<![CDATA[Protein secondary structure prediction (PSSP) is an important challenge in bioinformatics. Existing methods for PSSP are generally divided into three categories: neighbor-based, model-based, and meta-estimator-based methods, each using supervised or supervised learning methods model-based are often neural networks, hidden Markov models are available; they support vector machines and other machine learning techniques based on multiple sequence alignments and evolutionary data from increasingly large protein databases. This paper presents a powerful machine learning approach for PSSP, which is a new feature extraction method using sparse autoencoders to identify new protein features. The sparse autoencoder efficiently identifies new features in the training data and provides an accurate prediction of occurrences. Two machine learning methods are used: unsupervised learning methods based on sparse auto-encoders and semi-supervised learning methods using deep learning methods. Experimental results show that the deep learning method gets 86.719% accuracy on the test set, while the unsupervised pretraining method gets 85.853% accuracy on the training set after being improved by surface propagation. Fine-tuning and layer-wise pretraining significantly improve the performance of the proposed model. The results show that the deep learning method achieves an accuracy of 86.7% in the training set and 71.4% in the test set. In comparison, Sparse Autoencoders alone achieved an accuracy of 67%, demonstrating the effectiveness of the combination of these methods. This study highlights the role of advanced deep learning techniques in PSSP accuracy. Future research should consider using big data, exploring deep learning algorithms, and refining optimization methods to further encourage predictive performance in bioinformatics.]]>
