Genomic selection (GS) has revolutionized the field of plant breeding by enabling the selection of superior genotypes based on genomic estimated breeding values (GEBVs) derived from dense molecular marker information. Initially proposed by Meuwissen et al. (2001), GS bypasses the need for direct phenotypic selection, allowing for early and more efficient selection decisions, thereby shortening the breeding cycle and enhancing genetic gain. This methodology represents a fundamental shift in the breeder’s toolbox, moving from phenotype-based to genotype-driven decision-making, and has been successfully implemented in numerous crop breeding programs worldwide (Crossa et al., 2017; Desta and Ortiz, 2014).

Despite its transformative potential, the implementation of GS in real-world breeding programs faces several challenges. One key limitation is the variability in prediction accuracy across different environments and breeding populations. Factors such as genotype-by-environment G × E interactions, limited training population sizes, and the genetic architecture of traits of interest can significantly hinder the robustness of genomic predictions (Roorkiwal et al., 2016). Additionally, the cost and logistics of genotyping and phenotyping large populations remain substantial obstacles, particularly in resource-limited settings. These challenges necessitate the development of novel strategies to optimize the accuracy and applicability of GS.

In response to these challenges, a growing body of research has focused on improving the prediction accuracy of GS models. Strategies such as optimizing training population design (Rincent et al., 2012), incorporating G × E interactions (Jarquín et al., 2014), and applying advanced statistical learning techniques (Montesinos-López et al., 2018) have shown promising results. However, even with these improvements, the integration of additional layers of biological data offers a compelling avenue for further enhancing model performance.

The integration of multi-omics data—including genomics, transcriptomics, metabolomics, and proteomics—has emerged as a powerful strategy to enhance the performance of genomic prediction (GP) models in plant and animal breeding. These diverse yet complementary datasets provide a multidimensional view of the complex biological systems that govern phenotypic expression, enabling a more precise dissection of the genotype-to-phenotype relationship. Unlike single-layer genomic data, which often capture only a portion of the heritable variance, multi-omics approaches can account for regulatory, transcriptional, post-transcriptional, and metabolic interactions that influence trait architecture. For example, transcriptomic data capture gene expression levels across tissues or developmental stages, shedding light on functional genes and regulatory networks underlying complex traits (Guo et al., 2016; Azodi et al., 2020). Similarly, metabolomic profiles offer dynamic snapshots of cellular biochemical processes, which are often directly associated with phenotypic traits such as growth, stress response, or yield (Riedelsheimer et al., 2012; Wen et al., 2014). Proteomics data, although less frequently used due to technical constraints, provide critical insights into post-translational modifications and protein abundance, which are closely tied to phenotypic outcomes (Misra et al., 2019). The synergistic integration of these omics layers can substantially improve the explanatory power of prediction models, particularly for complex traits governed by multiple small-effect loci and their downstream interactions (Wang M. et al., 2024). Furthermore, machine learning and statistical modeling techniques have increasingly enabled the effective fusion of high-dimensional omics data into genomic selection pipelines, resulting in significant gains in predictive accuracy (Montesinos-López et al., 2022). As such, multi-omics integration not only enriches the biological relevance of genomic predictions but also facilitates more informed decision-making in breeding programs aimed at improving crop resilience, productivity, and nutritional quality.

Several studies have demonstrated the utility of integrating multi-omics data into GS models. For example, Azodi et al. (2020) showed that combining gene expression data with genomic information improved the prediction of complex traits in maize. Similarly, Riedelsheimer et al. (2012) found that metabolite profiles significantly contributed to the prediction of biomass traits in maize hybrids.

Wang et al. (2024b) developed an extensive multi-omics atlas for wheat, integrating transcriptomic, proteomic, phosphoproteomic, and acetylproteomic data across various tissues and developmental stages. This integrative approach enhanced the understanding of complex traits, including disease resistance and grain quality, highlighting the potential of multi-omics data to improve the predictive accuracy of genomic selection models.

Despite its potential, the statistical integration of heterogeneous omics datasets presents significant challenges. These arise from inherent differences in data dimensionality, measurement scales, noise levels, and patterns of missingness across various omics platforms. Additionally, capturing the intricate—and often nonlinear—interactions both within and between omics layers, and their combined influence on complex phenotypes, requires highly sophisticated analytical frameworks. Traditional linear models commonly employed in GS may lack the flexibility to adequately model these multidimensional relationships. Consequently, there is an increasing need to adopt advanced machine learning approaches, including deep learning architectures, kernel-based methods, and Bayesian hierarchical models, which offer greater adaptability and capacity to uncover hidden structures in complex biological data (Montesinos-López et al., 2019; Montesinos-López et al., 2022).

