What type of study is this?

September 5, 2025Open Access

Combining knowledge distillation and neural networks to predict protein secondary structure

LZLufei ZhaoLiaocheng University JLJingyi LiFujian Medical University BZBiao ZhangHebei University of Technology

Key Points

The proposed model significantly improves prediction accuracy for protein secondary structures, especially for octapeptides and tripeptides.
Using datasets TS115 and CB513, along with 15,078 protein data points from the PDB database, demonstrates the model's effectiveness and generalizability.
Deep learning models capitalize on complex sequence data to extract patterns that enhance secondary structure prediction accuracy.
Integration of techniques like knowledge distillation contributes to the model's strong performance in protein structure understanding.

Abstract

The secondary structure of a protein serves as the foundation for constructing its three-dimensional (3D) structure, which in turn is critical for determining its function and role in biological processes. Therefore, accurately predicting secondary structure not only facilitates the understanding of a protein's 3D conformation but also provides essential insights into its interactions, functional mechanisms, and potential applications in biomedical research. Deep learning models are particularly effective in protein secondary structure prediction because of their ability to process complex sequence data and extract meaningful patterns, thereby increasing prediction accuracy and efficiency. This study proposes a combined model, ITBM-KD, which integrates an improved temporal convolutional network (TCN), bidirectional recurrent neural network (BiRNN), and multilayer perceptron (MLP) to increase the accuracy of protein secondary structure prediction for octapeptides and tripeptides. By combining one-hot encoding, word vector representation of physicochemical properties, and knowledge distillation with the ProtT5 model, the proposed model achieves excellent performance on multiple datasets. To evaluate its effectiveness, two classic datasets, TS115 and CB513, containing 115 and 513 protein datasets, respectively, were used. In addition, 15,078 protein data points collected from the PDB database from June 6, 2018, to June 6, 2020, were used to further verify the robustness and generalizability of the model. This study improves prediction accuracy and provides an essential model for understanding protein structure and function, especially in resource-limited settings.

AI से पूछें

Bookmark

View Full Paper