Weighted Kernels Improve Multi-Environment Genomic Prediction

Xiaowei Hu, Brett F. Carver, Yousry A. El‐Kassaby, Zhu Lan, Charles ChenbioRxiv 2022 · DOI

TL;DR

This paper introduces a novel "weighted kernel" approach for genomic selection (GS) that significantly improves prediction accuracy across multiple environments by incorporating genetic association signals (like MAF and GWAS p-values) alongside genomic similarity. This method offers a more robust and flexible framework for predicting the performance of genetically diverse populations, crucial for accelerating crop and tree breeding programs.

What Problem Does This Paper Address?

Predicting how well a plant variety will perform across different environments is critical for breeding programs. However, traditional genomic selection (GS) models often struggle with "genotype by environment" (GxE) interactions, where the genetic basis of a trait changes depending on the environment. This leads to a drop in prediction accuracy when using single-environment models. While multi-environment (ME) models exist, they often rely on Gaussian kernels that primarily capture overall genomic similarity, neglecting valuable insights from genome-wide association studies (GWAS) that link specific genetic variants to traits. This paper aims to bridge this gap by developing a GS framework that leverages these GWAS insights to improve multi-environment prediction, especially for genetically heterogeneous populations.

Methods

The core innovation is a Weighted Kernel (WK) approach. Unlike standard Gaussian kernels (GK) that assign equal importance to all genetic markers, WK assigns weights to markers based on their minor allele frequency (MAF) and their association significance (p-values from single-marker association tests for both genotypic and GxE effects). This allows the model to prioritize markers that are more informative about trait variation and GxE interactions. The authors developed a Bayesian framework to implement these WK models, allowing for the estimation of various genetic parameters and providing flexibility for different population structures. They compared their WK models against standard single-environment (SE) and multi-environment Gaussian kernel (ME_GK) models.

Datasets

The study utilized two main datasets:

  • Winter Wheat DH Population: A Doubled Haploid (DH) population of 242 lines derived from the intercross of Duster and Billings varieties, evaluated over three years (2014-2016) in Stillwater, Oklahoma, USA. Traits included grain yield (GY), SDS sedimentation value (SDS), kernel weight (SKCSKW), and wheat protein (WHTPRO).
  • Interior Spruce Population: A breeding population of 1,126 38-year-old trees from three sites in British Columbia, Canada. Traits included height (HT), diameter at breast height (DBH), resistance to drilling (WDres), and wood density (WDx-ray).

Key Insights and Ideas

  • Beyond Simple Similarity: The study demonstrates that simply measuring overall genomic similarity (as done by Gaussian kernels) is insufficient. Incorporating specific genetic signals, like MAF and GWAS p-values, can significantly boost prediction accuracy.
  • Preserving Rare Variants: The MAF component of the WK is crucial for preserving the contribution of rare but potentially important favorable alleles, which are often underpowered in traditional GWAS.
  • Leveraging GxE Signals: The WK effectively captures the trait- and population-specific environmental effects by differentiating marker contributions across environments, leading to more robust predictions.
  • Flexibility for Heterogeneous Populations: The Bayesian framework and WK approach are particularly valuable for genetically heterogeneous populations (like half-sib families in spruce) where traditional methods struggle due to complex genetic backgrounds and lack of uniform material.

Key Results

  • Winter Wheat: The multi-environment weighted kernel (ME_WKMAF_Pvalue) achieved an overall gain in predictability of up to 31% for the winter wheat DH population compared to single-environment models. Significant improvements of 4-33% in prediction accuracy were observed for half-sib families.
  • Interior Spruce: While ME_GK showed improvements over SE_GK, the advantage of WK models was more pronounced for certain traits like DBH. The ME_WKMAF_Pvalue model generally outperformed other WK implementations across most traits.
  • Specific Trait Gains: For winter wheat, ME_WKMAF_Pvalue showed a 31% higher prediction accuracy than ME_GK for GY in 2016. For SDS, a substantial increase of 35% in prediction accuracy was observed with ME_GK over SE_GK.
  • Heritability: The study noted that traits with higher heritability generally showed better prediction performance.

Limitations and Future Work

  • Spruce Prediction Challenges: The advantage of ME_GK over SE_GK was less evident for Interior spruce, potentially due to a large amount of variance not accounted for in ME models and lower multi-site heritability estimates. The authors suspect that SNP predictors for spruce might be in weak linkage disequilibrium (LD) with underlying genes and QTLs.
  • Indifferent Performance for Some Traits: For some traits, like WDx-ray in spruce, the benefit of incorporating MAF and association signals was not observed, possibly due to indifferent prediction performance across all models.
  • Future Work: The authors suggest that their flexible Bayesian kernel methodology is the first GS framework capable of predicting the performance of highly genetically heterogeneous populations across environments without requiring genetically clonal material. This opens avenues for more accurate and efficient breeding programs for a wider range of species and conditions.

Summary updated

Abstract

Abstract Crucial to variety improvement programs is the reliable and accurate prediction of genotype’s performance across environments. However, due to the impactful presence of genotype by environment (G×E) interaction that dictates how changes in expression and function of genes influence target traits, prediction performance of genomic selection (GS) using single-environment models often falls short. Furthermore, despite the successes of genome-wide association studies (GWAS), the genetic insights derived from genome-to-phenome mapping have not yet been incorporated in predictive analytics, making GS models that use Gaussian kernel primarily an estimator of genomic similarity, instead of the underlying genetics characteristics of the populations. Here, we developed a GS framework that, in addition to capturing the overall genomic relationship, can capitalize on the signal of genetic associations of the phenotypic variation as well as the genetic characteristics of the populations. The capacity of predicting the performance of populations across environments was demonstrated by an overall gain in predictability up to 31% for the winter wheat DH population. Compared to Gaussian kernels, we showed that our multienvironment weighted kernels could better leverage the significance of genetic associations and yielded a marked improvement of 4-33% in prediction accuracy for half-sib families. Furthermore, the flexibility incorporated in our Bayesian implementation provides the generalizable capacity required for predicting multiple highly genetic heterogeneous populations across environments, allowing reliable GS for genetic improvement programs that have no access to genetically uniform material.
BibTeX
@misc{hu2022weighted,
  author = {Xiaowei Hu and Brett F. Carver and Yousry A. El‐Kassaby and Zhu Lan and Charles Chen},
  title = {{Weighted Kernels Improve Multi-Environment Genomic Prediction}},
  year = {2022},
  doi = {10.1101/2022.04.10.487783},
  howpublished = {bioRxiv},
}
Figure