HybridSeqNet: A Multimodal Approach Incorporating Convolutional and Long Short-Term Memory Networks for Comprehensive Structural Protein Classification

Soumen Sinha, Pranav Sunil, Neha Bharill, Om Prakash Patel

Abstract: In the field of bioinformatics and molecular biology, the accurate classification of proteins based on their sequences plays a crucial role in understanding their functions and interactions. This research paper presents a novel deep learning architecture, named HybridSeqNet, that leverages the strengths of Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) to enhance the accuracy of protein structure classification from sequence data. The proposed HybridSeqNet architecture capitalizes on the hierarchical features captured by CNNs in local regions of protein structure sequences and the temporal dependencies captured by LSTMs in longer-range contexts. This joint architecture synergistically integrates these two neural network paradigms, enabling the model to effectively learn intricate patterns within protein sequences. The dataset used contained more than 4,00,000 protein structure sequences. Various types of protein structures like DNA, RNA, hybrid DNA-RNA, etc. are considered for classification. We also compare our model with various existing models. HybridSeqNet achieves remarkable classification accuracy of 92.86% when compared to traditional models. The architecture’s effectiveness is demonstrated through comprehensive experiments on a huge dataset of protein structure sequences.