Coformer: a deep learning-based framework for cross-environment and multi-year cotton phenotype prediction and interpretation

Accurate prediction of key agronomic traits in cotton is crucial for advancing its genetic improvement and enabling breeding-by-design. However, when using high-dimensional genomic data (e.g., massive SNP markers) for prediction, traditional models often suffer from overfitting and poor generalization. To address this, we propose an innovative deep learning model—Coformer. Coformer is a hybrid Transformer–autoencoder model: a self-attention Transformer encoder captures long-range SNP dependencies and compresses high-dimensional genotypes into a compact latent representation, which is then decoded by a supervised prediction head to the target phenotype. Through rigorous evaluation on a combined multi-environment, multi-year dataset, Coformer maintains outstanding predictive robustness even without explicitly modeling environmental factors. The model integrates a normalization module and a linear projection layer to enable adaptive input processing and end-to-end training, effectively improving generalization across varying data dimensionalities. Importantly, Coformer is interpretable: it can precisely pinpoint key genetic loci that influence target traits, providing a solid theoretical basis for deeper insight into the genetic underpinnings of phenotypes and for implementing precision breeding. In addition, to accelerate application and dissemination, we have concurrently developed a browser-based Cotton Phenotype Prediction System (CPPS) that seamlessly integrates the full workflow of “data preparation–model training–result interpretation” into a unified graphical interface. The system supports batch processing and result visualization, substantially lowering the barrier for non-specialists and offering an efficient, user-friendly solution to bridge genomics research and breeding practice.