CWAGS: multi-trait genomic selection using channel weighted attention convolutional network Authors: Chunqing Cao, Farhan Bin Mohamed, Mohd Shahrizal Bin Sunar, Vei Siang Chan Published: 2026 ABSTRACT AbstractBackgroundGenomic selection serves as an effective approach to accelerate the improvement of agronomic traits in crops. However, as a core technique in modern crop breeding, genomic selection still faces many challenges in capturing complex interactions among genetic variants. This study proposes the Channel-Weighted Attention Genomic Selection Convolutional Network (CWAGS), a novel convolutional neural network specifically designed for genomic data. The major innovations define CWAGS: It employs a channel-weighted attention mechanism that reveals trait-specific genetic architectures through adaptive weight assignment to different genomic features. Second, it enhances computational efficiency through a depthwise separable convolution architecture. And integrates DropPath random depth regularization with residual connections to boost the model's generalization capability across diverse genetic backgrounds.ResultsAnalysis of channel attention weights demonstrates CWAGS's biological interpretability: different traits exhibit distinct genetic architectures, providing insights into genotype–phenotype relationships. In a comprehensive evaluation with four benchmark datasets, the CWAGS model improved average accuracy by 1.2%–4.8% compared with the suboptimal models. Channel weight attention analysis revealed distinct genetic architectures for yield, quality, and morphological traits, providing a reference for the development of deep learning frameworks for precision genomic selection.ConclusionsBy balancing prediction accuracy, and biological interpretability, CWAGS provides a reference framework for precision genomic selection. This framework facilitates crop genetic improvement through enhanced breeding efficiency.Keywords: Genomic selection, Channel-weighted attention, Deep learning, Multi-trait selection, Crop breeding Introduction Modern agriculture faces a combination of intertwined threats: climate change has led to frequent extreme weather events, arable land continues to shrink, pest and disease outbreaks recur, and soil degradation accelerates. These factors collectively threaten global food security [1, 2]. Traditional breeding techniques are time-consuming and inefficient, failing to satisfy the urgent demands for rapid crop improvement and precision agriculture [3]. Consequently, accelerating crop genetic enhancement and boosting genomic breeding efficiency have become the core objectives of crop breeding research [4]. Since crop phenotypes arise from complex regulatory interactions among genotypes, environments, and their interactions, important agronomic traits such as yield, disease resistance, and quality are typically controlled by multiple genes. These traits exhibit high genetic complexity and environmental dependence [5, 6]. Traditional phenotypic selection (PS) requires multi-location, multi-year field trials, which are not only time-consuming and labor-intensive but also fail to accurately assess a genotype's true genetic potential [7]. Consequently, the emergence of genomic selection (GS) technology offers a revolutionary approach to address these challenges [8, 9]. By establishing predictive models linking high-density single nucleotide polymorphism (SNP) markers in training populations to observed phenotypes, genomic selection enables the prediction and selection of superior genotypes without extensive field trials. This significantly shortens breeding cycles and enhances selection efficiency [10, 11]. The development of genetic marker breeding methods has undergone a major evolution from linear statistical models to deep learning approaches [12]. Early model implementations primarily relied on linear models, such as Regression-Based Regression for Best Linear Unbiased Prediction (RRBLUP), which assumes all SNP effects follow a normal distribution with equal variance and employs L2 regularization to address the issue of a marker count far exceeding the sample size [13]. The genomic best linear unbiased prediction (GBLUP) model utilizes genomic relationship matrices to capture genetic similarity among individuals [14, 15]. Concurrently, novel Bayesian methods introduce diverse prior distributions, permitting marker effect heterogeneity and thereby enhancing quantitative trait locus (QTL) detection capabilities [16]. Bayesian approaches, including BayesA and BayesB variants [47], create broader scope for improvement within the linear model framework which allows for more flexible modeling of genetic structures. However, these linear models have inherent limitations: its assumption of additive gene effects overlooks epistasis and gene-environment interactions; for traits with special genetic architectures, it is unable to effectively handle higher-order interaction effects [7, 17]. Machine learning in genomic selection To explore the application of machine learning in genomic selection [18], researchers employed Support Vector Regression (SVR) to map data into a high-dimensional feature space by constructing kernel functions, thereby extracting intrinsic associations between input features and response variables, enabling the modeling of complex nonlinear patterns [17]. Additionally, Random Forest (RF) was employed to construct and integrate multiple decision trees, enabling the handling of high-dimensional data and variable interactions where the splitting process implicitly simulates interactions among SNPs [19, 20]. By integrating the predictions from each tree, it avoids overfitting. More importantly, RF can capture nonlinear relationships without requiring