Estimation & Prediction

Genomic prediction through transfer learning: algorithms, software, and benchmarks

---

Abstract

Genomic prediction has been widely adopted in plant and animal breeding. Over time, breeding programs have generated large datasets comprising high-dimensional DNA genotypes linked to pedigree, phenotypes, and environmental information (e.g., feeding regimes or contemporary groups). These datasets can span tens of generations and may include data from several breeding lines and multiple environments.The massive datasets available, together with the dynamic nature of SNP effects (resulting from differences in allele frequencies, linkage disequilibrium, and possibly Gà—E), introduce important computational and statistical challenges. First, frequently updating SNP-effect estimates (e.g., weekly) represents a major computational burden. Second, if SNP effects differ across populations, generations, and environments, estimating them from all available data may be suboptimal.Transfer Learning (TL) is a machine learning technique used to improve a model's performance in a target population by leveraging knowledge from another population where the model has already performed well. We propose TL as an effective approach to address several challenges emerging in modern genetic evaluations.Although TL has been used to improve polygenic score prediction in humans, we currently lack TL algorithms for models involving genetic and non-genetic factors (e.g., contemporary groups) and for Gaussian processes (e.g., GBLUP and Single-Step GBLUP, SS-GBLUP). Therefore, in this study, we present novel TL methods using Gradient Descent with Early Stopping (GDES) for SNP-centric linear regression, GBLUP, and SS-GBLUP. For SNP-centric models, we present a simple extension of GDES that accommodates no-genetic fixed and random effects. For GBLUP and SS-GBLUP, we introduce a new TL approach based on a sequential decomposition of the G (or H) matrix.Using real breeding data, we benchmark GDES against standard GBLUP for prediction with multi-generation and multi-environment data. In the multi-generation benchmark, TL implemented through GDES, used SNP-effect estimates from earlier generations as initial values for a GDES algorithm that updates SNP effects using only the most recent data. By comparison, standard GBLUP recalculates SNP effects by refitting the model to all available data whenever new data are added.In the multi-environment benchmark, we used SNP-effect estimates from one environment as prior values for a TL approach that updated estimates using data from the target environment. We compare this approach with cross-environment, within-environment, and multi-environment GBLUP.Our results show that GDES can achieve prediction accuracy that is near identical than that of GBLUP-and in some multi-environment applications even better-while requiring only a fraction of the computing time and resources needed to estimate breeding values using all available data.

Keywords: 2026

How to Cite:

de los Campos, G., Wu, H., Pérez-Rodríguez, P., Toledo-Alvarado, H., Steibel, J., Cheng, H., Dekkers, J., Lourenco, D., Bermann, M., Chen, C., Holl, J. & Valente, B., (2026) “Genomic prediction through transfer learning: algorithms, software, and benchmarks”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2283141. doi: https://doi.org/10.31274/wcgalp.23487

Rights: 1