BERT-PPII: The Polyproline Type II Helix Structure Prediction Model Based on BERT and Multichannel CNN.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Polyproline type II (PPII) helix structure prediction is vital for understanding protein folding, drug targets, and protein functions.
• Existing prediction algorithms often suffer from insufficient protein sequence feature learning due to single encoding methods.

Purpose of the Study:

• To develop an improved algorithm for predicting PPII helix structures by enhancing protein sequence encoding.
• To leverage the capabilities of BERT and Convolutional Neural Networks (CNNs) for more effective feature extraction.

Main Methods:

• Proposed a BERT-based PPII helix structure prediction algorithm (BERT-PPII).
• Utilized BERT's CLS vector to capture global contextual information from amino acid residues.
• Employed CNN to extract local amino acid residue features, capturing important inter-residue interactions.
• Fused global features from BERT with local features from CNN for enhanced prediction.

Main Results:

• BERT-PPII demonstrated improved accuracy over the PPIIPRED method by 1% (strict) and 2% (less strict) on unbalanced datasets.
• Achieved AUC values of 0.826 (strict) and 0.785 (less strict) on balanced datasets.
• Obtained AUC values of 0.827 (strict) and 0.783 (less strict) on an independent test set.

Conclusions:

• The proposed BERT-PPII method significantly enhances the performance of PPII helix structure prediction.
• The fusion of BERT's global features and CNN's local features effectively captures complex protein sequence information.
• BERT-PPII offers a superior approach for predicting PPII helix structures compared to existing state-of-the-art methods.