Enhancing Protein Secondary Structure Prediction with a Bi-LSTM Model and Random Sampling Technique

Protein secondary structure prediction (PSSP) plays a critical role in understanding protein functions, interactions, and structural dynamics. Accurate prediction relies on classification schemes such as SST3 and SST8, which represent varying levels of structural detail. However, class imbalance remains a significant challenge, with abundant classes like Alpha-Helices (H) and Coils (C) overshadowing rare structures such as Pi-Helices (I) and Beta-Bridges (S). This imbalance often biases machine learning models, leading to poor predictive performance for minority classes. In this study, we address this issue using an oversampler, a technique that balances the dataset by generating synthetic instances for minority classes. Combined with a Bidirectional Long Short-Term Memory (Bi-LSTM) model, capable of capturing long-range dependencies in protein sequences, our approach improves prediction performance across all structural classes. Experimental results across SST3 and SST8 demonstrate significant improvements in accuracy, precision, and class-specific recall, establishing the effectiveness of our proposed methodology.