Fast numerical optimization for genome sequencing data in population biobanks.

Area of Science:

• Genomics
• Computational Biology
• Statistical Genetics

Background:

• Large-scale genome sequencing generates high-dimensional data, posing significant computational challenges.
• Existing general-purpose optimization tools often exhibit suboptimal performance for genetic data analysis in terms of speed and memory usage.

Purpose of the Study:

• To develop efficient computational solvers tailored for large-scale regularized regressions on genetic data.
• To address the memory and computational bottlenecks in analyzing millions of genetic variants from hundreds of thousands of individuals.

Main Methods:

• Developed a two-bit representation for genetic variant encoding, reducing memory requirements by 32-fold.
• Implemented an iteratively reweighted least square algorithm (snpnet-2.0) for Lasso regressions using the compact genetic matrix representation.
• Created a sparse genetic matrix implementation utilizing the two-bit encoding and a compressed sparse block format for parallelized matrix-vector multiplications.
• Developed an accelerated proximal gradient method (sparse-snpnet) for group Lasso on sparse genetic matrices.

Main Results:

• The developed solvers efficiently handle large-scale genetic matrices (1 million variants, 100,000 individuals).
• Achieved significant reductions in memory usage (less than 32GB) and computational time (within 10 minutes).
• Demonstrated the ability to solve linear, logistic, and Cox regression problems using Lasso and group Lasso.

Conclusions:

• The novel computational approaches provide highly efficient solutions for analyzing massive genomic datasets.
• These methods significantly improve memory and computational performance, facilitating large-scale genetic association studies.
• The snpnet R package offers powerful tools for researchers working with high-dimensional genetic data.