By 2050, the world will need to feed about ten billion people, requiring global agricultural production to increase by about 60 percent in the next 30 years. To tackle this challenge, plant breeders are developing new plant varieties with improved characteristics.

Breeding cycles are very complex, expensive and time-consuming. Each breeding cycle offers a huge range of possible crossings. In our use case, a breeder in a midsize seed company has 800 ⨉ 800 / 2 = 320,000 double-haploid maize crosses (hybrids) that could be made, but has space to evaluate only 2,500 hybrids in the field. Ranked by yield, only the top 50 ⨉ 50 plants would be crossed. Therefore, selecting the actual, best-performing hybrids is critical.

Beck’s wanted to evaluate the genetic potential of crosses, and assess the best performing ones to develop a new maize variety with higher yield. To achieve this goal, we built a model to predict all possible hybrid crosses (virtual hybrids) between double-haploid parents of a male and a female population with known and unknown phenotypes.

Beck’s has chosen Computomics’ Machine Learning-based solution xSeedScore to obtain the most accurate performance predictions for their breeding program for maize.

For each season, Beck’s sends Computomics genotypes and phenotypes. We predict the performance of all possible hybrids and uncover high-yielding combinations among the risky crosses that would never have been tried, because they are outside of the top-crossing scheme. xSeedScore identifies crosses which would not have been considered by standard statistical methods like BLUP.

Results show that our algorithms are able to double the prediction accuracy of the genetic potential of crops compared to state-of the-art methods, which means breeders and growers can make decisions with more confidence.

xSeedScore predicted phenotypic values for each hybrid using our proprietary kernel-based machine learning method, which is trained with the known parental double-haploid genotypes and their phenotype information of the hybrids. We evaluated and optimized the model using a cross-validation approach before predicting the offspring phenotypes.

Compared to BLUP models, the Computomics method takes complementing markers into account, so that hybrids receive predicted phenotypes that exceed those of their parents.

All results are visualized for easier and faster interpretation.

The Beck's team tested the crosses we recommended in the field and the results were exceedingly good – year after year. xSeedScore can deliver twice the prediction accuracy compared to standard statistical methods. Accurate prediction of all virtual crosses and identification and deselection of low performers before planting them significantly decreased testing time and considerably increased the potential to identify high performing crosses. This in total results in double genetic gain, shorter breeding cycle and increased efficiencies.