An Explainable Deep Learning Model for Clathrin Protein Prediction Using a DCT-Enhanced Position-Specific Scoring Matrix

INTRODUCTION: Clathrin protein (CP) plays a vital role in essential cellular processes, including endocytosis and signal transduction. Dysfunctions in CP have been associated with various diseases, including neurodegenerative disorders and cancer, making it a significant focus of biomedical research. METHODS: This study introduces Pred-CLGRUs, a novel computational framework for predicting Clathrin protein using Gated Recurrent Units (GRUs). Pred-CLGRUs utilize a Position-Specific Scoring Matrix (PSSM) in conjunction with the Discrete Cosine Transform (DCT) to enhance feature extraction while reducing noise and redundancy. A potential explanation for this mismatch could be that the Average Block-PSSM feature is derived from the Position-Specific Scoring Matrix (PSSM). The proposed model is trained using four deep learning architectures: GRUs, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Deep Neural Networks (DNNs). RESULTS: The PSSM-DCT input enabled GRUs to achieve 93.33% accuracy, producing a Matthews Correlation Coefficient value of 0.86. Pred-CLGRUs enabled GRUs to achieve an accuracy of 93.33%, sensitivity of 94.44%, and specificity of 92.22%, while maintaining a Matthews correlation coefficient of 0.86. DISCUSSION: We began by calculating PSSM to detect evolutionary changes. The compression method named DCT was applied within each PSSM during the second implementation step. CONCLUSION: The predictive performance can be significantly enhanced if the model can incorporate heterogeneous types of biological information, such as structure, evolution, and function. We plan to design a novel prediction model that combines Capsule Neural Networks (CapsNet) with DeepWalk to obtain structural features.