DiffBreed: automatic differentiation enables efficient gradient-based optimization of breeding strategies

Area of Science:

• Agricultural Science
• Computational Biology
• Machine Learning

Background:

• Differentiable programming frameworks (e.g., PyTorch, JAX) have advanced biological modeling by enabling joint parameter optimization.
• Existing agricultural breeding simulators lack differentiability, preventing integration with deep learning systems.
• This limitation hinders the optimization of complex breeding strategies.

Purpose of the Study:

• To introduce DiffBreed, a novel differentiable breeding simulator.
• To enable gradient-based optimization of breeding strategies for maximizing genetic gain.
• To facilitate the integration of breeding simulation into modern deep learning workflows.

Main Methods:

• Implemented DiffBreed as a Python module utilizing automatic differentiation.
• Evaluated DiffBreed's performance through gradient-based optimization of progeny allocation.
• Compared optimized strategies against a non-optimized equal allocation approach.

Main Results:

• DiffBreed successfully calculated gradient information via automatic differentiation.
• Gradient-based optimization refined progeny allocation strategies, leading to superior genetic gains.
• Demonstrated the effectiveness of DiffBreed in improving breeding outcomes.

Conclusions:

• DiffBreed represents a significant advancement in breeding simulation technology.
• The simulator's differentiable nature allows for seamless integration with deep learning.
• DiffBreed is poised to revolutionize breeding optimization in agricultural applications.