Abstract
Genetic prediction for complex traits is usually based on models including individual (infinitesimal) or marker effects. Here, we concentrate on models including both the individual and the marker effects. In particular, we develop a "Mendelian segregation" model combining infinitesimal effects for base individuals and realized Mendelian sampling in descendants described by the available DNA data. The model is illustrated with an example and the analyses of a public simulated data file. Further, the potential contribution of such models is assessed by simulation. Accuracy, measured as the correlation between true (simulated) and predicted genetic values, was similar for all models compared under different genetic backgrounds. As expected, the segregation model is worthwhile when markers capture a low fraction of total genetic variance.
Introduction
In quantitative genetics, predicting complex traits like crop yield or disease resistance typically relies on two distinct methodologies. The Individual (Infinitesimal) model utilizes pedigrees to estimate genetic values based on the average theoretical relatedness of relatives. Conversely, the Marker model (Genomic Selection) bypasses pedigree information to look directly at DNA variations, such as Single Nucleotide Polymorphisms (SNPs).
Neither approach is perfect. Marker models often suffer from "hidden heritability," failing to capture genetic variance when markers are not sufficiently close to relevant genes. Pedigree models rely on expected averages—assuming siblings share exactly 50% of their genes—ignoring the random "genetic lottery" (Mendelian sampling) that occurs during meiosis. Simply combining these models can lead to mathematical redundancy. This paper introduces a "Mendelian segregation" (MS) model designed to integrate these effects by using DNA markers to track actual realized inheritance.
The Mendelian Segregation (MS) Model
The core innovation of the MS model is the decomposition of an individual’s genetic value into two distinct parts: infinitesimal genetic contributions from the base population (ancestors) and realized Mendelian sampling in descendants. Instead of treating the random shuffle of genes from parent to offspring as an unknown error term, the MS model uses DNA marker data to estimate exactly how much of a parent's genome was passed down.
Technically, the model utilizes a Laplacian matrix ($M$) to represent the pedigree graph. While traditional models use a theoretical relationship matrix ($A$), the MS model replaces unknown sampling terms with a predictor ($s$) derived from observed marker genotypes ($W$). This allows the model to account for family structure and DNA-level inheritance simultaneously without "double-counting" genetic effects. As illustrated in the figure above, the total genetic value $u$ is derived by combining base effects $u_b$ and the realized sampling $s$ through the inverse matrix $M^{-1}$.
Evaluation and Results
The researchers evaluated the MS model using custom simulations (QMSim) and the QTLMAS 2008 public benchmark dataset. The simulations adjusted SNP density and the proportion of variance explained by markers to test the model's robustness.
Key results include:
- Superiority in Low-Information Scenarios: The MS model outperformed standard marker models by 12% in accuracy when markers only captured a small fraction (10%) of the total genetic variance. This suggests the model acts as a safety net when marker density is low.
- Benchmark Success: On the public QTLMAS dataset, the MS model achieved a correlation of 0.94 between predicted and true genetic values, significantly beating the standard Marker model (0.87) and the traditional Marker + Individual hybrid model (0.84).
- Refined Marker Estimates: The MS model was more effective at estimating the actual effects of individual markers ($r = 0.69$) compared to other models ($r = 0.56$), indicating better separation of true genetic signals from noise.
Discussion and Conclusion
The MS model provides a formal framework for moving from expected relatedness to realized relatedness. By using markers to find the specific differences in DNA sharing between relatives, the model picks up the "slack" when genomic data alone is insufficient.
However, real-world application faces hurdles. The current derivation assumes complete genotyping across the pedigree, which often requires complex data imputation in practice. Additionally, the model currently assumes genetic effects follow a normal distribution; incorporating non-normal distributions could improve performance for traits controlled by major genes. Future research will focus on optimizing the computational complexity of the required matrix inversions for large-scale breeding programs.
Summary updated