Michael Burns
University of Minnesota
Food-grade corn is an economically important crop in the United States and worldwide. However, in the United States, it makes up less than 2% of the corn grown by acre. This has led to reduced resources for improving food-grade corn, which affects the manufacturing process and downstream consumer satisfaction with final products. One trait that impacts many other quality traits is the amount of water absorbed during the nixtamalization process. Methods have been proposed for measuring this multi-impact trait, but either require large amounts of sample or were designed for inbreds. This work has developed a low-quantity prediction model specific to hybrids. It uses said model to assess the effects of grain traits (composition, yield), and parental combinations on a hybrid's nixtamalization moisture content.
For more information, please see the manuscript at: ADD DOI HERE
This project has developed a machine learning model to predict nixtamalization moisture content from NIR spectra of ground, raw maize kernels. This requires samples to undergo a grinding (destructive) and NIR scanning process and a benchtop cooking process (destructive) to collect the predictive features (NIR spectra) and predicted trait (nixtamalization moisture content). 300 samples were had 50g of sample allocated for grinding and scanning (requires at least 30g) and 200g of sample allocated for cook tests (2 reps of 100g). Samples were ground and scanned first to provide spectra for selecting diverse samples for training.
After samples were selected, initial cook tests were performed to train a preliminary model, which was used to select more samples to broaden the spectral diversity of the training set. The preliminary model (and subsequent models) were trained using supercomputing resources to determine the best combination of data preprocessing, model, hyperparameters, etc. This process was repeated 100 times to get an estimate for performance variation and allowed us to pick the best model combination available for the given dataset. After the best model for the final training set was chosen, the model was used to predict on the larger dataset of samples available.
Using these samples, we assessed the relationship between hybrids (using the model from this manuscript) and their inbred parents (using the model from Burns et al. 2021), environmental influences, and the ability to create genomic prediction models the can predict hybrid performance based on computational derivation of hybrid genomic markers.
- scripts:
- scripts used in the analysis for the publication
- data:
- data collected from cook tests and NIR scanning, as well as environmental and genetic information
- supplemental_materials:
- code developed largely as proof of concept that is added for historical accuracy of the analysis
- Moisture_Content_Predictor:
- A free to use, locally running RShiny application that can be implemented in various stages of corn breeding