Seed quality drives grain yield in Ethiopian and Senegalese sorghum: Insights from machine learning

Abstract

Accurately predicting grain yield remains a major challenge in sorghum breeding, particularly across genetically and geographically diverse germplasm. To address this, we applied a phenotype-informed machine learning (PIML) framework to analyze nine phenotypic traits in 179 Ethiopian and Senegalese accessions. Using hierarchical clustering and oversampling with ADASYN, we achieved high classification accuracy (0.99) for phenotypic group assignment. Grain yield prediction was most effective with a Neural Boosted model (NTanH(3)NBoost(8)), achieving a mean R2 of 0.36 and RASE (equivalent to RMSE) of 4.87. Feature importance analysis consistently identified seed weight and germination rate as the strongest predictors of grain yield, while disease resistance traits showed limited predictive value. These findings suggest that early selection based on seed quality traits may provide a practical strategy for improving sorghum yield under field conditions, especially in resource-limited environments.