GPFN: Prior-Data Fitted Networks for Genomic Prediction.

Area of Science:

• Agricultural Science
• Genetics
• Bioinformatics

Background:

• Genomic Prediction (GP) is crucial for selecting breeding candidates in livestock and crops.
• Classical linear models are popular for GP, but nonlinear methods like deep neural networks have shown limited advantages.

Purpose of the Study:

• To introduce the Genomic Prior-Data Fitted Network (GPFN) as a new paradigm for genomic prediction.
• To evaluate GPFN performance against established linear models in plant breeding.

Main Methods:

• GPFNs utilize amortized Bayesian inference by simulating large populations.
• This approach allows for immediate deployment without model training or tuning.
• Predictions are generated in a single inference pass.

Main Results:

• GPFNs significantly outperformed the linear baseline on 13 out of 16 traits across three plant populations and two crop species.
• On a complex structured prediction task, GPFNs matched linear model performance, outperforming it in one location.

Conclusions:

• GPFNs represent a significant advancement in genomic prediction methodology.
• This new direction has the potential to substantially increase selection accuracy, particularly in diverse populations.