BayesR3AD: Joint analysis of additive and dominance in Bayesian mixture models

⚕️

This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a farmer trying to predict how well your cows will perform in the future. Will they have healthy calves? Will they live long, productive lives? For decades, scientists have used a "recipe" to make these predictions. This recipe mostly looks at additive effects—think of this as adding up the good traits from the mother and the father, like mixing two colors of paint to get a new shade. If Mom is tall and Dad is tall, the calf is likely to be tall.

However, biology is messy. Sometimes, the combination of genes creates a surprise that isn't just the sum of the parts. This is called dominance. It's like baking a cake: if you add a little bit of vanilla, it tastes like vanilla. But if you mix vanilla and chocolate in a specific way, you get a flavor that neither ingredient has on its own. Or, sometimes, mixing two specific ingredients creates a bitter taste (a "recessive" bad trait) that only shows up when you have two copies of that specific ingredient.

For a long time, the scientific "recipe" (called BayesR3) ignored these mixing surprises. It assumed everything was just a simple sum. This worked okay for milk production, but for tricky traits like fertility (having babies) and survival (living long), it missed a lot of the magic.

The New Recipe: BayesR3AD

The authors of this paper created a new, upgraded recipe called BayesR3AD.

Think of the old recipe as a chef who only knows how to add ingredients. The new chef (BayesR3AD) knows how to add ingredients and how they interact (dominance) when mixed together.

Here is how the new method works, using a few simple analogies:

1. The "Smart Filter" (The Mixture Model)
Imagine you have a giant bag of 75,000 different spices (these are the genetic markers, or SNPs). Most of them do nothing at all. A few add a tiny pinch of flavor. A very few add a huge burst of flavor.

The old method tried to guess the flavor of every single spice.
The new method uses a Smart Filter. It looks at the data and asks: "Is this spice actually doing anything?"
If a spice (gene) has no effect, the filter says, "Ignore it."
If a spice has a big effect, the filter says, "Pay attention!"
Crucially, BayesR3AD has two filters: one for the "adding" effect and one for the "mixing" effect. If the mixing effect doesn't exist for a certain spice, the filter automatically turns it off, so it doesn't mess up the prediction.

2. The "Safety Net" (Robustness)
You might worry: "If I add a new ingredient to the recipe, won't it ruin the cake if I don't actually need it?"
The authors tested this by creating fake cow data.

Scenario A (No Mixing): They created cows where only the "adding" effect mattered. The new recipe (BayesR3AD) looked at the data, realized the "mixing" spice wasn't needed, and turned it off. It predicted just as well as the old recipe.
Scenario B (Mixing Exists): They created cows where the "mixing" effect was huge. The old recipe got confused, thinking the weird results were just random noise. The new recipe said, "Ah, I see the mixing effect!" and adjusted the prediction.
The Result: When mixing effects were present, the new recipe was 20% more accurate at predicting the future. That is a massive improvement in the world of breeding.

3. Finding the "Hidden Gems" (Real Data Results)
The team tested this on real Holstein cows (a popular dairy breed) looking at two traits: Calving Interval (how long it takes between having calves) and Survival (how long the cow lives).

The Big Discovery: They found a specific spot on the cow's genetic map (Chromosome 18) that acts like a "switch."
- For Calving Interval, they found a spot where having two copies of a gene made the cow less fertile, but having one copy (a mix) made her super fertile. This is a classic "heterozygote advantage"—like having a hybrid car that gets better mileage than either a gas or electric car alone.
- For Survival, they found another spot where the "mixing" effect helped the cow live longer.

Why Does This Matter?

In the past, if a farmer wanted to breed cows for better fertility, they might have missed out on the best cows because the old math couldn't see the "mixing" benefits.

Better Predictions: Farmers can now predict the total genetic worth of an animal more accurately, not just the "additive" part.
Avoiding Bad Breeds: It helps identify "recessive" bad genes (the bitter taste in the cake) that might hide in the population and only show up when two carriers are bred together.
No Downside: The best part is that this new method is safe. If a trait is purely about "adding" (like milk yield), the new method doesn't get confused; it just acts like the old method. But if a trait involves "mixing" (like fertility), it shines.

The Bottom Line

The authors built a smarter, more flexible tool for understanding cow genetics. It's like upgrading from a black-and-white TV to a high-definition color TV. You can still watch the black-and-white shows (additive traits), but now you can also see the vibrant, complex colors (dominance traits) that were previously invisible. This helps breeders make better decisions to create healthier, more fertile, and longer-lived herds.

1. Problem Statement

Genomic prediction in livestock breeding has historically relied on additive models (e.g., GBLUP or standard Bayesian regression like BayesR). While effective for many traits, these models ignore non-additive genetic effects, specifically dominance, which are known to significantly influence fitness-related traits such as fertility and survival.

Limitations of Current Methods: Ignoring dominance can bias additive variance estimates, misattribute non-additive signals to residual error, and reduce the accuracy of total genetic merit predictions.
Gap in Literature: While some methods (like GBLUP extensions) incorporate dominance via relationship matrices, there is a lack of robust Bayesian mixture models that explicitly estimate sparse, locus-specific SNP effects for both additive and dominance components simultaneously. Existing Bayesian mixture models (like BayesR) typically handle only additive effects.

2. Methodology: BayesR3AD

The authors developed BayesR3AD, an extension of the existing BayesR3 algorithm, to jointly model additive and dominance effects within a unified Bayesian mixture framework.

A. Statistical Model

The model uses a linear mixed model:
$y = \mu + Va + Td + e$
Where:

$y$ : Vector of pre-corrected phenotypes.
$V$ : Coded additive genotype matrix.
$T$ : Coded dominance genotype matrix.
$a, d$ : Vectors of additive and dominance SNP effects, respectively.
$e$ : Residual errors.

