Cotton Genomics and Genetics 2025, Vol.16, No.3, 148-162 http://cropscipublisher.com/index.php/cgg 157 comparison and found that: for quality traits such as fiber length and strength, the genomic prediction accuracy is high, and GS can be fully implemented in early-generation screening; for traits such as yield and maturity, the prediction accuracy is relatively low but still comparable to the efficiency of traditional family selection. Based on this, they suggested that public breeding projects can be divided into "two steps": first, introduce GS in fiber quality improvement to accelerate the cultivation of lines that meet industrial high-end requirements; then, with the accumulation of more environmental data and model improvement, expand GS to complex traits such as yield and stress resistance (Billings et al., 2022). It is worth mentioning that the study also compared the effects of different algorithms, including G-BLUP, BayesC, and random forests, and found that the stability of traditional linear models was slightly better when the amount of data was limited. But they also pointed out that machine learning and deep learning may have greater potential in the future. Some American breeders have begun to try to use simple artificial neural network models to simulate the combining ability of cotton hybrid combinations to predict which parent combinations are more likely to produce excellent offspring (Patil et al., 2023). These explorations of public projects have laid the foundation for the promotion of GS in cotton breeding. The experience of the United States also emphasizes the importance of model interpretability by breeders: they expect the prediction model to not only give results, but also indicate which markers or genes are most important in trait control, so as to verify with traditional genetic knowledge. Therefore, the US team often combines GS analysis with GWAS, incorporating significant markers into the model or annotating the markers with the highest model weight. This practice has increased breeders' trust in AI models and increased the adoption rate of GS results in practice. Overall, the US public breeding department has clarified the advantages and disadvantages of GS and its application boundaries through comparative analysis, providing a scientific basis for the implementation of technology. At present, they are promoting the establishment of a cotton breeding big data platform to integrate scattered historical breeding data and create conditions for the large-scale implementation of artificial intelligence-assisted breeding in the next step. 5.3 Intelligent design breeding systems in China's cotton breeding Compared with Europe and the United States, my country's practice of artificial intelligence-enabled cotton breeding started later but progressed rapidly. On the one hand, national scientific research institutions and universities are actively carrying out relevant research; on the other hand, enterprises and new R&D institutions have also joined in to develop intelligent breeding platforms. In recent years, the Cotton Research Institute of the Chinese Academy of Agricultural Sciences has laid out the "smart breeding" research direction, using its own cotton germplasm resource bank, phenotyping platform and molecular laboratory to explore the application of AI technology in the entire process of cotton breeding (Si et al., 2022). One of the representative achievements is the construction of a cotton whole genome selection breeding platform. A joint team from Zhejiang University and the Chinese Academy of Agricultural Sciences reported in 2023 that they integrated 32.5 Tb of multi-omics and phenotypic data, developed a central database for breeders to query gene expression, gene networks and epigenetic information, and established a cotton trait prediction model and decision support system on this basis. The platform is figuratively called the breeding "central kitchen". Breeders only need to input the genotypes of candidate parents, and the system can give the predicted performance and optimal selection plan of the hybrid combination on the target trait. Although the platform is still in the trial stage, it has shown great potential to shorten the breeding cycle (Zhao et al., 2024). Another eye-catching case is the cooperation between Lakeside Laboratory and Xinjiang Academy of Agricultural Sciences to use AI to crack the genetic mechanism of cotton stress resistance and apply it to breeding. They constructed a genome-wide DNA methylation map of 207 cotton varieties, identified 287 million single methylation polymorphic sites, and developed a deep learning model DeepFDML to predict which methylation variations affect gene expression, thereby discovering 43 key eQTM genes potentially involved in fiber development. More importantly, they successfully increased cotton fiber length after editing one of the genes through CRISPR. This achievement shows that AI can not only assist in selection, but also guide the discovery of gene editing targets to achieve "design breeding". In terms of the development of intelligent breeding systems, some domestic agricultural technology companies have also invested in it. Of course, the construction of my
RkJQdWJsaXNoZXIy MjQ4ODYzNA==