REML implementations of kernel-based genomic prediction models for genotype x environment x management interactions

arXiv:2501.06844 · stat.ME · Submitted 2025-01-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "REML implementations of kernel-based genomic prediction models for genotype x environment x management interactions".

Jane: Here's the gist of the paper. REML implementations of kernel-based genomic prediction models for genotype × environment × management interactions show that extending basic kernel models with environment-specific variances and nonlinear kernels increases prediction accuracy in sparse testing scenarios <ref:2501.06844#pg18> The paper demonstrates how these advanced models can be implemented in standard software to allow for direct restricted maximum likelihood (REML) estimation of all parameters <ref:2501.06844#pg20>.

Tom: Modeling G×E×M Interactions The research focuses on multi-environment, multi-management trial data, which involves genotypes evaluated across several locations and years under different management practices <ref:2501.06844#pg3> A common approach in analyzing such data is to consider the phenotypes of genotypes under various combinations of managements and environments as different traits within a classical multivariate quantitative genetics framework <ref:2501.06844#pg4>.

Jane: Complex, three-way genotype by environment by management (G×E×M) interactions typically require more flexible structures than simple covariance structures <ref:2501.06844#pg4>.

Tom: Kernel-Based Model Structures The paper details several kernel-based models used to represent the covariance matrix of the random vector of BLUPs, such as linear and nonlinear kernels <ref:2501.06844#pg6>.

Paper summary: Tom: So, this paper is about these kernel-based genomic prediction models for genotype by environment by management interactions, and what they claim is that extending basic kernel models with environment specific variances and nonlinear kernels increases prediction accuracy in sparse testing scenarios <ref:2501.06844#pg18>.

Jane: It sounds like they're taking standard kernel methods, which are used for genotype by environment interactions, and making them smarter by letting the variance change based on the management practice and the specific environment <ref:2501.06844#pg3>.

Lu: The main thing is showing how these advanced models can be implemented in standard software so you can do direct restricted maximum likelihood estimation of all parameters <ref:2501.06844#pg20>.

Meng: Standard software implementation is good for testing, but what I care about is whether it actually runs fast and scales when we get to real data, especially with those three-way interactions that are so common in plant trials <ref:2501.06844#pg4>.

Lalam: The research focuses on multi-environment, multi-management trial data where genotypes are evaluated across several locations and years under different management practices <ref:2501.06844#pg3>.

Tom: So, they're dealing with these complex three way interactions, genotype by environment by management, which is usually a lot harder than just simple covariance structures <ref:2501.06844#pg4>.

Jane: They approach this by looking at the phenotypes of genotypes under various combinations of managements and environments as different traits within a classical multivariate quantitative genetics framework <ref:2501.06844#pg4>.

Lu: They detail several kernel based models used to represent the covariance matrix of the random vector of BLUPs, including linear and nonlinear kernels <ref:2501.06844#pg6>.

Tom: I see they're looking at additive main effects, which treats all environment by management combinations as perfectly genetically correlated <ref:2501.06844#pg8>.

Jane: They also have order m factor analytic models, which use latent factors to model those genotype by environment by management interactions <ref:2501.06844#pg10>.

Meng: I'm curious about the single variance linear kernel and multiple variance linear kernel, which they use to model reaction norm structures with Hadamard products <ref:2501.06844#pg9>.

Lalam: Then there are single variance Gaussian and multiple variance Gaussian kernels that introduce nonlinearity via the radial basis function or RBF kernel <ref:2501.06844#pg9>.

Tom: The study investigates separating the genotype by environment by management interaction variances into components explained by structured random effects, like kernels or factor analytic structures, and components explained by diagonal lack of fit effects <ref:2501.06844#pg9>.

Jane: The results show that when you use more complex models, such as the multiple variance Gaussian kernel model, the environmental covariables explain a higher percentage of the total genotype by environment by management interaction variance <ref:2501.06844#pg10>.

Lu: And the multiple variance kernel models consistently outperform single-variance versions by a significant margin <ref:2501.06844#pg10>.

Tom: The implementation uses ASReml R version four point two, and they use the own function to specify custom covariance structures for random effects <ref:2501.06844#pg8>.

Jane: This framework lets you estimate the bandwidth parameter of Gaussian kernels directly using REML, which avoids those slow cross validation processes <ref:2501.06844#pg8>.

Paper summary: Meng: I wonder if that direct REML estimation for the bandwidth parameter is actually practical when we have massive datasets where traditional methods are already slowing us down <ref:two thousand five hundred one point zero six eight four four#pg8.

