Mutual information-based best linear unbiased prediction for enhanced genomic prediction accuracy

The use of genomic prediction (GP) in breeding programs has increased with the advancement of high-throughput sequencing. However, current methods face limitations: 1) genomic best linear unbiased prediction (GBLUP) assumes that all markers contribute equally to the genetic variance, which can limit accuracy for traits influenced by a few major genes; 2) Bayesian methods, although more flexible, often require intensive computation. To address these challenges, we present mutual information-based best linear unbiased prediction (MIBLUP), a novel framework that integrates marker selection and weighted genomic relationship matrices (GRM). MIBLUP employs a two-step approach: 1) using minimum redundancy maximum relevance and cross-validation to select informative markers as covariates and 2) constructing a mutual information-weighted GRM to prioritize markers with stronger trait associations. In simulations involving 7 scenarios with three heritability levels (0.2, 0.5, and 0.8), MIBLUP achieved up to 0.091 higher accuracy than GBLUP in four scenarios, with an average improvement of 0.056 across all scenarios. When evaluated on five real datasets (cattle, pig, loblolly pine, Duroc pig, and chicken), MIBLUP outperformed other methods such as GBLUP, Bayes R, BSLMM, LDAK-Bayes R, LDAK-Bolt, SLEMM, KAML, and BLUP|GA in overall prediction accuracy. Computationally, MIBLUP demonstrated substantially higher efficiency in both speed and memory usage compared with Bayesian methods such as Bayes R, making it more suitable for large-scale genomic datasets. Overall, MIBLUP shows promise as a reliable tool for enhancing GP performance in animal and plant breeding.