DeepPFT-GANCap: An explainable attention-based hybrid deep learning framework integrating protein language models and wavelet denoising-based evolutionary features for photosystem-II core protein prediction

Functional traits of plants are key to the study of ecological processes, adaptations of species, and the responses of ecosystems to alterations of the environment. Particularly, the Photosystem-II core protein (P-IICP) is vital for a plant's adaptability and efficiency of photosynthesis in several environments. The intricacy of sequence-function relations and the scope of conventional computational methods are hurdles in the precise characterization and prediction of such proteins. We propose an attention-based hybrid deep learning (DL) framework, DeepPFT-GANCap, which integrates attention mechanisms with Generative Adversarial Network, Capsule Network, and Bidirectional Temporal Convolutional Neural Network. Using advanced protein language models (ProtTrans-T5, UniRep, and ESM-2) in conjunction with a novel Discrete Cosine Transform-based evolutionary descriptor (PsePSSM-DCT) allows our model to extract highly discriminative features. To improve interpretability, we utilized a wrapper-based XGBoost-Forward Feature Selection (XGB-FFS) approach. The experimental results of DeepPFT-GANCap demonstrated consistently superior performance compared to competing predictors, exhibiting higher accuracy and good generalization across both training and independent test datasets, thereby validating the effectiveness of the proposed system. The SHapley Additive exPlanations (SHAP) analysis revealed the most influential features contributing to model predictions, emphasizing the critical functional role of P-IICP in photosynthesis and ecosystem monitoring. The study provides a scalable computational framework with advanced interpretability for monitoring plant functional traits to drive the science of trait-based ecology and the interaction of climate with ecosystems. • A novel explainable deep learning framework (DeepPFT-GANCap) is proposed for Photosystem II core protein prediction. • Multi-source protein language models and a frequency-domain evolutionary descriptor (PsePSSM-DCT) are integrated to capture complementary biological information. • A hybrid GAN–Capsule–BiTCNN architecture effectively models class imbalance, long-range dependencies, and hierarchical sequence relationships. • The framework provides biologically interpretable predictions with direct relevance to plant stress adaptation and ecosystem productivity analysis