Another critical aspect is the model tuning process. Although machine learning approaches are often highly competitive compared to traditional methods in predictive accuracy, they are frequently associated with complex and computationally intensive tuning procedures. This complexity can limit their practical applicability, especially in high-dimensional omics contexts. As such, the development of methodologies that strike a balance between predictive performance and user-friendly tuning remains a pressing research priority. Furthermore, the standardization of data preprocessing pipelines and the assurance of data quality across omics layers are essential for enhancing the reliability and reproducibility of integrative analyses.

Several recent studies have explored multi-omics integration using deep learning to predict phenotypic traits in crops and model species (Angermueller et al., 2016; Zingaretti et al., 2020; Montesinos-López et al., 2021). Although promising, most approaches have been limited by dataset size, environmental heterogeneity, or lack of benchmarking across model types. Our study addresses these limitations by evaluating predictive performance across three distinct datasets using standardized cross-validation procedures and multiple deep learning architectures.

In this research, we aim to address these challenges by integrating genomic, transcriptomic, and metabolomic data to explore alternative modeling approaches for improving GS methodology. The availability of these datasets enables the application of statistical and machine learning methods to predict complex traits by integrating genomic (G), transcriptomic (T), and metabolomic (M) data. By leveraging the complementary information provided by each omics layer, our goal is to enhance the prediction of complex agronomic traits in plant breeding. This integrative framework not only holds promise for improving model accuracy but also offers a deeper understanding of the biological mechanisms driving trait variation. We first evaluate how omics-based similarity among these datasets relates to trait variation.

Our study contributes to the growing body of evidence supporting multi-omics integration in GS and aims to identify modeling strategies that effectively harness the rich biological information embedded in diverse omics datasets. We explore conventional statistical learning methodologies capable of addressing the unique challenges of multi-omics integration, with an emphasis on practical implementation in plant breeding programs. Ultimately, this research aims to provide breeders with more accurate and biologically informed tools to accelerate genetic improvement.

We present the results in four sections. The first section reports the results for the Rice210 dataset. The second section provides the results for the Maize282 dataset, the third section provides the results for the Maize368 dataset, and the fourth section summarizes the results across datasets.

Despite the promise of GS in accelerating genetic gain in plant breeding programs, its widespread adoption in real-world applications remains constrained by several challenges. Key among these is the variability in prediction accuracy across traits, environments, and populations, often limiting the robustness and generalizability of the models (Crossa et al., 2017; Hickey et al., 2017). Moreover, the complex genetic architecture of many agronomic traits—often governed by numerous small-effect loci and subject to genotype-by-environment interactions—poses a significant barrier to achieving consistent and high predictive performance. Additionally, limitations in training population size, data quality, and the inability to fully capture the underlying biology of traits further impede the effective application of GS in breeding pipelines (Xu et al., 2017).

Our results align with recent developments in omics-based genomic prediction. Several studies have shown that combining genomics with transcriptomic or metabolomic data improves trait prediction, particularly when trait heritability is moderate or low. Compared with traditional models like GBLUP or kernel methods, deep learning architectures—especially those utilizing multi-layer feedforward or convolutional layers—can better capture nonlinear patterns inherent in omics data. The consistent superiority of MLP-based models in our results aligns with the findings of Zingaretti et al. (2020) and Montesinos-López et al. (2021).”

In response to these limitations, a variety of methodological strategies have been explored to improve the performance and reliability of GS. Among the most promising is the integration of multi-omics data, including genomics, transcriptomics, and metabolomics, which offer complementary layers of biological information. These integrative approaches aim to enhance the predictability and interpretability of GS models by capturing intermediate phenotypes and regulatory mechanisms that mediate the genotype-to-phenotype relationship (Fernie and Schauer, 2009; Acharjee et al., 2016). For example, transcriptomic data can provide insights into gene expression patterns linked to trait variability, while metabolomic profiles may reflect physiological states more directly related to phenotype expression. Such integrative models have shown potential to improve the biological relevance of predictions and increase their accuracy across different contexts (Azodi et al., 2020).

