPU aligner Arioc developed by Richard Wilton.
With 4 V100s, we reduce the runtime of bisulfite alignment on 500M reads from 45.32 hours to 13.84 hours while also dropping cost 19%. Depending on provisioned hardware, runtime can be further optimized at the expense of cost, dropping to a 9.4 hour runtime on 8 V100s with a ~7.5% increase in cost.
Notice the improvements are large even with smaller workloads, making GPUs appealing for team of all sizes. We are still exploring optimal configuration values, collecting benchmarks on newer GPUs and exploring discounts with different vendors.
Wilton has been developing this program for years and has released some great papers on the development and practical use of GPU-based bioinformatics code (also very responsive on GitHub). Please support and cite his work.
To ensure these results are at least as good as the gold standard CPU aligner Bismark, we compared the concordance of individual reads. More than 90% of the alignments produced by Bismark are also identified by ARIOC, with an additional 13.5% of the read alignments remaining specific to ARIOC. This benchmark also makes it easy to fine-tune the alignment parameters to achieve a trade-off between throughput and sensitivity.
Collaborate on our assay roadmap
We will continue to improve bisulfite sequencing analysis, but plan to extend the benefits of GPU implementations to all large assays. WGS and single-cell methylation are in the works. We encourage biotech teams that are interested in these cost and performance benefits to reach out to our team.
Shoot me an email at kenny@latch.bio.
This work was spearheaded by Harihara Muralidharan, a talented bioinformatics engineer at LatchBio.
Thank you to Rob Patro for his insightful comments at the scverse conference and Richard Wilton for his work.
 | A guest post by
|