Innovative techniques to speed up single-step genomic evaluations integrated into MiXBLUP 3.0 software

This research was initially performed by PhD student Alexander Freudenberg and Prof. Martin Schlather of the University of Mannheim, and their results were integrated by WUR-ABG researchers into MiXBLUP 3.0, a software for large genomic evaluations.

MiXBLUP 3.0

The software package MiXBLUP 3.0 (www.mixblup.eu) is developed by Wageningen University & Research in collaboration with the Natural Resources Institute Finland (LUKE), and it allows to efficiently estimate genomic enhanced breeding values in livestock using various linear mixed models. MiXBLUP 3.0 aims to be user-friendly while supporting efficient algorithms for solving large genomic evaluations. Originally developed for pedigree-based models, MiXBLUP has been extended to four different approaches for single-step genomic evaluations, allowing simultaneous analyses of all information for both genotyped and ungenotyped individuals. For all approaches, shared-memory parallelism is supported using tailor-made procedures and parallel libraries. The fast increase of genomic information, including up to several millions of SNP genotypes for the largest current datasets, however, limits the feasibility of routine single-step genomic evaluations within a specific timeframe with current software and algorithms.

Faster single-step evaluations through novel algorithms and GPUs

Researchers at the University of Mannheim investigate several techniques to improve the multiplication of the SNP genotype matrix by another matrix. This multiplication is indeed an important bottleneck for performing routine single-step breeding value estimation when the dataset includes millions of SNP genotypes. First, they developed a novel algorithm (called 5codes) for storing SNP genotypes based on combinatorics. This novel algorithm reduces data streams through the cache hierarchy on CPUs. Second, they implemented a procedure for the genotype matrix-matrix multiplication in CUDA that leverages the powerful HPC capabilities of modern Nvidia GPUs. Finally, researchers at WUR-ABG integrated the 5codes and GPU-based algorithms in MiXBLUP 3.0. Using a dataset including about 2.61 million genotyped animals – provided by the Irish Cattle Breeding Federation – they observed speed-ups by a factor of ca. 1.5 for the 5codes algorithm and ca. 2.5 for the GPU-based algorithm in comparison to the current approach. These speed-ups allowed MiXBLUP 3.0 to estimate genomic EBV for a single-step evaluation including 2.61 million genotyped animals in less than 2 hours.

Do you want more information about MiXBLUP 3.0? Please contact Jan ten Napel (jan.tennapel@wur.nl) or jeremie Vandenplas (jeremie.vandenplas@WUR.nl).