Multiple Protein Subcellular Locations Prediction Based on Deep Convolutional Neural Networks with Self-Attention...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein subcellular location prediction is vital for understanding protein function, disease diagnosis, and drug development.
• Existing methods struggle with feature representation and redundancy, especially for multi-label classification.
• Accurate prediction remains challenging due to complex feature extraction and data imbalance.

Purpose of the Study:

• To develop a novel deep learning method for predicting protein subcellular locations, including single and multiple sites.
• To enhance feature extraction and representation for improved prediction accuracy.
• To address challenges in multi-label classification and data imbalance in protein localization.

Main Methods:

• Integrated features from pseudo amino acid, amino acid index distribution, and physicochemical properties.
• Employed deep convolutional neural networks (CNNs) for high-dimensional feature extraction.
• Utilized self-attention mechanisms, a customized loss function, random k-label sets, and hybrid over/under-sampling techniques.

Main Results:

• The proposed deep CNN model achieved superior accuracy compared to three alternative classification methods.
• The method effectively extracts high-dimensional features, mitigating redundancy from preliminary features.
• The model demonstrates robust performance in predicting both single and multiple protein subcellular locations.

Conclusions:

• The novel deep CNN approach offers a significant advancement in protein subcellular location prediction accuracy.
• The integrated feature representation and deep learning architecture effectively capture complex biological information.
• This method provides a more reliable tool for bioinformatics research, aiding in disease and drug discovery efforts.