Genotype Coding:
Genotypes are transformed into additive ( $v_{ij}$ ) and dominance ( $t_{ij}$ ) covariates, standardized to mean 0 and unit variance. This ensures approximate orthogonality between additive and dominance components under Hardy-Weinberg equilibrium.

B. Mixture Priors

BayesR3AD employs a finite mixture of normal distributions for both additive ( $a_j$ ) and dominance ( $d_j$ ) effects:

Four Components: Each SNP effect is assigned to one of four distributions ( $k=1,2,3,4$ $k = 1, 2, 3, 4$ ) with variances $\sigma^2_k = \gamma_k \sigma^2_{global}$ $σ_{k}^{2} = γ_{k} σ_{g l o ba l}^{2}$ .
- $k=1$ : Spike at zero ( $\gamma_1 = 0$ ).
- $k=2,3,4$ : Increasing non-zero variances ( $\gamma_k \in \{10^{-4}, 10^{-3}, 10^{-2}\}$ ).
Separate Indicators: Additive and dominance effects have independent latent mixture indicators ( $z^{(a)}_j$ and $z^{(d)}_j$ ). This allows a SNP to have a non-zero additive effect but a zero dominance effect (and vice versa), or belong to different variance classes.
Mixture Proportions ( $\pi$ ): The proportions of SNPs in each variance class are estimated adaptively using Dirichlet priors during the MCMC process.

C. Sampling Algorithm (Blocked Gibbs Sampler)

The model uses a blocked Gibbs sampler to ensure efficient mixing:

Blocking: SNPs are partitioned into non-overlapping blocks.
Sequential Updates: Within each block, additive and dominance effects are updated one at a time, conditioning on the current residuals and other effects.
Conditional Updates: For each SNP, the algorithm calculates the posterior probability of belonging to each mixture component based on the conditional least-squares estimate of the effect. The effect is then sampled from the corresponding normal distribution (or set to zero if the null component is selected).
Variance Components: Global additive and dominance variances ( $\sigma^2_a, \sigma^2_d$ ) and mixture proportions are updated at the end of each outer iteration.

3. Key Contributions

Algorithmic Extension: Successfully extended the computationally efficient BayesR3 framework to handle joint additive-dominance modeling without compromising the ability to perform fine-mapping.
Adaptive Shrinkage: The model exhibits robust "self-regularizing" behavior. When dominance effects are absent, the model automatically shrinks dominance variance components toward zero, reverting effectively to an additive-only model.
Locus-Specific Inference: Unlike GBLUP-based dominance models that distribute variance genome-wide, BayesR3AD provides sparse, locus-specific estimates of dominance effects, enabling the identification of specific heterozygote advantages or recessive deleterious alleles.
Inbreeding Depression Estimation: The framework allows for the direct calculation of inbreeding depression (ID) from estimated dominance effects.

4. Results

A. Simulation Studies

Data: 227,942 Holstein genotypes (74,626 SNPs) with simulated phenotypes.
Scenario 1 (Additive Only): BayesR3AD correctly estimated near-zero dominance variance ( $\approx 0$ ) and maintained additive prediction accuracy identical to the additive-only BayesR3 model.
Scenario 2 (Additive + Dominance):
- Variance Recovery: BayesR3AD accurately recovered true dominance variance ( $\hat{\sigma}^2_d \approx 0.44$ vs. true $0.44$). In contrast, the additive-only model misattributed dominance variance to the residual, inflating residual variance and underestimating heritability by ~12%.
- Prediction Accuracy: BayesR3AD improved prediction accuracy by +0.1011 (approx. 19.7% relative increase) compared to BayesR3 when dominance was present.
- Fine Mapping: The model successfully identified causal loci for both additive and dominance effects, with high concordance between true and estimated top SNPs.

B. Real Data Analysis (Holstein Cattle)

Traits: Calving Interval (CI) and Direct Survival.
Variance Estimates:
- Dominance variance was small but detectable: ~1.4% of genetic variance for CI and ~3.0% for Survival.
- Inbreeding depression was estimated as positive for CI (longer intervals with inbreeding) and negative for Survival (reduced survival with inbreeding), consistent with biological expectations.
Genomic Findings:
- BTA18 (Chromosome 18): Identified a major additive QTL at 57.82 Mb (consistent with previous GWAS) and a significant dominance QTL at 44.37 Mb for Calving Interval. The dominance signal at 44.37 Mb suggests a heterozygote advantage (negative dominance effect implies homozygotes have longer calving intervals).
- Survival: Detected a dominance signal near 43.15 Mb on BTA18, near the RGS9BP gene.
Prediction: While dominance variance was low in real data, BayesR3AD showed marginal improvements in prediction accuracy over BayesR3 (e.g., 0.245 vs 0.243 for CI) without destabilizing additive estimates.

5. Significance and Conclusion

Robustness: BayesR3AD is a "safe" extension for breeding programs. It delivers significant gains when dominance is a major factor (simulated data) but imposes no penalty (and maintains accuracy) when dominance is negligible (real fertility data).
Biological Insight: The ability to pinpoint specific loci with dominance effects (e.g., recessive deleterious mutations or heterozygote advantages) provides actionable insights for managing genetic diversity and avoiding inbreeding depression.
Applicability: While validated on Holstein cattle, the method is directly applicable to other species, particularly plants and animals where non-additive effects play a larger role in fitness and yield traits.
Computational Efficiency: The method is computationally tractable, requiring only a ~40-70% increase in runtime and ~2x memory compared to the additive-only model, making it feasible for large-scale genomic evaluations.

In summary, BayesR3AD represents a practical and powerful advancement in genomic prediction, bridging the gap between sparse Bayesian modeling and the complex reality of non-additive genetic architectures.