explicit feature engineering. Thus, these machine learning algorithms can represent nonlinear relationships between genotypes and phenotypes effectively [17, 20]. However, these models often require computationally intensive hyperparameter optimization, are prone to overfitting on ultra-high-dimensional genomic data, and exhibit poor interpretability, making it difficult to derive biological insights [21]. Deep learning in genomic selection Therefore, the introduction of deep learning technologies has opened new avenues for genomic selection research [11, 22]. This is primarily because deep learning (DL) techniques can extract complex multidimensional features from large datasets and process intricate nonlinear and high-dimensional feature information [12, 23, 24]. Consequently, Ma et al. [25] for the first time applied convolutional neural networks (CNNs) to genomic selection in wheat and maize. Liu et al. [26] combined CNNs with genome-wide association studies (GWAS) to develop a deep convolutional neural network-based DLGWAS method for soybean phenotypic selection. Concurrently, Abdollahi-Arpanahi et al. [27] compared multiple deep learning architectures and found that convolutional neural networks exhibit advantages in capturing local linkage disequilibrium patterns. This research collectively confirmed the potential of deep learning techniques to capture complex nonlinear patterns. During their research, Wang et al. [28] employed Principal Component Analysis (PCA) dimensionality reduction to process multi-omics data, applying it to their proposed DNNGP model, which demonstrated robust predictive performance across diverse plant species and outperformed traditional methods in genomic selection tasks. Its architecture incorporates a three-layer convolutional neural network with batch normalization and dropout regularization. This model demonstrated that batch normalization, early stopping, and ReLU activation functions are three effective techniques for optimizing genomic selection, overcoming the limitations of traditional linear models in modeling non-additive genetic effects. To overcome the limitations of sequential models, Xie, Weng, et al. [29] developed the PNNGS model employing a parallel convolutional architecture that enhance genomic prediction capabilities. By utilizing multi-scale convolutional kernels to extract genetic features at different scales, this approach effectively overcomes the inherent information bottleneck of serial architectures and mitigates the single kernel size selection issue. Attention mechanisms, originally developed for sequence-to-sequence models in natural language processing, enable neural networks to dynamically focus on the most relevant parts of the input data when making predictions. In essence, an attention mechanism assigns learnable importance weights to different input features or positions, allowing the model to 'attend' to informative signals while down-weighting noise or irrelevant information. In the context of genomic prediction, attention mechanisms serve two primary purposes: (1) improving predictive performance by enabling the model to selectively emphasize genomic regions with stronger phenotypic associations, and (2) enhancing model interpretability by providing quantitative importance scores for different genomic features, which can be linked back to biological knowledge about gene function and regulation, thereby providing clues for in-depth biological research. To validate the underlying genetic principles, researchers have turned their focus to attention mechanisms, thereby enhancing the interpretability of genomic data models and advancing genomic prediction [30]. Attention mechanisms not only improve model performance but also strengthen the model's ability to focus on key genomic loci [31]. Building upon this foundation, Wu et al. [32] developed the GPformer model by integrating self-attention mechanisms with positional encoding techniques to capture long-range dependencies within SNP sequences. This model successfully fuses GWAS information as prior knowledge with deep learning, simultaneously demonstrating that biological prior knowledge can be effectively incorporated into attention mechanisms. The CropFormer model proposed by Wang et al. [33] achieves significant improvements across multiple crops including maize and wheat by modeling complex interactions among SNPs through a multi-head self-attention mechanism [34]. Concurrently, explorations of hybrid architectures demonstrate substantial potential in combining convolutional neural networks (CNNs) with attention mechanisms [23]. The WheatGP model developed by Wang et al. [35] integrates the strengths of convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), effectively capturing both local and global features in wheat. Gao et al. [36] noted that integrating advanced attention mechanisms (such as self-attention and cross-attention) into genomic data processing enhances models' ability to focus on relevant features across different data types. These models demonstrate the immense potential of deep learning in modern breeding. However, deep learning methods still face limitations that require urgent resolution. The continuous increase in model complexity imposes a significant computational burden, accompanied by dimensionality issues in genomic data calculation [37–40]. Furthermore, models often overlook the differential contributions of various features to specific traits during training. This makes it challenging to identify key genomic regions associated with traits, limiting their application in functional genomics. To address these challenges and research gaps, this paper proposes the Channel-Weighted Attention Genomic Selection Convolutional Network (CWAGS), which