Lalam: They discuss potential extensions like combining kernels with factor analytic models to fit the IFA LMM model and constructing different environmental kernels based on subsets of environmental covariates or growth stages <ref:two thousand five hundred one point zero six eight four four#pg8.

Tom: The conclusion is that modeling management specific effects of environmental covariables, maybe using Gaussian kernels with management specific bandwidths, might be better than just assuming equal slopes for multiple managed treatments <ref:two thousand five hundred one point zero six eight four four#pg1.

Jane: So the big picture here is that this framework provides a straightforward way of integrating phenomics, enviromics, and genomics into one linear mixed model <ref:two thousand five hundred one point zero six eight four four#pg1.

Lu: They also showed that extending basic kernel models that assume a single genetic variance for all environments to models with environment specific variances results in higher prediction accuracies <ref:two thousand five hundred one point zero six eight four four#pg9.

Meng: But they also flagged that not all genotype by environment can be predicted from environmental features, and the explanatory power of those features depends on the management practice following a biologically intuitive pattern <ref:two thousand five hundred one point zero six eight four four#pg9.

Tom: That's interesting because the results revealed that total genetic variance was higher under favorable management conditions, like high nitrogen and irrigated trials, which aligns with what we see in plant breeding experiments <ref:two thousand five hundred one point zero six eight four four#pg9.

Jane: So what this means for someone who just listens to the show is that it's not as straightforward as fitting a regression or reaction norm type model with what are essentially equal slopes for multiple managed treatments <ref:two thousand five hundred one point zero six eight four four#pg9.

Lu: The gains come from using kernels to avoid estimating many covariance parameters by assuming the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:two thousand five hundred one point zero six eight four four#pg8.

Tom: That’s because you’re using those kernels to avoid estimating many covariance parameters by assuming the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:two thousand five hundred one point zero six eight four four#pg8.

Jane: The use of nonlinear Gaussian kernels introduces flexibility through a single parameter often referred to as the bandwidth or scale parameter <ref:two thousand five hundred one point zero six eight four four#pg8.

Meng: It's that the REML estimation proposed here extends prior approaches by allowing the bandwidth of the kernel to be estimated directly using REML, avoiding cross validation <ref:two thousand five hundred one point zero six eight four four#pg8.

Lu: The authors also recognize that combining kernels with a random effect with factor analytic covariance structure would lead to a different parameterization of the IFA LMM model proposed by Tolhurst et al <ref:two thousand five hundred one point zero six eight four four#pg8.

Tom: So they've shown easy to use REML implementations of linear and nonlinear kernel based genomic prediction models for genotype by environment by management interactions in standard mixed model software <ref:two thousand five hundred one point zero six eight four four#pg1.

Jane: The paper shows that extending the basic kernel models that assume a single genetic variance for all environments to models with environment specific variances results in higher prediction accuracies <ref:two thousand five hundred one point zero six eight four four#pg9.

Paper summary: Lu: They also show that models with heterogeneous variances capture more of the genotype by environment by management interaction variance than the single variance versions, especially if combined with a nonlinear kernel <ref:two thousand five hundred one point zero six eight four four#pg9.

Tom: So what we learned is that using kernels to avoid estimating many covariance parameters by assuming that the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:two thousand five hundred one point zero six eight four four#pg8.

Jane: The ability to combine kernels with a random effect with factor analytic covariance structure would lead to a different parameterization of the IFA LMM model proposed by Tolhurst et al <ref:two thousand five hundred one point zero six eight four four#pg8.

Lu: The methodology involves a two stage modeling approach where individual environment by management combinations are first analyzed to obtain the best linear unbiased predictions or BLUPs in y, and then the data can be analyzed in a second stage with linear mixed models <ref:two thousand five hundred one point zero six eight four four#pg8.

Meng: That's a two stage modeling approach where individual environment by management combinations are first analyzed to obtain the best linear unbiased predictions or BLUPs in y, and then the data can be analyzed in a second stage with linear mixed models <ref:two thousand five hundred one point zero six eight four four#pg8.

Tom: And it's the modeling of the covariance matrix of this random vector u that is primarily concerned with <ref:two thousand five hundred one point zero six eight four four#pg8.

Jane: It’s about how we link those records to the nested genotype by environment by management BLUPs in u, and modeling that covariance matrix is what matters most <ref:two thousand five hundred one point zero six eight four four#pg8.

