Generative AI for predictive breeding: hopes and caveats.

Area of Science:

• Artificial Intelligence
• Genomics
• Computational Biology

Background:

• Generative AI (genAI) creates realistic synthetic data, offering an alternative to traditional simulation.
• While often focused on phenotypes, genAI can also generate synthetic environments and genotypes.
• GenAI may overcome limitations of standard simulations, such as rigid genotype-phenotype mapping assumptions.

Purpose of the Study:

• To explore the potential of generative AI in creating synthetic environments and genotypes.
• To discuss key features of popular generative models and their application in biological data generation.
• To highlight the benefits of integrating generative AI with conventional simulation for genomic prediction and breeding strategies.

Main Methods:

• Discussion of generative models: autoregressive models, generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, and flow-based models.
• Exploration of latent space properties in generative models for interpretability and linking simulation with AI.
• Conceptualization of hybrid approaches combining conventional simulation with genAI for data augmentation.

Main Results:

• Generative AI can produce synthetic data for environments and genotypes, expanding beyond phenotype generation.
• Latent spaces in generative models offer interpretability and bridge simulation and AI.
• Augmenting genomic prediction models with realistic synthetic data can enhance inference and performance.

Conclusions:

• Generative AI offers a powerful data-driven approach to complement traditional simulation in biological research.
• Hybrid models integrating conventional simulation and generative AI are promising for in silico evaluation of predictive breeding strategies.
• Future directions include simulating novel genotypes and using generative models for phenotype prediction conditional on genotype and environment.