enhances prediction accuracy by optimizing adaptive feature channel weights tailored to different agronomic traits. To comprehensively validate CWAGS's effectiveness, we conducted a systematic evaluation across four representative datasets. Comparative experiments against six benchmark methods demonstrate that CWAGS achieves competitive performance across diverse datasets and traits, demonstrating robustness under varied crops, sample sizes, and environmental conditions. This method can be applied in genomic selection, providing tools for precision breeding. Methods Datasets To comprehensively evaluate CWAGS performance across different genetic backgrounds and environmental conditions, this study employed four public genomic datasets representing distinct crop species, population sizes, marker densities, and environmental conditions: Rice 44 K, Wheat 33 K, Wheat 599, and Rice 469. Together, these datasets form a comprehensive evaluation framework encompassing challenges ranging from simple to complex scenarios, single to multiple environmental conditions, and low-dimensional to high-dimensional data. Thus, the model's performance can be systematically assessed under various demanding conditions. The Wheat 33 K dataset, which contains 2,000 Iranian bread wheat landraces, was provided by the International Maize and Wheat Improvement Center (CIMMYT). The original dataset comprises 8,416 Mexican landraces and 2,403 Iranian landraces, genotyped using 33,709 Diversity Array Technology (DArT) markers for genome-wide analysis. This dataset is characterized by large sample size and rich genetic diversity [41]. This study focuses on six core agronomic traits extracted from the original phenotyping files: grain hardness (GH), grain length (GL), grain width (GW), grain protein content (GP), thousand-grain weight (TKW), and test weight (TW). The original marker matrix contained no missing values, thus no additional marker filtering was performed. For the Wheat 599 dataset, CIMMYT's global breeding program provided 599 wheat lines. These lines were tested across four environments, and each environment represented different growing conditions [42]. The genotyping was done with 1,447 DArT polymorphic markers. We focused on grain yield (GY) as the target trait. Because the lines were grown in multiple environments, we used BLUPs calculated from mixed linear models to get reliable estimates of genetic effects. This helped us account for environmental variation. We ended up with 1,279 markers after quality control. The Rice 44 K dataset includes 413 rice inbred lines collected from 82 countries [43]. Originally, there were 44,100 SNP markers, but SNPs with low quality—that is, low call rate and low allele frequency across all samples—were removed from the dataset; thus, 36,901 high-performing SNPs with call rate > 70% and minor allele frequency (MAF) > 0.01 were retained and used for all analyses. We selected six traits for prediction: Seed number per panicle (SNPP), Flag leaf length (FLL), Seed length (SL), Plant height (PH), Florets per panicle (FPP), and Brown rice seed length (BRSL). These traits cover both morphological features and yield components. The Rice 469 dataset contains 469 rice varieties that were planted in Hangzhou and Lingshui in China [44]. Genotyping was performed using Illumina chips designed specifically for this study, which gave us 5,291 SNP loci. The six traits we analyzed were: tiller angle (TA), plant height (PH), flag-leaf length (FLL), flag-leaf width (FLW), flag-leaf length/width ratio (FLLRW), and panicle number (PN). CWAGS architecture Overall architecture design The genome to phenotype prediction model proposed in this study is based on a deep convolutional neural network architecture. It integrates a channel-weighted attention mechanism, which fully accounts for the high-dimensional nature of genotype data and the complexity of crop phenotype expression. The model utilizes a step-by-step feature refinement architecture, as shown in Fig. 1A, primarily comprising an input layer, feature extraction layer, deep backbone module (main module), dropout layer, flattening layer, linear output layer, and final phenotype prediction layer. This architecture enhances feature representation capabilities and nonlinear modeling capacity while maintaining a controllable parameter size, showing powerful bioinformation capture and generalization capabilities. CWAGS architecture with hierarchical views. A Overall pipeline from SNP genotypes to phenotype selection. B A main block from (A) using depthwise separable convolution and channel weighted attention, with Batch Normalization (BN) and Swish in a residual connection. C Channel weighted attention (CWA) module from (B), consisting of global average pooling followed by feature convolutions that produce per-channel weights The model takes preprocessed SNP data as input. The input data enters the feature extraction layer, which consists of a convolutional layer, a batch normalization layer, and a Swish activation function, designed to perform preliminary feature extraction on the raw gene sequences. Subsequently, the data is modeled for deep level representation through multiple stacked main modules (N × main modules). The Swish activation function is defined as: The smoothness and non-monotonicity of the Swish function offer distinct advantages in processing genomic data: it preserves negative value information while offering smoother gradients, thereby facilitating the learning of cumulative genetic effects. The backbone module adopts a feature extraction architecture [45]. By expanding feature channels, the network extends into the feature space of genotype–phenotype mapping relationships, combined with the selective action of the