Predicting protein secondary structure is essential for understanding protein function, folding mechanisms, and drug discovery applications. Although recent deep learning models, particularly those based on transformer architectures, have improved prediction performance, many approaches rely on computationally expensive fine-tuning or lack explicit mechanisms for sequential refinement at the residue level. In this work, we propose E2BNet, a hybrid deep learning framework that integrates pretrained ESM2 protein language model embeddings with a lightweight Bidirectional Long Short-Term Memory (BiLSTM) module for efficient residue-level sequence modeling. Unlike approaches that rely solely on transformer outputs or extensive fine-tuning, E2BNet introduces a feature integration strategy in which contextual embeddings from ESM2 are refined through a BiLSTM layer to enhance local sequential coherence and improve prediction consistency near secondary-structure boundaries. This design allows the model to leverage rich global protein representations while maintaining low additional computational overhead. Experimental evaluation on multiple benchmark datasets, including CASP12, PDB-PISCES-2018, and ProteinNet, demonstrates that E2BNet achieves strong performance in 8-state secondary structure prediction, improving weighted F1-score and Q8 accuracy compared with several existing approaches. These results highlight the effectiveness of combining pretrained protein language representations with lightweight sequential refinement for accurate and scalable protein secondary structure prediction.
Droba et al. (Mon,) studied this question.