Back

Why Accuracy Metrics Fall Short in Comparing Phenomic and Genomic Prediction Models

Wang, F.; Feldmann, M. J.; Runcie, D. E.

2025-01-14 genetics
10.1101/2025.01.09.632209 bioRxiv
Show abstract

Phenomic selection is a new paradigm in plant breeding that uses high-throughput phenotyping technologies and machine learning models to predict traits of new individuals and make selections. This can allow breeders to evaluate more plants in higher throughput more accurately, resulting in faster rates of gain and reduced labor costs. However, phenomic prediction models are frequently benchmarked against genomic prediction models using cross-validation to demonstrate their usefulness to breeders. We argue that this is inappropriate for two reasons: 1) differences in the accuracy statistic measured by cross-validation do not reliably indicate differences in the accuracy parameter of the breeders equation, which we show analytically and through re-analysis of data from three representative phenomic prediction studies, and 2) phenomic and genomic selection tools influence other parameters of the breeders equation, so comparing accuracy, even if done properly, is insufficient to advocate for one approach over the other. We conclude that phenomic selection may be useful, but comparisons of accuracy between genomic prediction and phenomic prediction models are not.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.