channel attention mechanism. This design enables the model to capture diverse genetic patterns within the feature space, including additive and dominant effects. The module employs depthwise separable convolutions instead of traditional convolutions, enhancing local feature extraction while reducing model parameters. Subsequently, a Channel Weighted Attention (CWA) module recalibrates feature importance across channels. This attention mechanism selects feature combinations most relevant to target traits, expanding the model's expressive power. Finally, a residual addition operation is introduced within the module to facilitate cross-layer feature information flow, mitigating gradient vanishing issues and enhancing training stability. To address overfitting risks associated with network depth, each backbone module integrates a DropPath random depth mechanism to strengthen model generalization. The backbone module architecture is illustrated in Fig. 1B. Within the model architecture, the Channel Weighted Attention (CWA) module plays a pivotal role in the backbone structure, enhancing the model's responsiveness to key gene channels. This module first extracts feature representations along the channel dimension, capturing their overall contribution to the current sample. Subsequently, it performs channel-level feature attention extraction through a lightweight mapping structure. Convolution reduces the dimensionality to form low-dimensional embedding vectors. Following nonlinear transformation via activation functions, another convolutional operation restores the channel dimension, ultimately generating channel weight coefficients. These channel weight vectors act upon the original features, dynamically adjusting the importance weights of each channel. Through this mechanism, the model automatically focuses on feature channels that contribute more significantly to phenotype prediction while suppressing redundant or noisy signals. This guides the model to concentrate on genetic loci or signals that contribute more substantially to the current phenotype prediction task, thereby enhancing generalization capability and efficiency. The module structure is illustrated in Fig. 1C. denotes Global Average Pooling, W1 and W2 represent the dimension reduction and expansion transformation matrices separately, and denotes the Sigmoid function. Therefore, the total parameter count for depthwise separable convolution and the parameter reduction ratio are [46]: where denotes the convolutional kernel parameters and denotes the bias parameters. Depthwise separable convolution decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. The depthwise convolution performs the convolution operation independently for each input channel, using a separate convolution kernel for each channel. The number of parameters is: Pointwise convolution uses convolutions to map the output of depthwise convolution to the desired number of output channels: Therefore, the total parameter count for depthwise separable convolution and the parameter reduction ratio are: The model training process employs Mean Squared Error (MSE) loss function, defined as: where represents true phenotype values, represents model predictions, and is the sample size. The MSE loss function captures the continuous nature of quantitative traits, providing stable gradient signals. Combined with adaptive learning rate scheduling and multiple regularization strategies, it ensures robust convergence across diverse datasets. By stacking these backbone modules, the model achieves transformation from genotype sequences to deep level phenotype related features. Subsequently, the network undergoes Dropout layer processing to mitigate overfitting issues. Through flattening layers, features are transmitted to linear layers for regression mapping, ultimately outputting phenotype prediction results. The entire network supports end-to-end training, adaptable to multiple crops, diverse environments, and various phenotype prediction tasks, demonstrating structural flexibility and multi-task training capabilities. Methods of comparison To ensure a complete and impartial performance evaluation of CWAGS, we have built a benchmark framework covering three critical stages in the development process of genomic selection methods. Six benchmark methods are selected for comparison due to their extensive application in genomic selection and their proven availability across diverse crop species, including RRBLUP [13], Bayesian approaches [47], RF [20], SVR [17], DNNGP [28], and PNNGS [29]. Evaluation metrics utilized For the predictive assessment of continuous traits (such as yield and growth characteristics), we employ the Pearson correlation coefficient (PCC) as the primary evaluation metric [28]: where and represent observed and predicted phenotype values respectively, while and denote their respective means. The PCC value ranges from −1 to 1, with values closer to 1 indicating a stronger positive linear relationship between predicted and observed values. All models were trained using tenfold cross-validation to ensure robust performance evaluation and minimize overfitting bias. Final performance metrics quantify prediction stability by calculating the average across all 10 folds. Parameter settings and experimental validation of CWAGS This study was conducted on a platform equipped with NVIDIA hardware. The experimental environment utilized the PyTorch deep learning framework, employing CUDA for training acceleration. To ensure reproducibility of experimental results, the random seed was fixed throughout the entire experiment. The AdamW optimizer was employed. Compared to the traditional Adam optimizer, this algorithm decouples weight decay from gradient updates, enabling more