NASGP: An efficient and interpretable neural architecture search framework for plant genomic prediction

Machine learning and deep learning techniques have been extensively adopted in genomic selection (GS) to accelerate the identification of superior genotypes and shorten breeding cycles. However, most existing methods rely heavily on manually designed algorithmic models, which require substantial domain expertise, are time-consuming and labor-intensive, and often suffer from limited robustness and poor interpretability. To mitigate the aforementioned drawbacks, this work proposes NASGP, a dedicated neural architecture search (NAS) framework tailored for plant phenotypic prediction and subsequent breeding analytical workflows. The proposed framework incorporates Bayesian optimization, a surrogate-assisted model evaluation pipeline, and a mutual information-driven dual-channel adaptive weighted feature selection strategy (MIDAFS), jointly boosting predictive performance and model building efficiency. Comprehensive validations were performed on 28 phenotypic traits derived from six plant datasets covering wheat, maize, rice, sorghum, blueberry and spruce, which verify the robust predictive capacity of NASGP for complex polygenic phenotypes. The results show that NASGP outperforms other genomic selection methods. Its average prediction accuracy is 15.39% higher than that of the second-best model, Support Vector Regression (SVR), and its computational speed is 2.94 times faster. Furthermore, NASGP can identify molecular markers closely associated with phenotypic variation, providing novel insights into the genetic mechanisms underlying complex traits. In summary, this study not only offers a novel and efficient tool for plant genomic selection but also opens new avenues for future breeding practices. NASGP is freely accessible at https://github.com/NAS-AIBreeding/NASGP .