Protein secondary structure prediction using a small training set (compact model) combined with a Complex-valued...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein secondary structure prediction (SSP) is crucial for understanding protein function.
• Current SSP methods face limitations in generalization despite advances.
• Errors in SSP impact downstream protein structure prediction pipelines.

Purpose of the Study:

• To develop a novel approach for training SSP classifiers to improve generalization.
• To investigate structural factors contributing to prediction inaccuracies.
• To enhance the accuracy and reliability of protein secondary structure prediction.

Main Methods:

• A heuristic approach selected a compact model of 55 proteins from the CB513 dataset.
• Protein residue states were represented as probability matrices using C-Alpha, C-Beta, Side-chain (CABS) energy calculations.
• A Fully Complex-valued Relaxation Network (FCRN) classifier was trained on the compact model.

Main Results:

• The compact model achieved accuracies of ~81% on G Switch proteins, outperforming existing literature methods.
• Blind testing demonstrated superior generalization compared to traditional cross-validation.
• Analysis revealed hydrogen bond contacts as a source of Coil-Sheet misclassifications.

Conclusions:

• The selection of training data is critical for classifier generalization in SSP.
• Improved methods distinguishing backbone from other hydrogen bonds are needed to reduce Coil-Sheet errors.