effective regularization. To prevent overfitting, the weight decay regularization coefficient was set to 0.001. An adaptive learning rate adjustment mechanism was employed to optimize the initial learning rate for different datasets, implementing multiple scheduling strategies during training. The model training cycle was set to 400 epochs to ensure sufficient convergence, while an early stopping mechanism was deployed to prevent overfitting. Results Phenotypic correlation analysis In this study, we employed Pearson's correlation coefficient as a statistical metric to evaluate associations between different phenotypes, as well as visualized these relationships through heatmaps. Phenotypic correlation analyses across four datasets revealed patterns of genetic correlations among different trait classifications (Fig. 2). Phenotypic correlation analysis across four datasets. Pearson correlation coefficients among phenotypic traits in (A) Rice 44 K, (B) Wheat 33 K, (C) Wheat 599, and (D) Rice 469 datasets. Circle size represents the absolute value of correlation coefficients, with colors indicating the direction and magnitude of correlations: blue circles denote positive correlations, red circles denote negative correlations, and color gradient from light to dark represents increasing correlation strength, with darker shades indicating stronger associations Within the Rice 44 K dataset, we observed significant correlation patterns among phenotypic traits. Traits related to grain exhibited strong correlations, with seed length (SL) showing the most significant association with brown rice seed length (BRSL). The number of florets per panicle (FPP) also demonstrated a highly positive correlation with the number of seed number per panicle (SNPP), indicating that FPP represents a potential upper limit for grain number. flag leaf length (FLL) showed moderate correlations with plant height (PH) and number of florets per panicle (FPP), suggesting these traits are associated yet controlled by distinct genetic factors. Wheat 33 K gene data revealed correlations among wheat quality traits. thousand kernel Weight (TKW) showed significant positive correlations with both grain width (GW) and grain length (GL), with correlation coefficients of 0.76 and 0.57, respectively. grain protein (GP) generally showed low correlations with other physical traits (correlation coefficients ranging from −0.33 to −0.36), exhibiting a negative correlation trend. This may reflect the trade-off between physical and quality traits in wheat breeding. The wheat 599 dataset focused on analyzing the performance of the same phenotype across different environments. Results indicate that the ENV1 environment showed lower correlations with the other three environments, while ENV2 and ENV3 exhibited the highest phenotypic correlation (0.66), suggesting relatively similar environmental conditions or similar genotype response patterns to these environments. However, ENV1 showed significantly lower correlations with other environments, indicating that ENV1 may represent a unique ecological environment, leading to more pronounced genotype × environment interaction effects. In the Rice 469 dataset, flag leaf length (FLL) exhibits a highly positive correlation (0.77) with flag leaf length/width ratio (FLLRW), while flag leaf width (FLW) shows a negative correlation (−0.55) with FLLRW, consistent with expected biological patterns. flag leaf width (FLW) showed a negative or low correlation with panicle number (PN). tiller angle (TA) exhibited extremely low correlations with other phenotypes. Analysis of phenotypic correlations across four datasets revealed: (1) Functionally related traits typically showed high positive correlations, such as associations among grain physical traits; (2) Quality traits and yield traits may show negative correlations, reflecting trade-offs in breeding practice; (3) The same phenotype exhibits variations across different environments, highlighting the importance of genotype × environment interactions; (4) Certain traits demonstrate relative independence, potentially offering opportunities for independent selection. CWAGS compared with other mainstream methods In experimental evaluations on the Wheat 599 dataset, the CWAGS model displayed predictive capabilities under complex genotype-environment interactions (Fig. 3A). Overall performance showed that CWAGS achieved an average prediction accuracy of 0.5368, outperforming RR-BLUP (0.3407, + 57.6%), Bayesian (0.4794, + 12%), RF (0.4868, + 10.2%), SVR (0.5137, + 4.5%), DNNGP (0.4944, + 8.6%), and PNNGS (0.5123, + 4.8%). In the ENV1 environment, CWAGS achieved the highest prediction accuracy of 0.6023. This environment exhibits a distinct performance gradient: CWAGS (0.6023) > PNNGS (0.5857) > DNNGP (0.537). All deep learning methods outperformed machine learning approaches SVR (0.5825) and RF (0.5507), as well as traditional methods Bayesian regression (0.5097) and RR-BLUP (0.3465). Under ENV2 conditions, predictive accuracy declined across all methods, although CWAGS achieved the highest accuracy among the compared methods with a correlation coefficient of 0.537. ENV3 posed severe predictive challenges, causing significant performance drops for all approaches. RR-BLUP achieved only a correlation coefficient of 0.2861, indicating severe limitations of linear assumptions under complex genotype by environment interactions. Nevertheless, CWAGS outperformed RR-BLUP by 61.3% and PNNGS by 4.4%. In the ENV4 environment, CWAGS achieved a correlation coefficient of 0.5565. Machine learning methods performed relatively well in ENV4 (RF: 0.511, SVR: 0.5292). The channel attention mechanism adapts feature weights in response to environmental variations. BLUP predictions (CWAGS: 