Lu: The authors show that extending basic kernel models that assume a single genetic variance for all environments to models with environment specific variances results in higher prediction accuracies <ref:two thousand five hundred one point zero six eight four four#pg9.

Meng: And they also flag that not all genotype by environment can be predicted from environmental features and that the explanatory power of those features for genotype by environment interaction variance depends on the management practice following a biologically intuitive pattern <ref:two thousand five hundred one point zero six eight four four#pg9.

Tom: It's about how we link those records to the nested genotype by environment by management BLUPs in u, and modeling that covariance matrix is what matters most <ref:two thousand five hundred one point zero six eight four four#pg8.

Jane: So the gains come from using kernels to avoid estimating many covariance parameters by assuming that the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:two thousand five hundred one point zero six eight four four#pg8.

Meng: And the REML estimation proposed in this work extends prior approaches by allowing the bandwidth of the kernel to be estimated directly using REML, avoiding cross validation <ref:two thousand five hundred one point zero six eight four four#pg8.

Tom: That's because you’re using those kernels to avoid estimating many covariance parameters by assuming that the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:two thousand five hundred one point zero six eight four four#pg8.

Jane: The paper provides easy to use REML implementations of linear and nonlinear kernel based genomic prediction models for genotype by environment by management interactions in standard mixed model software <ref:two thousand five hundred one.

Conclusion: Tom: So we've been looking at these kernel-based models for genotype by environment by management interactions, and now we're wrapping up this paper on REML implementations of those models.

Jane: It’s about how they made these advanced genomic prediction models practical to use in standard software for direct maximum likelihood estimation.

Lu: Basically, they took complex kernel structures, like the nonlinear Gaussian kernels, and showed you can estimate all the parameters directly using REML without having to do slow cross-validation.

Meng: That’s a big deal for production because it means we don't have to run massive validation loops just to get model parameters right.

Tom: And what they found is that by making the models more flexible, especially by letting the variance change based on management practice, you get better prediction accuracy in sparse testing scenarios.

Jane: So, if you’re working with plant trials or anything similar where you have genotype and environment factors, this suggests that ignoring those specific interactions can really hurt your results.

Lu: They also pointed out that modeling the management-specific effects of environmental variables, perhaps using Gaussian kernels tuned for different management groups, is often better than just assuming equal slopes across all treatments.

Tom: That makes sense from an engineering standpoint; treating every management treatment exactly the same when it's actually different is a mistake.

Jane: And they also noted that while these models are powerful, you can't predict everything from just the environmental features alone; the explanatory power really depends on which management practice you followed.

Meng: I worry about that dependency; if we can’t even predict it fully, how much does this model actually save us when we’re deploying it?

Lu: The authors showed a two-stage modeling approach where they first get best linear unbiased predictions for each environment by management combination and then use those in a second stage linear mixed model.

Tom: That means they aren't trying to solve the whole G by E by M puzzle all at once; they’re breaking it down into manageable steps.

Jane: It gives us a clear framework for integrating phenomics, enviromics, and genomics into one single mixed model setup.

Lu: And there's potential to combine these kernels with factor analytic structures for the IFA LMM model, which opens up new ways to parameterize how we look at those latent environmental factors.

Tom: We've seen a lot of work on this, but making it accessible via REML is what makes this paper useful for people actually trying to build and test these models right now.

Jane: So the big implication is that we can use these sophisticated tools to better understand how management practices shape the genetic response in different environments.

Lu: And next up, we're going to look at some of the specific kernel types they tested—like those Hadamard products and RBF kernels—to see exactly where that extra power comes from.

Killian A.C. Melsen, Salvador Gezan, Daniel J. Tolhurst, Fred A. van Eeuwijk, Carel F.W. Peeters

Mathematical & Statistical Methods group - Biometris, Wageningen University & Research, VSN International, The Roslin Institute and Royal (Dick) School of Veterinary Science, University of Edinburgh

stat.ME

Submitted: 2025-01-12

Updated: 2026-11-13

Code: https://github.com/KillianMelsen/kernels_2026

Project page: https://biometris.github.io/statgenGWAS/index.html

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 95/100

The gist: REML implementations of kernel-based genomic prediction models for genotype × environment × management interactions show that extending basic kernel models with environment-specific variances and…

