The latest version of ASReml, Release 4.3, introduces several computational and modelling enhancements relevant to the analysis of multi-environment trial (MET) and genomic data. These developments were driven by the growing demand to fit models with larger datasets — more environments, more genotypes, and increasingly large dense genomic relationship matrices (GRMs). Key highlights include eigen-model transformation for large genomic animal models, genetic model trimming for MET analyses, construction of dominance and epistatic matrices from marker data, formation of the H (hybrid) matrix for single-step GBLUP, and extended tools for constraining variance parameters. The key enhancements are summarised below, with full technical details available in the updated user documentation.
Faster Factor Analytic Models in ASReml 4.3
Factor analytic (FA) models are widely used in MET analyses to model complex genotype-by-environment covariance structures. ASReml 4.3 introduces the new rrdk() (reduced rank + diagonal) structure for factor analytic models, which decomposes the extended FA model into two independent terms: rrk() and diag(). This reparameterisation of Σ = ΓΓ’ + Ψ, while mathematically equivalent to xfak(), leads to sparser mixed model equations and delivers significantly faster computation, particularly when interacting with large genomic relationship matrices (GRMs) in multi-environment trial analyses.
Environment-Specific GRM Trimming for Faster MET Analyses
In MET analyses, it is common for different environments to present different subsets of genotypes. ASReml 4.3 introduces genetic model trimming for multi-environment trial (MET) analyses via the sat() structure. Rather than fitting a single dense GRM across all environments, trimming creates environment-specific reduced GRM matrices containing only the genotypes present at each site. This yields identical variance parameter estimates to the untrimmed model, but with dramatically lower computing times (in one example, reducing iteration time from 116.7 seconds to just 33.3 seconds).
Eigen-Transformation of the GRM for Computationally Efficient Multi-Trait Genomic Models
Fitting large genomic animal models across many traits is computationally intensive, particularly when a dense GRM dominates the mixed model equations. ASReml 4.3 addresses this with an Eigen-transformation, which applies a singular value decomposition (SVD) of the GRM to convert the genomic component of the coefficient matrix to a diagonal structure, while leaving variance parameter estimates unchanged. The result is substantially faster computation. In one example, a conventional bivariate analysis taking 8,313 seconds was reduced to just 813 seconds using this eigen-transformation.
Unifying Pedigree and Genomic Data: H Matrix Construction and Fitting for ssGBLUP
It is common in many breeding programmes that a fraction of individuals have genomic information, while the remainder have reliable pedigree records. Proper analyses of these data require a relationship matrix that integrates both sources of information. ASReml 4.3 introduces the !HINV qualifier to construct the H (hybrid) matrix inverse, by merging the inverses of the numerator relationship matrix with the genomic relationship matrix. This approach enables ASReml to fit single-step GBLUP (ssGBLUP) analyses in a single unified model within ASReml.
Dominance and Epistatic Relationship Matrices for Partitioning Genetic Components
Standard genomic analyses often rely on estimating additive effects, but understanding the full genetic architecture often requires estimation of non-additive effects. ASReml 4.3 introduces the !DOMINANCE and !EPISTATIC qualifiers, which construct dominance (GD) and epistatic (GAA, GAD, GDD) relationship matrices directly from SNP marker data. This enables the fitting of more complete genetic models that explicitly partition the total genetic variance into additive, dominance, and epistatic components.
Generating and Saving the Mixed Model Coefficient Matrix and its Inverse
The coefficient matrix C of the mixed model equations contains rich information about model effects, relevant for additional and enhancing statistical inference. ASReml 4.3 introduces a set of new qualifiers to generate and save the C matrix and its inverse directly. All matrices are stored in sparse format, with only known non-zero cells written, making them practical to use even for large models.
Efficient GRM Management via Binary Storage in ASReml 4.3
Repeatedly reading and recomputing large genomic relationship matrices can be time-consuming in iterative or multiple analyses. ASReml 4.3 introduces binary storage formats for the GRM and its inverse via !SAVEGIV, writing matrices in single or double precision binary files. Once saved, these can be read back directly in subsequent runs, eliminating redundant re-computation.
Flexible Variance Component Constraints
Constraining variance parameters to simple equality is often insufficient for complex models. ASReml 4.3 introduces the VCC directive, which defines formal linear relationships between variance structure parameters. The options for relationships include equality, opposite sign, or custom scaling. This gives analysts precise control over parameter constraints, enabling more flexible, parsimonious and meaningful user’s defined variance models.