Deep Learning Approaches for Protein Secondary Structure Prediction

This paper presents a novel hybrid model comprising Evolutionary Scale Modeling (ESM), Convolution Neural Network (CNN) and Long Short Term Memory (LSTM) network for prediction of protein secondary structures from coil (C), helix (H), and sheet (E)—from amino acid sequences using deep learning techniques. Each architecture leverages unique strengths, with LSTMs capturing long-range dependencies, CNNs extracting local spatial patterns, and ESM enhancing contextual understanding of sequences. The hybrid model was trained and tested using two key datasets: the UniProt dataset and the pdb-intersect-pisces dataset, which provide a rich source of protein sequences and structural information. The proposed model achieved an accuracy of 89.22%, demonstrating robust performance in protein secondary structure prediction.