MMGS: A novel genomic prediction framework to integrate genotype, environment and their interactions for multi-environment breeding trials

Mingjia Zhu, Zeyu Zheng, Wei Liu, Yu Han, Wenjie Mou, Tongming Yin, Xiaogang Dai, Huaitong Wu, Yongzhi Yang, Yanjun Zan, Jianquan LiuHorticulture 2026 · DOI

TL;DR

This paper introduces MMGS, a novel framework called PEI that improves genomic prediction accuracy by explicitly modeling the complex interactions between genotypes and environments. This new approach outperforms traditional methods, especially in predicting performance in novel environments, and is made accessible through an open-source R package.

What Problem Does This Paper Address?

Predicting how well different plant varieties (genotypes) will perform across various locations and climates (environments) is crucial for breeding crops and trees that can withstand changing conditions and meet global food demands. Traditional methods often struggle to capture the intricate interplay between genotype and environment (GxE). While the "reaction norm" (RN) framework has been used to model this, it often assumes linear relationships and may not fully leverage the genetic basis of plasticity. Existing methods also tend to rely on a limited set of statistical estimators, missing out on the potential of more advanced machine learning techniques. This gap limits our ability to accurately predict performance, especially in novel or unseen environments, hindering breeding efficiency.

Methods

The core of this work is the Polygenic Environmental Interaction (PEI) framework. Unlike the classical Reaction Norm (RN) framework, PEI directly models GxE interactions by integrating genotype information with environmental covariates (ECs). A key step involves using the Critical Environmental Regressor through Informed Search (CERIS) algorithm to identify the most informative environmental index that explains the most phenotypic variation.

PEI then constructs a relationship matrix that includes pairwise interactions between genetic markers and this identified environmental index. This allows for a more comprehensive modeling of genotype-environment dynamics. The framework is designed to be flexible, incorporating an ensemble of 15 different estimators, ranging from traditional parametric methods (like G-BLUP, BayesA) to semi-parametric (RKHS) and non-parametric/machine learning approaches (SVM, Random Forest, GBM). This broad selection of estimators allows for accounting for diverse genetic architectures and potential non-linear environmental responses.

The authors also developed an R package called MMGS to make this framework readily available to researchers and breeders.

Datasets

The study utilized three distinct datasets to benchmark the PEI framework:

  1. Maize CUBIC population: A multi-parental advanced inter-cross population with 1404 inbred lines phenotyped across five diverse locations in China. This dataset represents a complex genetic architecture.
  2. Inbred wheat population: A set of 287 inbred lines from the International Maize and Wheat Improvement Center (CIMMYT), phenotyped across nine locations globally (India, Pakistan, Nepal, Bangladesh, Iran, Egypt, Sudan, and Mexico). This dataset represents a different population structure and broader environmental range.
  3. F₁ willow population: An outbred population of 501 F₁ lines phenotyped across three locations in Sichuan province, China. This dataset was used to assess performance in hybrid populations.

These datasets allowed for evaluation across different population structures (inbred vs. outbred, diverse heterotic groups) and various traits with distinct genetic architectures.

Key Insights and Ideas

  • Direct GxE Modeling is Powerful: Explicitly modeling the interaction between genotype and environment, rather than just accounting for them separately or linearly, leads to significant improvements in prediction accuracy.
  • Environmental Index is Key: Identifying the most influential environmental factor (e.g., daily highest temperature for some traits) is crucial for effectively capturing GxE.
  • Ensemble of Estimators Matters: The choice of statistical estimator can significantly impact prediction accuracy, and a flexible framework that can accommodate various approaches (especially machine learning) is beneficial.
  • Hybrid Populations are Tricky: While PEI excels in inbred lines, its performance can be less consistent in hybrid populations, potentially due to the greater influence of non-additive genetic effects (epistasis) in such crosses.
  • Novel Environments are Predictable: PEI shows a strong advantage in predicting how genotypes will perform in environments they haven't been tested in before, which is a critical challenge in breeding.

Key Results

  • Superior Accuracy: The PEI framework consistently outperformed the classical Reaction Norm (RN) framework across multiple traits and populations. For example, in the maize population, PEI showed an average increase in accuracy of 10%, 8.7%, and 20.6% for days to anthesis (DTA), plant height (PH), and ear weight (EW), respectively. Similar improvements were observed in the wheat population.
  • Predicting Novel Environments: PEI demonstrated higher predictive accuracy in 53.3% of predictions for novel environments in the maize population, with an average improvement of 16.5%. For PH, PEI consistently outperformed RN, with an average improvement of 21.7%.
  • Estimator Variability: The optimal estimator varied depending on the trait, population, and framework used. For instance, EN was optimal for DTA in RN, while LASSO was best for PEI.
  • Willow Population Nuance: In the F₁ willow population, PEI showed a reduction in average accuracy for PH and GD compared to RN, but achieved a 20.6% increase for defoliation time (DT). This highlights the importance of population structure.
  • Computational Cost: PEI requires more computational resources than RN, especially with a large number of environments, due to the creation of larger super-matrices. However, the use of advanced estimators helps manage computational burden.

Limitations and Future Work

  • Hybrid Population Performance: The paper acknowledges that PEI's performance can be less optimal in hybrid populations where epistasis plays a larger role. Future work could explore how to better model these non-additive genetic effects within the PEI framework.
  • Environmental Index Sensitivity: While PEI is less sensitive to the choice of environmental index than RN, the framework still relies on a single index. Future research could investigate incorporating multiple environmental indices to further enhance model performance.
  • Computational Scaling: For extremely large datasets (high dimension of population size x markers), computational cost can increase substantially. The authors suggest considering alternative estimators in such scenarios and advise users to be mindful of this trade-off.
  • Benchmarking: The authors attempted to benchmark against other frameworks like GEFormer and AutoGS. While GEFormer was successfully benchmarked, AutoGS was unavailable due to software updates, indicating a need for continued comparison with emerging methods.

The release of the MMGS R package is a significant step towards enabling broader adoption and further development of these advanced genomic prediction methods.

Summary updated

Abstract

Abstract Accurately predicting the performance of trees and crops across diverse and changing climates is essential for matching genotypes to both current and future environments. Yet modelling the complex interplay among genotype, environment, and phenotype in multi-environment trials remains a major challenge. Here, we introduce a unified framework, polygenic environmental interaction (PEI), directly models genotype-by-environment interactions through integrating genotypes and environmental covariates. We implemented an ensemble of 15 estimators spanning parametric, non-parametric, and machine-learning approaches. We then benchmarked our framework against the classical reaction norm (RN) using three genetically distinct populations and three traits with variable genetic architectures. Furthermore, we released an open-source R package, Multiple-Environments Genomic Selection (MMGS), on GitHub. Together, our study offers a flexible and computationally efficient approach for multi-environment genomic prediction, enhancing breeding efficiency, providing deeper insights into modelling the genotype-environment-phenotype continuum.
BibTeX
@inproceedings{zhu2026mmgs,
  author = {Mingjia Zhu and Zeyu Zheng and Wei Liu and Yu Han and Wenjie Mou and Tongming Yin and Xiaogang Dai and Huaitong Wu and Yongzhi Yang and Yanjun Zan and Jianquan Liu},
  title = {{MMGS: A novel genomic prediction framework to integrate genotype, environment and their interactions for multi-environment breeding trials}},
  booktitle = {Horticulture},
  year = {2026},
  doi = {10.1093/hr/uhag035},
}
Figure