Improving protein fold recognition using triplet network and ensemble deep learning.

Area of Science:

• Computational biology
• Bioinformatics
• Structural bioinformatics

Background:

• Protein fold recognition is crucial for predicting protein structure and function.
• Deep learning (DL) has advanced protein fold recognition, but current methods using bottleneck layers are indirect and inefficient.
• Existing DL methods often rely on suboptimal intermediate representations for fold classification.

Purpose of the Study:

• To develop a novel computational framework for improved protein fold recognition.
• To overcome the limitations of indirect feature representations in current DL-based methods.
• To enhance the sensitivity and accuracy of predicting protein fold types.

Main Methods:

• Developed FoldNet, a DL model using triplet loss to directly optimize protein fold embeddings.
• Implemented FoldTR, a residue-contact-assisted predictor leveraging FoldNet embeddings.
• Proposed FSD_XGBoost, an ensemble DL method combining FoldNet embeddings with features from SSAfold and DeepFR.

Main Results:

• FoldNet directly optimizes protein embeddings, placing similar folds closer in the embedding space.
• FoldTR demonstrated improved protein fold recognition performance.
• FSD_XGBoost achieved a Top 1 sensitivity of 74.8% at the fold level, a ~9% improvement over state-of-the-art methods.

Conclusions:

• Directly optimizing protein fold embeddings is more effective than using intermediate bottleneck layers.
• Ensemble DL methods combining complementary fold-specific features significantly boost recognition accuracy.
• The developed framework offers a substantial advancement in protein fold recognition.