Cost-effective GPU-grid for genome-wide epistasis calculations.

Area of Science:

• Computational genetics
• Bioinformatics
• High-performance computing

Background:

• Genotype studies traditionally focused on single nucleotide polymorphism (SNP) effects due to computational limitations.
• Epistasis, or gene-gene interactions, is crucial for understanding complex traits and diseases but computationally intensive.
• Advancements in computing power have recently made epistasis analysis more feasible.

Purpose of the Study:

• To develop a computationally efficient and cost-effective in-house solution for analyzing genetic epistasis.
• To leverage the parallel processing capabilities of Graphics Processing Units (GPUs) for genotype studies.
• To address the need for analyzing complex genetic interactions with confidential patient data.

Main Methods:

• Porting sequential epistasis calculations to GPUs using CUDA.
• Implementing massive parallelization on a local, GPU-based grid architecture.
• Comparing the performance and cost-effectiveness of GPU versus CPU parallelization.

Main Results:

• A cost-effective local grid combining consumer-level GPUs was established.
• The GPU-based approach demonstrated superior price/performance compared to cluster-based systems.
• A single GPU achieved performance comparable to 200 CPU cores for epistasis calculations.

Conclusions:

• The developed GPU-based approach is effective for problems amenable to massive parallelization.
• This method enables more comprehensive genotype studies, including epistasis.
• The project provides accessible code and ongoing tool development to facilitate the shift to parallel algorithms.