parSMURF, a high-performance computing tool for the genome-wide detection of pathogenic variants.

Area of Science:

• Genomic Medicine
• Computational Biology
• Machine Learning

Background:

• Genomic medicine and computational biology face challenges with big data and imbalanced datasets.
• Positive examples (e.g., pathogenic variants) are often a small fraction of the total data.
• Classical methods struggle to identify rare pathogenic variants and manage large genomic datasets.

Purpose of the Study:

• To develop a method that addresses big and imbalanced genomic data challenges.
• To improve the prediction of rare pathogenic variants in genomic medicine.
• To accelerate computational analysis of genomic data.

Main Methods:

• parSMURF utilizes a hyper-ensemble approach with oversampling and undersampling techniques.
• Parallel computational techniques (MPI, OpenMP) are employed for big data management and speed.
• Bayesian optimization facilitates efficient hyper-parameter tuning.

Main Results:

• parSMURF successfully manages big genomic data by partitioning it across computing nodes.
• Achieved state-of-the-art results on synthetic and real-world genomic datasets.
• Demonstrated an 80-fold speed-up compared to sequential versions.

Conclusions:

• parSMURF is a scalable parallel machine learning tool for big and imbalanced genomic data.
• Its parallelization enables efficient fitting of complex genomic problems.
• Available in C++ OpenMP and C++ MPI/OpenMP hybrid versions for diverse computing environments.