Adjusting for Spatial Effects in Genomic Prediction

Xiaojun Mao, Somak Dutta, Raymond K. W. Wong, Dan NettletonJournal of Agricultural Biological and Environmental Statistics 2020 · DOI · arXiv

Abstract

In plant breeding, genomic prediction aims to estimate phenotypic performance from genotypic data. However, environmental heterogeneity within field trials—such as soil gradients or moisture levels—often introduces spatial correlations that confound genetic signals. This paper introduces a Gaussian Random Field (GRF) framework that simultaneously models genetic markers, subpopulation structures, and spatial correlations. By accounting for the structured "microenvironment" of neighboring plants, the GRF model demonstrates superior accuracy and more reliable genotype ranking compared to standard Mixed Linear Models (MLM).

Introduction

The primary challenge in genomic selection is distinguishing a plant's genetic potential from environmental noise. Standard models often treat spatial variation as random, unstructured error. In reality, factors like soil nutrient gradients or shade create a structured microenvironment where neighboring plants exhibit similar phenotypes regardless of their genetic distance. Ignoring these spatial effects can lead to "false positives," where inferior genotypes are selected simply because they were grown in favorable field locations.

Methodology: The Gaussian Random Field Framework

The authors propose a GRF model utilizing an additive covariance structure. This approach decomposes phenotypic variance into three distinct kernels:

  1. Marker Kernel ($C_g$): Captures genotypic relationships using SNP data. To manage high dimensionality, the FarmCPU algorithm is employed to select the most significant SNPs.
  2. Subpopulation Kernel ($C_b$): Accounts for broader family effects within the population, capturing similarities that individual SNPs might overlook.
  3. Spatial Kernel ($C_s$): The core innovation, utilizing a stationary autoregression model on a regular lattice. This allows the model to "smooth" spatial effects, treating a plant's performance as a function of its specific $(x, y)$ coordinates in the field.
Figure 1: Example of a field layout illustrating the spatial distribution of experimental plots. The GRF model uses these coordinates to account for localized environmental correlations.

The model parameters are estimated via Maximum Likelihood Estimation (MLE), providing the flexibility to adapt to various field geometries, from long narrow rows to complex square plots.

Experimental Evaluation

The study utilized two primary datasets to validate the GRF model:

  • Maize (NAM Panel): 4,660 recombinant inbred lines (RILs) across 25 subpopulations, measuring $CO_2$ emissions.
  • Wheat: 384 advanced lines measuring grain yield and kernel weight under varying irrigation conditions.

Key Results

The GRF model consistently outperformed benchmarks such as Compressed MLM (CMLM), Row-Column (RC) models, and Moving-Means (MVNG). In maize trials with high spatial noise, GRF achieved an accuracy of 0.4904, compared to 0.1930 for CMLM. Similarly, for irrigated wheat grain yield, GRF reached an accuracy of 0.7575, a massive improvement over CMLM's 0.2091.

Discussion and Conclusion

A critical insight from this research is that spatial noise is highly structured. By quantifying the ratio of spatial variance to genetic variance ($\gamma$), researchers can determine the extent of environmental "pollution" in a trial. Furthermore, the GRF model excels at ranking genotypes—identifying the "Top 10" plants—which is the ultimate goal for breeders.

While the model is highly effective for contiguous grids, future work will address its performance across disconnected spatial domains (e.g., multiple distant trial sites) and optimize computational efficiency for larger datasets. Ultimately, integrating spatial effects into genomic prediction provides a more robust path for identifying truly superior genetics in diverse field conditions.

Summary updated

BibTeX
@article{mao2020adjusting,
  author = {Xiaojun Mao and Somak Dutta and Raymond K. W. Wong and Dan Nettleton},
  title = {{Adjusting for Spatial Effects in Genomic Prediction}},
  journal = {Journal of Agricultural Biological and Environmental Statistics},
  year = {2020},
  doi = {10.1007/s13253-020-00396-1},
}
Figure