A normalized differential sequence feature encoding method based on amino acid sequences.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein interactions are fundamental to cellular processes like apoptosis and immune response.
• Accurate prediction of protein interactions is crucial for understanding cellular functions and disease mechanisms.

Purpose of the Study:

• To develop and evaluate an optimized coding method for human protein interaction prediction.
• To compare the performance of ensemble learning methods using different dimensionality reduction techniques.

Main Methods:

• A novel coding method, normalized difference sequence characteristics (NDSF), was developed using amino acid sequence positional and correlational features.
• Principal Component Analysis (PCA) and Local Linear Embedding (LLE) were employed for dimensionality reduction of the NDSF-coded protein sequences.
• Four ensemble learning algorithms (AdaBoost, Extra trees, LightGBM, XGBoost) were assessed for classification performance.

Main Results:

• The NDSF coding method demonstrated higher accuracy and reduced loss and coding time compared to the MOS method.
• Linear dimensionality reduction (PCA) yielded significantly better classification accuracy than nonlinear reduction (LLE).
• The XGBoost classifier, utilizing NDSF features reduced by PCA, achieved a peak accuracy of 99.2%.

Conclusions:

• NDSF combined with PCA and XGBoost presents a highly effective strategy for classifying human protein interactions.
• This approach offers improved accuracy and efficiency for protein interaction prediction in bioinformatics.
• The findings highlight the importance of feature coding and dimensionality reduction in machine learning for biological data analysis.