Parallel mutual information estimation for inferring gene regulatory networks on GPUs

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Mutual information quantifies variable similarity, crucial in diverse fields like computational biology and machine learning.
• Traditional histogram-based estimators lack precision; B-spline methods offer competitive quality with lower complexity.

Purpose of the Study:

• To accelerate B-spline mutual information estimation using graphics hardware.
• To enable efficient inference of gene regulatory networks (GRNs) from large datasets.

Main Methods:

• Developed a parallel algorithm using the Compute Unified Device Architecture (CUDA) programming model.
• Implemented the CUDA-MI algorithm for B-spline mutual information estimation on GPUs.
• Utilized CUDA-MI for GRN inference from microarray data.

Main Results:

• Achieved speedups up to 82x on a single GPU compared to multi-threaded CPU implementations for large datasets.
• CUDA-MI produced higher quality GRNs than existing methods (ARACNE, TINGe) in less time.
• Demonstrated significant acceleration by exploiting GPU compute capabilities.

Conclusions:

• CUDA-MI is open-source software written in CUDA and C++.
• The method provides substantial speedup over sequential implementations.
• Leverages common CUDA-enabled GPUs for efficient computation.