Scenario-based machine learning prediction of trait performance and stability in crested wheatgrass under climate change

Objectives: Climate change intensifies drought and salinity in semi-arid ecosystems, increasing demand for predictive, data-integrated tools in climate-resilient breeding. Crested wheatgrass (Agropyron cristatum (L.) Gaertn) is a stress-tolerant perennial grass essential for forage, soil conservation, and rehabilitation. This study combined morphological traits, genetic markers, and environmental variables within a machine learning framework to predict trait performance and genotype stability under climate-change scenarios. Methods: Plant materials of crested wheatgrass were collected from multiple geographic locations and subsequently evaluated under a common garden experimental design to minimize environmental variability and isolate genetic effects. Morphological and genetic data were collected across multiple growing seasons (2021–2025). A two-stage modeling approach was implemented: first, algorithms models were evaluated using Leave-One-Year-Out cross-validation. Second, trait-specific best-performing models were selected to integrate putative markers, edaphic, and climate variables for scenario-based predictions under a moderate and severe climate scenario. Genotype stability was assessed using directional change and coefficient of variation, integrated into a network-based multi-criteria scoring system. Results: Model–trait matching improved predictive reliability. Canopy diameter and growth habit were optimally predicted using K-Nearest Neighbors; however, the other traits were more accurately predicted by Neural Boosted models. In the study, scenario-based clustering revealed that high-performing traits do not necessarily correspond to climate resilience. Seven genotypes exhibited consistent responses across climate scenarios according to the multi-criteria scoring system. Conclusions: These results suggest that aligning algorithm selection with trait characteristics can enhance predictive accuracy, potentially facilitating the development of decision-support tools that aid breeders in data interpretation and field implementation.