Trait importance in predicting Glycine max seed viability decline: A machine learning approach

Predicting seed viability decline is crucial for both farmers and seed conservators. For seed conservation institutions such as Genebanks, the high cost of germination tests poses a significant challenge. Reducing these costs through accurate prediction of time to viability loss is therefore desirable. This study aimed to identify traits that enhance the accuracy of prediction models for soybean seed viability. Analysis of germination test data accumulated by the NARO Genebank using a Gradient Boosting Machine (GBM) to account for non-linear relationships and complex interactions, revealed that seed size was the most important trait, followed by the number of main stem nodes. In contrast, seed coat colour exhibited a weak correlation with viability decline. The prediction model incorporating seed size, number of main stem nodes, and initial germination rate achieved a C-index of 0.65, compared with 0.58 for the model based solely on initial germination rate ( p = 0.0014). Given the ease of data acquisition, seed size is a useful explanatory variable in predicting decline of soybean seed viability. Further improvement in prediction accuracy is expected to be achieved by integrating post-storage germination data into the current model, which primarily uses seed size, number of main stem nodes and initial germination.