Key concepts

Kernel-Based Models
These are mathematical structures used to represent the covariance matrix of genetic predictions (BLUPs). Instead of assuming simple relationships, kernels allow the model to capture complex patterns, such as how genetic effects change across different environments or management practices. They provide a flexible way to model how different factors interact.
G×E×M Interactions
This refers to the complex relationship where a genotype's performance depends not just on its genetics (G) and the environment (E), but also on how it is managed (M). The study uses these interactions to understand how management practices influence genetic potential across various locations and years.
REML Estimation
Restricted Maximum Likelihood estimation is a statistical method used to estimate model parameters. In this context, it allows researchers to directly estimate all parameters of the complex kernel models without needing slow cross-validation processes, making the modeling process more efficient and direct.
Heterogeneous Variances
This means allowing different environments or management practices to have unique levels of variance in their effects. The paper shows that using models with these varying variances, especially combined with nonlinear kernels, captures a larger portion of the total G×E×M interaction variance compared to simpler models assuming equal variances.

Terminology

Summary

REML implementations of kernel-based genomic prediction models for genotype × environment × management interactions show that extending basic kernel models with environment-specific variances and nonlinear kernels increases prediction accuracy in sparse testing scenarios <ref:2501. p. 006844#pg18> The paper demonstrates how these advanced models can be implemented in standard software to allow for direct restricted maximum likelihood (REML) estimation of all parameters <ref:2501. p. 006844#pg20>.

Modeling G×E×M Interactions

The research focuses on multi-environment, multi-management trial data, which involves genotypes evaluated across several locations and years under different management practices <ref:2501.06844#pg3> A common approach in analyzing such data is to consider the phenotypes of genotypes under various combinations of managements and environments as different traits within a classical multivariate quantitative genetics framework <ref:2501. p. 106844#pg4>. Complex, three-way genotype by environment by management (G×E×M) interactions typically require more flexible structures than simple covariance structures <ref:2501. p. 106844#pg4>.

Kernel-Based Model Structures

The paper details several kernel-based models used to represent the covariance matrix of the random vector of BLUPs, such as linear and nonlinear kernels <ref:2501.06844#pg6>. Key models include:

  1. Additive main effects (ADD), which treats all E×M combinations as perfectly genetically correlated <ref:2501. p. 406844#pg8>.

  2. Order m factor analytic (FA-m) models, which use latent factors to model G×E×M interactions <ref:2501. p. 406844#pg10>.

  3. Single variance linear kernel (SV-LK) and Multiple variance linear kernel (MV-LK), which model reaction norm structures using Hadamard products. p. 4

  4. Single variance Gaussian kernel (SV-GK) and Multiple variance Gaussian kernel (MV-GK), which introduce nonlinearity via the radial basis function (RBF) kernel. p. 2

Variance Partitioning and Performance

The study investigates the separation of G×E×M interaction variances into components explained by structured random effects, such as kernels or FA structures, and components explained by diagonal lack of fit (LOF) effects <ref:2501.06844#pg9>. The results show that the environmental covariables explain a higher percentage of the total G×E×M variance when using more complex models, such as the multiple variance Gaussian kernel model. p. 9 Furthermore, the multiple variance kernel models consistently outperform single-variance versions by a significant margin <ref:2501. p. 406844#pg10>.

Implementation and Extensions

The models are implemented using ASReml-R version 4.2, utilizing the own function to specify custom covariance structures for random effects. p. 7 The framework allows for estimating the bandwidth parameter of Gaussian kernels directly using REML, avoiding slow cross-validation processes. p. 2 Potential extensions discussed include combining kernels with factor analytic models to fit the IFA-LMM model and constructing different environmental kernels based on subsets of environmental covariates or growth stages.

Conclusion

The paper concludes that modeling management-specific effects of environmental covariables, possibly using Gaussian kernels with management-specific bandwidths, may be more appropriate than assuming equal slopes for multiple managed treatments. The framework provides a straightforward way of integrating phenomics, enviromics, and genomics in a single LMM. p. 0

Pruning redundant reasoning tokens

The paper shows that extending the basic kernel models that assume a single genetic variance for all environments to models with environment-specific variances results in higher prediction accuracies. The authors have shown that models with heterogeneous variances capture more of the G×E×M interaction variance than the single variance versions, especially if combined with a nonlinear kernel. p. 2

How the training signal is built

The methodology involves a two-stage modeling approach where individual E×M combinations are first analyzed to obtain the BLUEs in y, and then the data can be analyzed in a second stage with LMMs <ref:2501.06844#pg8>. For example, the second stage LMM is of the form y = Xβ + Zu + ϵ <ref:2501. p. 306844#pg8>. The random effect design matrix Z links records to the nested G×E×M BLUPs in u, and it is the modeling of the covariance matrix of this random vector u that is primarily concerned with <ref:2501. p. 306844#pg8>.

