SLEMM: million-scale genomic predictions with window-based SNP weighting

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic data is growing exponentially, making large-scale genomic prediction computationally challenging.
• Accurate genomic prediction is crucial for breeding programs in plants and livestock.

Purpose of the Study:

• To introduce SLEMM (Stochastic-Lanczos-Expedited Mixed Models), a novel software tool designed to overcome computational hurdles in genomic prediction.
• To enhance prediction accuracy through the implementation of SNP weighting within the SLEMM framework.

Main Methods:

• SLEMM utilizes an efficient implementation of the stochastic Lanczos algorithm for Restricted Maximum Likelihood (REML) within a mixed model framework.
• The study involved extensive analyses on seven public datasets across plant and livestock species, comparing SLEMM with other genomic prediction methods.
• Simulations were conducted on datasets with up to 3 million individuals and 1 million SNPs to assess computational performance.

Main Results:

• SLEMM with SNP weighting demonstrated superior predictive ability compared to GCTA's empirical BLUP, BayesR, KAML, and LDAK's BOLT and BayesR models on public datasets.
• In analyses of nine dairy traits from ~300k cows, SLEMM showed comparable prediction accuracies to other methods, with KAML failing to process the data.
• Simulation studies confirmed SLEMM's computational advantage over counterparts for million-scale genomic predictions.

Conclusions:

• SLEMM provides a computationally efficient solution for large-scale genomic predictions, achieving accuracy comparable to established methods like BayesR.
• The software's ability to handle millions of individuals and SNPs makes it a valuable tool for modern genomic selection.
• SLEMM is publicly available, facilitating its adoption in plant and animal breeding research.