0.5276) further validate the model's ability to capture stable genetic effects across environments. This adaptive adjustment pattern indicates that CWAGS not only captures predictive signals but also identifies environment-specific genetic response patterns. Genomic prediction accuracy comparison across seven methods. A Prediction accuracy (Pearson correlation coefficient) across four environments (ENV1–ENV4) in the Wheat 599 dataset. B Prediction accuracy for six traits in the Wheat 33 K dataset: GH, grain hardness; GL, grain length; GW, grain width; GP, grain protein; TKW, thousand kernel weight; TW, test weight. Methods compared include RRBLUP, ridge regression best linear unbiased prediction; BAYESIAN, Bayesian regression; RF, random forest; SVR, support vector regression; DNNGP, deep neural network for genomic prediction; PNNGS, parallel neural network for genomic selection; and CWAGS, channel weighted attention genomic selection The Wheat 33 K dataset validated CWAGS's capability to process high-dimensional data (Fig. 3B). The average Pearson correlation coefficient across six quality traits reached 0.6733, surpassing the other compared methods on average. Compared to the next-best deep learning approach, CWAGS outperformed DNNGP (0.6565) by 2.5% and PNNGS (0.6596) by 2.1%. In grain length (GL) prediction, CWAGS led with a correlation coefficient of 0.7385, indicating grain length may be regulated by relatively simple genetic structures. Conversely, grain protein (GP) prediction proved more challenging, with all methods showing reduced accuracy; however, CWAGS maintained relatively high performance at 0.5613. Predictions for grain physical traits (GH, GL, GW) showed high consistency with thousand kernel weight (TKW) predictions. For grain hardness (GH) and grain width (GW) prediction, CWAGS achieved 0.6793 and 0.7466 respectively, demonstrating the modeling capability for yield-related traits. For the important yield trait of thousand-grain weight, the improvement rate was 2.5% modest improvement than DNNGP (0.6597) and 2.4% modest improvement than PNNGS (0.6603). The stability demonstrated by this technology across related traits indicates that CWAGS successfully captures the common genetic basis regulating grain development. Differences in high-dimensional data processing capabilities were particularly pronounced on the Wheat 33 K dataset. CWAGS effectively reduced computational complexity through depthwise separable convolutions, while the channel attention mechanism enabled selection of informative markers. Performance variations across traits with different genetic architectures levels provided important biological insights. Traits with high genetic architectures (e.g., GL, GW) generally exhibit higher prediction accuracy (> 0.7), whereas moderate genetic architectures quality traits (e.g., GP) show relatively lower prediction accuracy (< 0.6). CWAGS maintains a relative advantage across all trait types, particularly excelling in improving special genetic architectures traits, highlighting its relative advantages in handling complex traits. In the Rice 44 K dataset, CWAGS demonstrated consistently competitive performance across six wheat quality traits (Fig. 4A). Based on cross-trait average prediction accuracy, CWAGS achieved 0.728, surpassing RRBLUP, BAYESIAN, RF, SVR, DNNGP, and PNNGS by 7.5%, 6.4%, 7.4%, 17.8%, 1.8%, and 1.2%, respectively. Prediction accuracy was higher for plant height (PH) and brown rice seed length (BRSL), partly due to these traits' genetic architectures, which facilitated model identification of relevant features. CWAGS achieved 0.8112 for PH prediction and 0.8308 for BRSL prediction, ranking first among all methods for both traits. In contrast, SVR performed poorly for certain traits (PH: 0.5675, FLL: 0.4089), reflecting the limitations of kernel function selection in predicting high-dimensional genomic data. Prediction of the important factor seed number per panicle (SNPP) proved challenging, with correlation coefficients ranging between 0.57 and 0.65 across all methods. CWAGS (0.6468) outperformed RR-BLUP (0.5912) by 9.4%. Prediction accuracy for flag leaf length (FLL) was generally low (all methods ranged from 0.41 to 0.58), but deep learning methods significantly outperformed others. The CWAGS model achieved 0.5782, representing a 41.4% improvement over SVR (0.4089) and an 11.0% improvement over RF (0.5208). Such pronounced performance differences suggest flag leaf length may involve complex genetic regulatory mechanisms, confirming the relative advantage of deep learning in capturing intricate genetic structures. Genomic prediction accuracy comparison across seven methods. A Prediction accuracy (Pearson correlation coefficient) for six traits in the Rice 44 K dataset: SL, seed length; PH, plant height; BRSL, brown rice seed length; SNPP, seed number per panicle; FLL, flag leaf length; FPP, florets per panicle. B Prediction accuracy for six traits in the Rice 469 dataset: PH, plant height; FLL, flag leaf length; FLW, flag leaf width; TA, tiller angle; FLLRW, flag leaf length/width ratio; PN, panicle number In the genomic prediction evaluation using the Rice 469 dataset, performance analysis of the CWAGS model for six key rice agronomic traits based on average prediction accuracy (Fig. 4B) revealed that CWAGS achieved the highest mean prediction accuracy among the compared methods (0.6757), next was PNNGS (0.6561), while RR-BLUP performed the worst (0.5529). CWAGS outperformed RRBLUP, BAYESIAN, RF, SVR, DNNGP, and PNNGS by 22.2%, 6.2%, 7.4%, 14.6%, 3.2%, and 3.0%, compared respectively. Plant height (PH) prediction yielded the highest accuracy across all methods, reflecting this trait's relatively simple genetic architecture. CWAGS achieved a