Improved protein contact prediction using dimensional hybrid residual networks and singularity enhanced loss function

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Deep residual learning significantly advances protein contact prediction.
• Accurate prediction of protein contacts is crucial for understanding protein structure and function.

Purpose of the Study:

• To develop a novel deep residual learning model for enhanced protein contact prediction.
• To improve model performance by introducing a hybrid 1D/2D convolutional residual block and a specialized loss function.

Main Methods:

• Developed DRN-1D2D, a protein contact prediction model incorporating a novel hybrid 1D and 2D convolutional residual block.
• Introduced a new loss function that emphasizes easily misclassified residue pairs to improve training.
• Evaluated the model on CASP11, CAMEO, and membrane protein datasets, comparing it against reference models.

Main Results:

• The hybrid residual block and enhanced loss function demonstrably improved protein contact prediction performance.
• DRN-1D2D achieved superior results compared to two in-house reference models on multiple datasets.
• Further validation on CASP13 and CASP14 free modeling targets showed DRN-1D2D outperformed six state-of-the-art models.

Conclusions:

• The proposed DRN-1D2D model, featuring a dimensional hybrid residual block and singularity-enhanced loss function, significantly enhances protein contact prediction.
• DRN-1D2D establishes a new state-of-the-art in protein contact prediction accuracy.
• The study confirms the efficacy of the novel architectural and loss function components for improving predictive performance.