Random forest regression to predict Farinograph traits from GlutoPeak output in wheat wild relative backcross lines

Authors: Price, John; Guttieri, Mary; Stutz, Sydney

Submitted to: Cereal Chemistry
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/30/2024
Publication Date: N/A
Citation: N/A

Interpretive Summary: The functional quality of winter wheat for bread making varies widely among varieties. Therefore a key goal for wheat breeders is to ensure that a given breeding line can produce flour with satisfactory quality. Within winter wheat, the dough mixing properties are of paramount importance. The primary test for flour quality in the United States is the Farinograph, which requires large quantities of flour and can take 30 minutes to complete a single test. This limits the number of breeding lines that breeders can test, which slows breeding progress. A new device, called the GlutoPeak, became available in 2015. The Glutopeak requires very small sample sizes and runs very quickly. Developing methods to relate the output of the GlutoPeak test to those of the Farinograph will improve the utility of this high-throughput device as a tool for breeding. Here, we use a machine learning approach to develop models to predict key Farinograph parameters from GlutoPeak measurements. The breeding lines result from crossing bread wheat with a wild ancestor, and so contain high genetic diversity for quality, which could be difficult to predict. The machine learning models predicted key Farinograph traits with moderate accuracy and would be sufficient to differentiate between high and low-quality lines. These machine learning models, and the machine learning approach, will be useful breeders in making better selection decisions early in the breeding process.

Technical Abstract: Background and Objectives: Flour quality is a key target of hard winter wheat breeding. The Farinograph is important for assessing quality prior to cultivar release in the United States, but large sample size requirements and long test times render it impractical for early-stage selection relative to the GlutoPeak. To improve GlutoPeak utility for breeding, we calculated new parameters from the GlutoPeak raw data and used random forest regression to predict key Farinograph parameters in a winter wheat population containing wheat wild relative introgressions. Findings: The key quality parameters of absorption, bake absorption, tolerance stability, and mixing tolerance index were moderately well predicted (R2 ranging from 0.488-0.745). Classification of samples as acceptable or unacceptable for mixing tolerance index and tolerance stability was more accurate than prediction of numeric values. Conclusions: New features calculated from the GlutoPeak raw data were useful predictors of quality. Prediction accuracies are sufficient to improve breeding populations. Significance and Novelty: This study is the first to use wheat wild relative introgressions in GlutoPeak Farinograph prediction. It is the first to generate features from raw data and is one of the few random forest models for quality prediction. Improved ability to cull poor quality lines early in the breeding pipeline can support efficient wheat cultivar development.