MFTrans: A multi-feature transformer network for protein secondary structure prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Accurate protein secondary structure prediction (PSSP) is vital for understanding protein function, drug discovery, and disease diagnostics.
• Existing methods face challenges in capturing complex protein sequence features for precise PSSP.

Purpose of the Study:

• To introduce MFTrans, a deep learning-based multi-feature fusion network designed to enhance PSSP accuracy and efficiency.
• To leverage advanced deep learning architectures for improved feature extraction and integration in PSSP.

Main Methods:

• MFTrans employs a Multiple Sequence Alignment (MSA) Transformer combined with a multi-view deep learning architecture.
• It integrates diverse features: MSA, sequence, evolutionary, and hidden state information via multi-feature fusion.
• A hybrid CNN-BiGRU network further refines high-level features post-fusion.

Main Results:

• MFTrans demonstrates superior generalization ability compared to state-of-the-art PSSP models.
• Achieved an average performance improvement of 3% on public benchmarks (CASP12-14, TEST2016/2018, CB513).
• Case studies confirm advanced performance in predicting mutation sites.

Conclusions:

• MFTrans offers a significant advancement in PSSP, improving prediction accuracy and efficiency.
• The model's performance opens new possibilities for drug discovery, disease diagnosis, and protein science research.