Heterogeneous computing architecture for fast detection of SNP-SNP interactions

Area of Science:

• Bioinformatics
• Computational Biology
• Genetics

Background:

• Genome-wide association studies (GWAS) generate vast datasets, presenting significant computational challenges for gene-gene interaction discovery.
• Exhaustive analysis of millions of single nucleotide polymorphisms (SNPs) can take weeks or months with traditional methods.
• Massively parallel hardware, including Graphics Processing Units (GPUs) and Many Integrated Core (MIC) coprocessors, offers potential for substantial runtime reduction.

Purpose of the Study:

• To develop and evaluate a heterogeneous, accelerated software module for SNP-SNP interaction discovery.
• To compare the performance and programming effort of GPU and Intel MIC architectures for this task.
• To integrate the accelerated module into the existing SNPsyn web-based program.

Main Methods:

• Development of a heterogeneous software module leveraging both GPU and Intel MIC architectures.
• Implementation of SNP-SNP interaction discovery algorithms on these parallel processing platforms.
• Comparative analysis of execution times and programming complexity between GPU, MIC, and single-threaded CPU implementations.

Main Results:

• The new accelerated module achieved an order of magnitude reduction in execution time compared to the single-threaded CPU version.
• GPU implementation (Nvidia Tesla K20) demonstrated twice the performance of the MIC architecture (Intel Xeon Phi P5110).
• GPU implementation required significantly more programming effort than the MIC architecture.

Conclusions:

• General-purpose GPUs offer mature, powerful parallel processing capabilities for computationally intensive tasks like SNP-SNP interaction discovery, though they demand greater programming investment.
• The newer MIC architecture, while less performant, offers a more accessible programming model and a versatile architecture suitable for diverse computational problems.