The integration of omics data into genomic prediction frameworks is thus emerging as a critical frontier for enhancing predictive accuracy and model interpretability. By leveraging synergistic information from genomics and other omics layers, researchers can better account for the biological complexity of traits, ultimately leading to more precise selection decisions. This approach aligns with systems biology paradigms, which emphasize the interconnected nature of biological data and advocate for a holistic perspective in predictive modeling (Sandhu et al., 2022). Moreover, integrating omics data enables breeders to gain mechanistic insights into trait architecture, supporting both prediction and discovery and offering a dual benefit to crop improvement programs.

Interestingly, the performance gain from integrating omics varied by dataset. In rice, where metabolomic and transcriptomic signals are rich and traits show complex regulation, omics integration yielded substantial gains. In contrast, in the chickpea dataset, where transcriptomic features may have less variance relative to genomic information, the improvement was more modest. These differences illustrate the context dependency of multi-omics modeling, highlighting that data type relevance varies by species and trait architecture.”

Using only genomic information (model M1) led to relatively lower prediction accuracy compared to integrated models (M19–M22), especially in datasets with high-dimensional transcriptomics. This highlights the added value of capturing gene expression patterns or metabolite activity, which can serve as proximal indicators of phenotypic variance. Our study reinforces the view that multi-omics layers provide complementary biological information not captured by markers alone.

Our empirical results underscore the value of an integrative approach to genomic prediction. By incorporating transcriptomic and metabolomic data along with genomic information, we observed a notable increase in heritability across all three datasets. However, a significant improvement in prediction accuracy was observed in only one dataset. This gain in accuracy is not merely incremental—it offers compelling empirical evidence for the critical importance of multi-omics integration within genomic prediction frameworks. These findings highlight the advantages of expanding the data foundation of GS models beyond genomics alone and support the systematic inclusion of all available omics data when feasible. As plant breeding moves further into a data-driven era, leveraging the full spectrum of biological information is no longer optional but essential for achieving the next generation of genetic gain.

Integrating genomic, transcriptomic, and metabolomic data did not, however, enhance prediction accuracy in two of the datasets, thus indicating that multi-omics integration is not universally advantageous. This result highlights key statistical challenges inherent to combining heterogeneous data types. Omics datasets often differ in scale, dimensionality, noise levels, and correlation structures, making integration complex. These differences can complicate model training, particularly in high-dimensional settings where the number of features vastly exceeds the number of observations, increasing the risk of overfitting and reducing generalizability. Additionally, the contribution of each omics layer may vary across samples, with some data sources offering useful signals in certain cases while adding noise in others. Such variability underscores the need for careful preprocessing, feature selection, and model design to fully harness the potential of multi-omics approaches.

Furthermore, the presence of multicollinearity within and across omics layers can inflate variance estimates and obscure the identification of truly informative predictors. Another critical issue is the potential for redundant or weakly informative signals in additional omics layers, which may dilute the predictive power when not appropriately weighted or regularized. The statistical challenge of determining which features contribute meaningfully to the prediction—and how to combine them effectively—requires sophisticated modeling strategies such as dimension reduction, penalization, or multi-view learning approaches.

These findings underscore the need for rigorous and statistical frameworks that can manage the complexity and high dimensionality of multi-omics data while also capturing their complementary information. Addressing these challenges is essential to fully harness the promise of integrative omics in genomic prediction and avoid misleading conclusions based on suboptimal integration methods.

In general, despite the demonstrated efficiency of the integration of multi-omics data in prediction, there is still considerable room for improvement in learning from these different datasets. For this reason, the integration of genomics, transcriptomics, and metabolomics represents a transformative strategy to overcome current limitations in genomic prediction. Our findings highlight the empirical and theoretical advantages of this approach and provide strong justification for future breeding programs to adopt multi-omics data integration as a core component of predictive breeding methodologies.

The integration of genomics, transcriptomics, and metabolomics into genomic prediction models has shown promise in enhancing predictive accuracy, although improvements have been observed only in certain datasets compared to models relying solely on genomic data. This variability underscores both the potential and the complexity of incorporating multi-omics information into prediction frameworks. Although empirical evidence highlights the added value of multi-omics integration, it also reveals significant challenges in achieving effective data fusion and optimal model performance. Nonetheless, the incorporation of multi-omics data holds great potential to enhance predictive power and enable more informed decision-making in plant breeding programs. These findings emphasize the transformative capacity of holistic, data-driven strategies in modern breeding efforts. We strongly advocate for continued research aimed at developing and refining integrative multi-omics frameworks. Their effective implementation could substantially improve the identification of superior candidate lines and accelerate genetic gain in genomic selection programs.