Phenotype-driven screening and machine learning integration for developing high-performance Miscanthus genotypes for saline-alkaline soils

Abstract

Cultivation of dedicated bioenergy crops on marginal lands is critical for sustainable development, yet breeding climate-resilient genotypes remains a major bottleneck. This study introduces a powerful, integrated framework for accelerating the selection of high-performance Miscanthus hybrids adapted to saline-alkaline soils. The framework uniquely combines phenotypic classification (scatter vs. cespitose plant types) with explainable machine learning (ML). An Extra Trees Regressor (ETR) model, optimized through hyperparameter tuning, achieved high predictive accuracy for biomass (R² = 0.916) in a diverse hybrid population. Crucially, SHapley Additive exPlanations (SHAP) analysis revealed two distinct, data-driven ideotypes: the optimal cespitose type maximizes tiller number, whereas the optimal scatter type prioritizes individual tiller biomass and stem diameter. SHAP also uncovered critical physiological trade-offs, identifying potential thresholds for key traits beyond which biomass gains diminish. Applying this framework, we screened 216 genotypes and identified five superior lines. Subsequent two-year field trials on both farmland and saline-alkaline soil validated our approach, with genotype S88 demonstrating superior yield and stability under stress. This study not only establishes a scalable strategy for industrial crop breeding but also offers novel insights into the architectural determinants of yield under marginal conditions.

Plant phenotyping relevance

植物形態形質と機械学習を統合した選抜フレームワークが研究の中心で、バイオマス予測、形態型分類、SHAPによる形質解釈、圃場検証まで実施しているため、単なる生物学的実験のルーチン測定ではない。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.