SAIGE-GPU: accelerating genome- and phenome-wide association studies using GPUs.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Genome-wide association studies (GWAS) are computationally intensive, particularly for admixed populations.
• Existing methods like SAIGE, while effective for generalized linear mixed-model GWAS, are CPU-bound, limiting their application for large-scale phenome-wide association studies.
• The computational demands hinder accessibility for many research groups.

Purpose of the Study:

• To develop a GPU-accelerated version of SAIGE (SAIGE-GPU) to overcome computational limitations.
• To enable efficient GWAS for large biobank-scale datasets, including diverse and admixed populations.
• To reduce the time and cost associated with large-scale genetic association analyses.

Main Methods:

• Developed SAIGE-GPU by replacing CPU-intensive matrix operations with GPU-optimized kernels.
• Implemented distributed genetic relationship matrix calculations across GPUs and communication layers.
• Optimized variant association testing using multi-core and multi-trait parallelization.

Main Results:

• SAIGE-GPU achieved a 5-fold speedup in mixed model fitting on supercomputing and cloud platforms.
• The method was successfully applied to 2,068 phenotypes from 635,969 participants in the Million Veteran Program (MVP).
• Deployment on Google Cloud Platform and Azure demonstrated substantial cost and time savings.

Conclusions:

• SAIGE-GPU significantly enhances the efficiency of large-scale GWAS, especially for admixed populations.
• The GPU acceleration makes complex genetic analyses more feasible and accessible.
• The tool offers practical benefits in terms of computational time and cost for researchers.