Summary
A bioRxiv preprint introduces ARCHER-LD, a multi-GPU tool for calculating linkage disequilibrium across very large sequencing datasets. The authors report speedups of about eightfold to 28-fold over PLINK in benchmark tests and complete a 10-quadrillion-element LD matrix in under six hours on 512 GPUs.
A bioRxiv preprint presents ARCHER-LD, a software method that uses distributed graphics-processing units (GPUs) to calculate linkage disequilibrium (LD) between every pair of genetic variants in very large sequencing datasets. The authors report that the approach completed selected calculations between about eight and 28 times faster than PLINK, a widely used genetics tool, in tests on whole-genome sequencing data.
LD describes the statistical association between variants in a population. It is used in analyses such as statistical fine-mapping, which seeks to narrow down genomic regions and identify variants most likely to be relevant to a trait or disease. Researchers generally prefer LD calculated from the same dataset being analysed, but the number of variant pairs grows approximately as N²/2 for N variants. That makes all-pair calculations increasingly expensive as sequencing datasets grow.
A GPU method for all-pair LD calculations
ARCHER-LD calculates R2, a commonly used measure of the strength of LD, for every variant pair in a dataset. The method distributes the work across multiple GPUs and is implemented in Python using CuPy, a library that enables GPU-accelerated numerical computation. The authors have made the tool publicly available through a GitHub repository linked from the preprint.
The computational problem is particularly large for biobank-scale whole-genome sequencing. Such datasets can contain hundreds of millions of variants, creating a very large number of pairwise calculations. Because of that cost, many studies use an external reference panel, such as the 1000 Genomes resource, rather than calculating LD directly from their own participants.
The authors say ARCHER-LD can make it practical to use dataset-specific LD instead. They argue that this could support more population-specific analyses, including for groups that are underrepresented in existing reference databases.
Reported benchmark results
In one test, ARCHER-LD processed chromosome 22 from the 30x whole-genome sequencing version of the 1000 Genomes dataset. The chromosome contained about 1.8 million variants. Eight NVIDIA RTX 2080 Ti GPUs, each with 12 GB of memory, completed the calculation in less than 10 minutes. The same calculation using PLINK on a 64-thread Intel Xeon Gold 6338 processor took more than 80 minutes, which the authors describe as an approximately eightfold speedup.
A larger test used the Penn Medicine Biobank, which contains 57,170 samples. For approximately 1.38 million variants on chromosome 1, ARCHER-LD completed the LD calculation in 49 minutes, compared with 22.7 hours for PLINK. The reported speedup in this experiment was approximately 28-fold.
The preprint also reports a full-scale demonstration on the 30x whole-genome sequencing 1000 Genomes dataset, containing about 120 million variants and approximately 2,500 samples. Using 512 NVIDIA A100 GPUs with 40 GB of memory each on the Polaris supercomputer at the US Department of Energy's Argonne Leadership Computing Facility, the team calculated an LD matrix containing more than 10^16 elements—over 10 quadrillion values—in under six hours.
These results are performance measurements from the specified datasets, software comparison and computing hardware. The work is presented as a bioRxiv preprint, and the reported timings therefore describe the authors' implementation and test environments rather than a universal speed guarantee. Its main technical contribution is a way to apply multi-GPU computing to a calculation whose cost rises rapidly with the number of variants, potentially allowing researchers to work with LD derived from their own large genomic datasets.