CuGen: A GPU-accelerated framework for large-scale genomics

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Large-scale genomic analyses are computationally intensive and CPU-bound.
• Adjusting for confounding factors in biobank-scale data presents significant computational challenges.

Purpose of the Study:

• To present CuGen, a GPU-accelerated framework for efficient large-scale genomics.
• To enable computationally expensive analyses like genome-wide association studies (GWAS) and fine-mapping on massive datasets.

Main Methods:

• Utilized UltraLasso, a hierarchical application of univariate-guided sparse regression (uniLasso), for variant selection.
• Introduced the .cugen file format for memory-optimized GPU data handling.
• Developed a general GPU-accelerated genomics toolkit for prediction, QC, analysis, and visualization.

Main Results:

• Achieved robust leave-one-chromosome-out (LOCO) confounding control with a compact, phenotype-informed active set (<30,000 variants).
• Demonstrated CuGen's efficacy on UK Biobank data (up to 408,624 individuals), completing GWAS and fine-mapping in ~10 minutes on a single GPU.
• Showcased efficient scaling for phenome-wide analyses with sublinear resource consumption.

Conclusions:

• CuGen significantly reduces computational costs and time for biobank-scale genomic analyses.
• The framework enables efficient GWAS and in-sample fine-mapping with effective confounding control.
• CuGen provides a versatile GPU-accelerated toolkit for various genomics tasks, facilitating large-scale research.