How can we utilize pedigree information for machine-learning-based genetic evaluations?

Purpose

The aim of this project is to investigate how pedigree information can be represented and utilized in machine-learning models for genetic evaluation.

The project will explore whether alternative representations of pedigrees can help machine-learning models recover genetic information and predict breeding values, particularly when traits are influenced by non-additive genetic effects.

Background

Genetic evaluations are a central component of modern selective breeding programs. They are used to estimate the genetic merit of individuals and predict which animals are most likely to pass desirable traits to the next generation.

Traditional genetic evaluations are generally based on parametric statistical models in which pedigree information is incorporated through expected genetic relationships between individuals. These approaches are well established and particularly suited to additive genetic effects.

Machine-learning methods provide a different modelling framework and may offer opportunities to capture more complex relationships, including non-linear and potentially non-additive genetic effects. However, pedigree information does not naturally occur in the tabular or vector-based format commonly used by machine-learning algorithms.

An important question is therefore how pedigree structure can best be translated into informative features for machine learning. Possible approaches range from relationship matrices and ancestor-based representations to graph-based descriptions of the pedigree.

Project description

The project will primarily use simulated breeding populations, allowing the underlying genetic architecture and true genetic values of individuals to be known.

The student will simulate populations with different pedigree structures and genetic architectures and compare alternative ways of incorporating pedigree information into genetic evaluation models.

Depending on the interests of the student, the project may include:

• Simulating breeding populations and pedigrees under different genetic architectures.
• Constructing different pedigree-derived representations, such as relationship matrices, ancestry-based features or pedigree graphs.
• Training and evaluating machine-learning models using these representations.
• Investigating how well different models recover additive and non-additive genetic effects.
• Exploring the performance of the approaches under different population structures, heritabilities and mating designs.
• Potentially applying selected methods to real breeding datasets.

The exact scope can be adapted depending on the student's interests and programming experience. The project combines quantitative genetics, statistics, simulation and machine learning, with considerable room for methodological experimentation.

Specifications

Suitable for: MSc students in Animal Science (animal breeding track), as well as students in plant or forest breeding, bioinformatics, statistics, data science or related fields.

Some experience with programming in R, Python or Julia and basic statistical knowledge is helpful. Previous experience with machine learning is not required.

Curiosity, an interest in quantitative genetics and modelling, and motivation to learn are more important than extensive prior programming experience.

The student will gain experience with:

• Quantitative genetics.
• Basic simulation of breeding populations.
• Pedigree-based genetic evaluation.
• Machine-learning methods.