基于深度森林的水稻表型预测方法OA

[Objective]Rice genome prediction facilitates precision breeding design;however,current approaches face challenges such as high-dimensional genomic data,nonlinearity,and poor interpretability.Therefore,developing a novel algorithmic framework that balances prediction accuracy,generalization capability,and biological interpretability offers a solution to these challenges.[Methods]This study proposes and implements an iterative adaptive deep forest-based rice phenotype prediction method(Iterative Adaptive Deep Forest,IADF).The method consists of four collaboratively functioning modules:a single nucleotide polymorphism(SNP)data processing module that controls SNP data quality and reduces noise.Second,the Deep Forest module utilized its unique cascaded forest structure to effectively capture complex nonlinear gene-gene interactions with-out gradient driving.Third,the Iterative Feature Reweighting Mechanism(IFRM)module served as an intelligent guidance framework,simulating the exploration process through a learning-feedback-screening positive feedback loop,and dynamically focused the model's computational resources on the feature subset with the highest contribution.Finally,the Automated Hy-perparameter Optimization module employed the Optuna framework to systematically and collaboratively optimize the key pa-rameters of the Deep Forest engine and the IFRM,ensuring the optimal performance of the entire integrated model.[Results]The experimental results demonstrated that the IADF model exhibited significantly superior Pearson Correlation Coefficients(PCC)to benchmark models such as LightGBM,XGBoost,and DeepCCR across four key agronomic traits:panicle angle,panicle strength,flag leaf angle,and thousand-grain weight.The IADF model achieved an average prediction accuracy of 0.53,representing an approximate 29%improvement over the second-best model,with the highest prediction PCC reaching 0.70 for the thousand-grain weight trait.Furthermore,root mean square error(RMSE)analysis further validated the model's robustness in reducing prediction bias.Ablation experiments confirmed the necessity of the synergistic collaboration among SNP filtering,the iterative feature weighting mechanism,and the hyperparameter optimization modules.Additionally,parame-ter analysis revealed a nonlinear relationship between the number of input features and model performance,with optimal perfor-mance achieved when retaining 3000-3500 key SNPs.[Conclusion]The prediction accuracy of the proposed method was im-proved to varying degrees in terms of increasing PCC and reducing RMSE compared with LightGBM,XGBoost,and other models,which provides an important technical path for realizing accurate and efficient rice molecular breeding.