Abstract: This study developed a network structure-adaptive model for precise wheat phenotype prediction, named the adaptive structure genomic prediction network (AS-GPNet), to address the limited adaptability of fixed-architecture deep learning models across different phenotype prediction tasks. By constructing a dynamic network-capacity adjustment mechanism, AS-GPNet adaptively adjusts its network structure and training parameters according to the target phenotype. This strategy alleviates overfitting and underfitting across different phenotype prediction tasks, thereby improving prediction accuracy and model stability. First, a genomic prediction network integrating convolutional neural networks (CNNs), multi-head self-attention, and residual modules was constructed. The convolutional layers extract local patterns from high-dimensional genetic markers, the attention mechanism captures long-range relationships among marker features, and the residual modules enhance deep feature transmission and training stability. Second, a self-adaptive adjustment method for network structure and training parameters was designed. A tree-structured Parzen estimator (TPE) was used to jointly search for suitable network structures and training parameters according to the target phenotype. Independent optimization was performed for each trait so that the network capacity could better match the requirements of the corresponding prediction task. Finally, a composite loss function integrating mean squared error (MSE) and Pearson correlation coefficient (PCC) was designed to jointly constrain prediction error and the consistency between predicted and observed phenotypes. AS-GPNet was evaluated on independent test sets from the Wheat599 and Wheat2000 datasets and compared with GBLUP, BayesB, DNNGP, WheatGP, and Cropformer. In addition, five-fold cross-validation was conducted on the development data excluding the independent test sets to assess the model’s sensitivity to different data partitions. On Wheat599, AS-GPNet achieved the highest PCC in all four environments, with an average PCC of 0.64, representing a 6.20% improvement over the second-best model, DNNGP. In the GY1 environment, AS-GPNet achieved a PCC of 0.70, exceeding DNNGP, WheatGP, and Cropformer by 6.06%, 11.11%, and 12.90%, respectively. These results indicate that AS-GPNet can adaptively match model capacity to different environmental prediction tasks and maintain stable predictive performance. On Wheat2000, AS-GPNet achieved the highest PCC for all six agronomic traits, including grain hardness (GH), grain protein content (GP), test weight (TW), grain length (GL), grain width (GW), and thousand kernel weight (TKW). Its average PCC across the six traits was 0.72, 5.90% higher than that of WheatGP. The PCC for GL reached 0.80, whereas that for GP reached 0.61. For Wheat599, AS-GPNet achieved the highest coefficient of determination (R²) and the lowest root mean square error (RMSE) for GY1 and GY3. On Wheat2000, the highest R² and lowest RMSE were obtained for TKW, GW, GL, GH, and GP. The differences among the optimized architectures for different traits further confirmed that a single fixed network configuration is not equally suitable for all phenotype prediction tasks. In addition, evaluation on the Maize1404 dataset showed that AS-GPNet could also be applied to maize phenotype prediction, indicating potential applicability across crops. Structural ablation experiments further verified the effectiveness of the adaptive design. Compared with the best fixed-architecture variant, AS-GPNet increased the average PCC by 4.47% on Wheat599 and 3.36% on Wheat2000, demonstrating that the performance gain resulted from adaptively adjusting network depth and structural composition rather than simply increasing model complexity. To enhance biological interpretability, Shapley additive explanations (SHAP) were used to quantify the contribution of input single nucleotide polymorphism (SNP) markers. For days to anthesis (DTA) in maize, nine of the top 20 SNPs ranked by mean absolute SHAP value were located within quantitative trait loci (QTL) identified by sGWAS and hGWAS, indicating that AS-GPNet captured genetic signals relevant to DTA prediction and that the learned feature importance had biological plausibility. In summary, this study established a network structure-adaptive genomic prediction model that jointly optimizes network structure and training parameters. AS-GPNet provides an accurate, stable, and interpretable framework for phenotype prediction and offers technical support for intelligent crop breeding.