GPFN: Prior-Data Fitted Networks for Genomic Prediction

Abstract

Genomic Prediction (GP) methods predict the breeding value of unphenotyped individuals in order to select parental candidates in breeding populations. Among models for GP, classical linear models have remained consistently popular, while more complex nonlinear methods such as deep neural networks have shown comparable accuracy at best. In this work we propose the Genomic Prior-Data Fitted Network (GPFN), a new paradigm for GP. GPFNs perform amortized Bayesian inference by drawing hundreds of thousands or millions of synthetic breeding populations during the prior fitting phase. This allows GPFNs to be deployed without requiring any training or tuning, providing predictions in a single inference pass. On three populations of crop plants across two different crop species, GPFNs perform significantly better than the linear baseline on 13 out of 16 traits. On a challenging between-families NAM prediction task, the GPFN performs significantly better in 3 locations while only falling behind in one. GPFNs represent a completely new direction for the field of genomic prediction, and have the potential to unlock levels of selection accuracy not possible with existing methods.