BACKGROUND

Breeding programs need to improve multiple characteristics of their crops simultaneously, including yield, disease resistance, and quality. However, when traits are genetically correlated, selection on one trait leads to indirect selection on other traits. Classic quantitative genetics theory provides a solution, defining theoretically optimum weights of two (or more) traits for simultaneously improving both. However, actual breeding programs differ from the simplifying assumptions underpinning these theories in several critical ways:

• Genetic parameters are not directly observable and are often estimated with considerable uncertainty
• Some traits can be measured more precisely, cheaper, or at different growth stages than other traits
• Responses to selection are proportional to the available genetic variation in the breeding population
• Correlations between traits can dramatically hinder or improve genetic gain.

A holistic, simulation-based approach to breeding program optimization could address all of these challenges, and provide precise and explainable recommendations to breeders that take into account the specifics of each individual program. A simulation framework can help 1) quantify and visualize the overall range of possible outcomes of different selection strategies across multiple correlated traits given the uncertainty in genetic value estimation, Mendelian assortment, and economic value; 2) select lines to cross that optimize short-term genetic gains and future genetic diversity, and 3) allocate phenotyping resources including training population selection to optimize genetic value prediction accuracy for target traits and lines.

GOALS

• Develop a simulation tool to rapidly simulate outcomes of breeding decision frameworks over multiple generations and calibrate the simulations around the current state of the UC Davis strawberry and pepper breeding programs
• Quantify the probability of discovering specific genotypes
• Optimize breeding program designs for both of the previous goals using the principles of Sequential Design of Experiments, including which lines to cross (using which selection criteria), which to phenotype, and how to allocate limited resources between these objectives

IMPACT

• Breeding programs are critical to a sustainable and high quality food system.
• In silico simulation can provide recommendations to breeders on how to allocate resources across the multiple components of a breeding program, making the process more efficient, faster, and more robust.
• While details of every breeding program in every crop are unique, the processes and decision points of breeding programs are universal and AI tools we create here will be useful for any breeding program.