Where the gains come from

The gains are derived from using kernels to avoid estimating many covariance parameters by assuming that the covariance matrix of the random vector of BLUPs can be expressed as a Kronecker or Hadamard product of appropriately sized and known kernels <ref:2501.06844#pg8>. The use of nonlinear Gaussian kernels introduces flexibility through a single parameter often referred to as the bandwidth or scale parameter. p. 2 The REML estimation proposed in this work extends prior approaches by allowing the bandwidth of the kernel to be estimated directly using REML, avoiding cross-validation. p. 2

The authors show that not all G×E can be predicted from environmental features and that the explanatory power of these features for G×E interaction variance depends on the management practice following a biologically intuitive pattern. The results reveal that total genetic variance was higher under the favorable management conditions, i. p. 9e., high nitrogen and irrigated trials, which matches what is commonly found in plant breeding experiments. p. 1 The authors also indicate that modeling G×E×M data may not be as straightforward as fitting regression or reaction norm type models with what are essentially equal slopes for multiple managed treatments. p. 0

The implementation uses the own function in ASReml-R to specify custom covariance models associated with random effects in the LMM. This function allows for modeling management and environment factors using either a single variance for each level of management or unique variances for all environment-management combinations. p. 4 The external variance function must have the arguments order and kappa, which take the number of factor levels and parameter estimates at a particular iteration, respectively. p. 2

The ability to combine kernels with a random effect with factor analytic covariance structure would lead to a different parameterization of the IFA-LMM model proposed by Tolhurst et al. This combination offers a relatively straightforward implementation of what is essentially linear or nonlinear regression on known covariables, combined with linear regression on an additional set of latent environmental covariables. p. 5 The authors also recognize that modeling the secondary traits or HTP features in a G×E context can be easily implemented by estimating correlations between traits instead of managements. p. 5

The paper provides easy-to-use REML implementations of linear and nonlinear kernel-based genomic prediction models for G×E×M data in standard mixed model software. The authors have shown that extending the basic kernel models that assume a single genetic variance for all environments to models with environment-specific variances results in higher prediction accuracies. p. 4 p. 4 p. 4 The authors also show that models with heterogeneous variances capture more of the G×E×M interaction variance than the single variance versions, especially if combined with a nonlinear kernel. p. 2 p. 2 p. 2

The paper provides easy-to-use REML implementations of linear and nonlinear kernel-based genomic prediction models for G×E×M data in standard mixed model software. The authors have shown that extending the basic kernel models that assume a single genetic variance for all environments to models with environment-specific variances results in higher prediction accuracies. The authors also show that models with heterogeneous variances capture more of the G×E×M interaction variance than the single variance versions, especially if combined with a nonlinear kernel.

The paper provides easy-to-use REML implementations of linear and nonlinear kernel-based genomic prediction models for G×E×M data in standard mixed model software. The authors have shown that extending the basic kernel models that assume a single genetic variance for all environments to models with environment-specific variances results in higher prediction accuracies. The authors also show that models with heterogeneous variances capture more of the G×E×M interaction variance than the single variance versions, especially if combined with a nonlinear kernel.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can perform restricted maximum likelihood REML estimation of all parameters in linear mixed models using kernel-based approaches for genotype by environment by management interactions (G×E×M).

  2. The improved system can model heterogeneous genetic variances for all environments, which is shown to maximize the amount of genetic variance captured by environmental covariables and perform best in prediction settings.

  3. The system can estimate the bandwidth parameter of Gaussian kernels directly using REML, avoiding the need for slow cross-validation processes, which improves efficiency compared to prior optimization methods.

  4. The system can partition G×E×M interaction variances into structured effects explained by kernels or factor analytic structures and components explained by diagonal lack of fit effects, providing a clear separation of variance sources.

  5. The improved system can model environment-specific rates of shrinkage using multiple variance linear kernel models, which has been shown to consistently outperform the single variance versions by a significant margin.

  6. The system can be extended to combine kernels with factor analytic models, leading to a straightforward implementation of the IFA-LMM model, which separates G×E into a component predictable by environmental covariables and one that cannot be predicted.

  7. The system can incorporate feature selection by constructing different environmental kernels using subsets of environmental covariates, allowing researchers to determine the relevance of the different sets of features for predicting G×E.

Related papers