TorchGWAS : GPU-accelerated GWAS for thousands of quantitative phenotypes

Area of Science:

• Bioinformatics
• Computational Biology
• Genetics

Background:

• Modern bioinformatics generates thousands of quantitative phenotypes, creating computational bottlenecks for traditional genome-wide association studies (GWAS).
• Existing GWAS tools are inefficient for phenotype-rich screening workflows that reuse genotype data across numerous traits.

Purpose of the Study:

• To develop TorchGWAS, a high-throughput framework for association testing in large phenotype panels using hardware acceleration.
• To provide efficient GWAS screening for phenotype-rich datasets.

Main Methods:

• Implemented in Python, TorchGWAS utilizes hardware acceleration (NVIDIA A100 GPU) for association testing.
• Supports linear GWAS and multivariate phenotype screening with NumPy, PLINK, and BGEN genotype inputs.
• Includes internal covariate adjustment and sample identifier alignment.

Main Results:

• TorchGWAS achieved a 300- to 1700-fold increase in phenotype throughput compared to traditional methods.
• Processed 20,480 phenotypes in 20 minutes on a single GPU, compared to fastGWA's ~100 seconds per phenotype on a multi-core CPU.
• Demonstrated practicality for large-scale GWAS screening in phenotype-rich settings.

Conclusions:

• TorchGWAS significantly enhances the efficiency of genome-wide association studies for large numbers of quantitative phenotypes.
• Enables practical and rapid screening of thousands of traits, overcoming computational bottlenecks in modern bioinformatics.
• Provides a valuable tool for genetic research in complex, phenotype-rich datasets.