SE-BLTCNN: A channel attention adapted deep learning model based on PSSM for membrane protein classification

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Accurate membrane protein classification is crucial for understanding protein function.
• Traditional experimental methods are time-consuming and costly.
• Existing bioinformatics methods often require diverse data types, which may not always be available.

Purpose of the Study:

• To develop efficient deep learning models for membrane protein prediction using limited data.
• To evaluate the performance of models utilizing only position-specific scoring matrices (PSSM).
• To investigate the impact of channel attention and data preprocessing on prediction accuracy.

Main Methods:

• Developed two deep learning models: BLTCNN (BiLSTM-TextCNN) and SE-BLTCNN (Squeeze-Excitation embedded BiLSTM-TextCNN).
• Both models utilize PSSM as the sole input feature.
• Conducted ablation studies to validate the contribution of BiLSTM and channel attention.

Main Results:

• BLTCNN achieved state-of-the-art average precision of 96.2% on benchmark datasets using only PSSM.
• SE-BLTCNN achieved 95.7% average precision, outperforming BLTCNN after correcting for excessive zero-padding in training data.
• Identified excessive zero-padding as a significant factor affecting SE-BLTCNN performance.

Conclusions:

• Deep learning models, particularly BLTCNN, can effectively predict membrane protein types using only PSSM data.
• The SE-BLTCNN model demonstrates potential for higher accuracy with appropriate data preprocessing.
• These findings offer a computationally efficient alternative to experimental methods for membrane protein classification.