correlation coefficient of 0.8653 for plant height prediction (Rice 469), comparable to PNNGS (0.8643) and DNNGP (0.8614). All three deep learning methods show similar performance for this trait. Notably, even the traditional RR-BLUP achieved a high prediction accuracy of 0.8336, indicating that plant height is primarily governed by additive genetic effects. In contrast, the prediction results for tillering angle (TA) exhibited a distinctly different pattern. Prediction accuracy significantly declined across all methods, with CWAGS reaching a maximum of only 0.4617 and RR-BLUP achieving merely 0.2758. This generally low prediction accuracy suggests TA may possess genetic architectures or be strongly influenced by environmental factors. Nevertheless, CWAGS still achieved a 21.1% overall improvement compared to other models. In predicting flag leaf-related traits, CWAGS demonstrated consistent advantages. Its accuracy reached 0.6637, 0.6924, and 0.6708 for flag leaf length (FLL), flag leaf width (FLW), and panicle number (PN), respectively. When predicting the derived trait flag leaf length-to-width ratio (FLLRW) using the CWAGS model, the accuracy reached 0.7005, demonstrating the model's capability in handling complex traits. CWAGS identifies genomic hotspots The CWAGS model provides guidance for identifying candidate genes associated with agronomic traits by analyzing SNP importance scores in the training model. Through systematic analysis of six agronomic traits in the Rice 469 dataset, we examined the interpretability of the CWAGS model and identified trait-specific genomic hotspot regions. The SNP significance distribution analysis (Fig. 5) enables us to systematically evaluate the genomic distribution patterns of key genes associated with different traits. Using the 99th percentile (red line) and 95th percentile (orange line) as significance thresholds, different agronomic traits exhibit distinct genomic distribution patterns, reflecting differences in their underlying genetic regulatory mechanisms. Concentrated distribution patterns suggest that these traits are likely regulated predominantly by a small number of genes with substantial effects. In contrast, complex agronomic traits display dispersed distribution characteristics, indicating that these traits may be subject to the coordinated regulation of multiple genes. Genome-wide SNP importance scores identified by CWAGS across six rice agronomic traits. Manhattan plots showing the distribution of average SNP importance scores across tenfold cross-validation for flag leaf length (FLL), flag leaf width (FLW), flag leaf length-to-width ratio (FLLRW), plant height (PH), panicle number (PN), and tiller angle (TA). Each point represents a single SNP, with position on the x-axis and importance score on the y-axis. Red points highlight the top 5 most important SNPs for each trait. Horizontal dashed lines indicate statistical thresholds: red line (99th percentile) and orange line (95th percentile) To validate the reliability of SNP importance assessment, we analyzed the stability of the top 10 most important SNPs across cross-validation folds (Fig. 6). Overall, the candidate SNPs for PH, PN, and FLL showed relatively small variation across folds, with narrower interquartile ranges and fewer outliers; FLLRW exhibited intermediate stability; while FLW and TA displayed relatively larger fold-to-fold variation, with wider box widths and a greater number of outliers. Cross-fold stability of top 10 SNPs across six rice agronomic traits. Box plots show the distribution of importance scores for the top 10 most important SNPs across tenfold cross-validation for six traits. Each box represents the interquartile range (25th-75th percentile) of importance scores across folds, with the median shown as the central line, and outliers are shown as individual points To evaluate the biological relevance of the candidate loci identified by CWAGS, we cross-referenced the candidate SNPs exceeding the 99th percentile significance threshold for each trait with the significant loci reported by Lu et al. [44] in rice populations. The results showed that some candidate SNPs for the six traits corresponded to significant loci for the respective traits reported by Lu et al. [44]: the candidate SNP seq-rs3945.1 for tiller angle (TA) identified by CWAGS corresponded to a reported association locus (Chr8, 20.83 Mb, P = 1.00 × 10–6, R2 = 5.94%); seq-rs5399 and seq-rs489 for flag leaf length (FLL) corresponded to reported loci (Chr11, 27.87 Mb, P = 3.60 × 10–5, R2 = 4.52% and Chr1, 32.54 Mb, P = 1.95 × 10–3, R2 = 2.08%, respectively); seq-rs669.1 for flag leaf length-to-width ratio (FLLRW) corresponded to (Chr1, 43.85 Mb, P = 9.00 × 10–5, R2 = 4.09%); seq-rs1877.1 for flag leaf width (FLW) corresponded to (Chr3, 36.25 Mb, P = 1.97 × 10–4, R2 = 3.76%); seq-rs3208.1 for panicle number (PN) corresponded to (Chr7, 4.89 Mb, P = 2.84 × 10–4, R2 = 3.56%); and seq-rs3012 for plant height (PH) corresponded to (Chr6, 22.32 Mb, P = 1.56 × 10–4, R2 = 3.51%). It should be noted that, unlike genome wide association analysis methods that rely on population level statistical association testing, CWAGS extracts features from genotype data through deep learning. Its data-driven feature ranking can, to a certain extent, capture genomic signals that are consistent with those identified by statistical association analysis, providing a preliminary basis for cross referencing between the two approaches. This result suggests that CWAGS may serve as a complementary perspective to genome wide association analysis, offering a reference for the prioritization of candidate regions and gene mapping. Discussion The study findings validate the effectiveness of these technological innovations while