parSMURF

parSMURF applies hyper-ensemble machine learning and high-performance computing to identify deleterious or pathogenic single nucleotide variants (SNVs) in large, class-imbalanced genomic datasets, including non-coding regions.

Key Features:

• Hyper-Ensemble Approach: Integrates multiple machine learning models into a hyper-ensemble to improve predictive accuracy on highly imbalanced datasets where pathogenic examples are a minority.
• Oversampling and Undersampling Techniques: Uses oversampling of rare (pathogenic) instances and undersampling of abundant (neutral) instances to mitigate class imbalance.
• Parallel Computational Techniques: Partitions large datasets across nodes with MPI and accelerates intra-node computation with OpenMP to enable parallel processing on HPC clusters.
• Bayesian Optimization for Hyperparameter Tuning: Applies Bayesian optimization to automatically tune algorithm hyperparameters for improved model performance.
• High Scalability and Flexibility: Provides a C++ OpenMP version for single workstations and a hybrid C++ MPI/OpenMP version for distributed HPC environments, supporting multi-level parallelization across multi-core and cluster systems.
• Performance: Demonstrated ability to handle millions of examples and to achieve up to an 80× speed-up compared with sequential methods.

Scientific Applications:

• Mendelian disease-associated SNV detection: Identification and prioritization of pathogenic SNVs associated with Mendelian diseases, including variants in non-coding regions.
• GWAS hit prioritization in non-coding regions: Prioritization of potentially deleterious genome-wide association study (GWAS) hits located in non-coding genomic regions.
• Large-scale genomic studies and benchmarking: Analysis and benchmarking on millions of examples using synthetic datasets and real variant collections to evaluate performance on big genomic data.

Methodology:

Combines hyper-ensemble (ensemble learning), oversampling and undersampling, Bayesian optimization for hyperparameter tuning, and multi-level parallelization using OpenMP and MPI; implemented in C++ with OpenMP and hybrid MPI/OpenMP versions.