Integrating Taguchi Design and Machine Learning Models for Trait Stability and Predictive Modeling of Forage Quality in Grass pea (Lathyrus spp.)

Grass pea ( Lathyrus sativus L.) is a resilient legume traditionally used for both human and animal consumption, valued for its drought tolerance and adaptability to marginal soils. However, its utilization is limited due to the neurotoxin β-ODAP, emphasizing the need for selecting genotypes with improved nutritional profiles and reduced toxicity. This study evaluated the forage quality of four naturally occurring Lathyrus species, including one endemic, in the Rize province of Turkey, using a combination of statistical and machine learning approaches. Results regarding different forage traits of Lathyrus species for two years were analyzed by Taguchi Design of experiment. Results revealed L. pratensis as the most robust genotype for critical traits such as crude ash ratio, crude protein ratio, and K/(Ca+Mg) based on signal-to-noise ratio. Application of machine learning models like Random Forest and Light Gradient Boosting Machine models were also used tp predict forage traits. Results illustrated high predictive accuracy for mineral and digestibility-related traits (R² > 0.97). In general, random forest model was superior than light gradient boosting machine model for fiber traits. However, both models failed to predict dry matter intake. The integration of Taguchi design and ML models exhibited most efficient genotype based on forage traits prediction. This dual-framework approach highlights the potential of combining traditional experimental design with modern artificial intelligence tools to support data-driven breeding, optimize forage quality, and promote sustainable livestock production in diverse environments.