revealing their distinctive characteristics and value in genomic data processing. However, CWAGS does not universally outperform all baseline methods across every trait or dataset configuration [48]. In a systematic evaluation across four datasets, channel attention weights demonstrated biological relevance: for instance, in the Rice 44 K dataset, weight patterns for yield-related traits (SNPP, FPP) exhibited significant differences compared to morphological traits (PH, FLL), directly reflecting genetic structural variation among traits. Through depthwise separable convolutions, CWAGS maintains predictive accuracy while reducing the number of model parameters, reflecting technical optimization grounded in structural analysis of genomic data. Compared to standard convolutional neural networks, achieving parameter reduction while maintaining accuracy demonstrates that judicious inductive bias is more crucial than blindly increasing model capacity. The architectural design of CWAGS embodies the integration of deep learning with domain knowledge. Each design decision is grounded in the unique properties of genomic data: specifically, in terms of the method's generalization capability and cross-species transferability. CWAGS demonstrates stable performance advantages across both rice and wheat species with differing levels of genetic complexity. Importantly, the prediction accuracy achieved by CWAGS on the Wheat 33 K dataset shows strong correspondence with the narrow-sense heritability (h2) values reported in the original dataset publication by Crossa et al. [41]: high-heritability traits such as grain length (h2 = 0.881), grain width (h2 = 0.848), grain hardness (h2 = 0.839), and thousand-kernel weight (h2 = 0.833) consistently achieved high prediction accuracy (PCC > 0.67), while the lower-heritability trait grain protein (h2 = 0.625) showed the lowest accuracy (PCC = 0.5613). CWAGS still requires further refinement in several aspects. One key challenge lies in its sample size requirements. Deep learning models typically demand large-scale training datasets to achieve optimal performance. Our analysis indicates that while CWAGS maintains stable performance with sufficient samples, its predictive accuracy declines when applied to rare germplasm resources with limited data. To address this, we are exploring transfer learning strategies that leverage features from models pre-trained on large datasets to enhance small-sample prediction capabilities. Currently, CWAGS can identify key candidate gene types and is applicable for: developing markers for molecular marker-assisted selection; designing gene editing strategies for precision breeding improvement; and constructing genetic diversity evaluation systems for core germplasm resources. This provides insights into understanding genotype–phenotype mapping relationships. Based on CWAGS study, we are integrating new interpretable modules to develop a gene selection scoring system. Future developments will involve integrating multi-omics data. The next-generation CWAGS will integrate multi-level data such as genomic and transcriptomic information. By designing multi-fusion model architectures that combine various data types, predictive accuracy can be enhanced. Architecture innovations grounded in field knowledge are more valuable than simple model applications. CWAGS provides reference tools for precision breeding by balancing technological innovation with practical needs, computational efficiency with biological interpretability. As genomic sequencing costs continue to decline and computational power advances, deep learning-based approaches such as CWAGS may offer increasing utility in practical breeding programs. Research should deepen the integration of biological understanding with artificial intelligence to develop smarter, more interpretable, and versatile genomic prediction tools. Focusing on biological knowledge integration and strategic application will unlock the true advantages of deep learning's unique capabilities. This approach not only enables precise genotype-to-phenotype predictions but also achieves phenotype-to-structure knowledge, marking the evolution of genomic selection from predictive to interpretive. It will play a key role in precision crop breeding. Acknowledgements We express our gratitude to all the investigators and participants of this study. We acknowledge the support provided by the Media and Game Innovation Centre of Excellence. Authors' contributions CC and FM designed the study. CC performed the data analysis, developed the deep learning framework, and drafted the manuscript. FM supervised the research and provided critical insights. MS and VSC guided the writing of the manuscript. All authors read and approved the final manuscript. Funding This research did not receive funding. Data availability The datasets analyzed in this study are publicly available from the following sources: The Wheat599 dataset is available in the BGLR R package (https://CRAN.R-project.org/package=BGLR). The Wheat33K dataset is available at (https://data.cimmyt.org/dataset.xhtml?persistentId=hdl:11529/10548918). The Rice44K dataset is available at (http://www.ricediversity.org/data/sets/44kgwas). Rice469 dataset dataset is available at (10.5061/dryad.cp25h). Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets analyzed in this study are publicly available from the following sources: The Wheat599 dataset is available in the BGLR R package (https://CRAN.R-project.org/package=BGLR). The Wheat33K dataset is available at (https://data.cimmyt.org/dataset.xhtml?persistentId=hdl:11529/10548918). The Rice44K dataset is available at (http://www.ricediversity.org/data/sets/44kgwas). Rice469 dataset dataset is available at (10.5061/dryad.cp25h).