Protein Contact Map Prediction Based on ResNet and DenseNet

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in protein science

Background:

• Protein structure modeling is crucial when homologous structures are unavailable.
• Ultradeep residual neural networks (ResNets) are popular for residue-residue contact prediction due to their ability to capture contextual information.
• Accurate contact prediction aids in determining protein three-dimensional structures.

Purpose of the Study:

• To develop a novel deep neural network framework for enhanced protein contact prediction.
• To improve the accuracy of protein contact map prediction compared to existing methods.
• To integrate new input features and network architectures for better performance.

Main Methods:

• A hybrid deep neural network framework combining ResNet and DenseNet was proposed.
• 1D ResNet was utilized for processing sequential features, incorporating Position-Specific Frequency Matrix (PSFM) alongside PSSM, SS3, and solvent accessibility.
• Outer concatenation combined sequential and pairwise features, followed by DenseNet for final processing.

Main Results:

• The proposed framework demonstrated improved prediction accuracy for protein contact maps.
• The novel network architecture and the inclusion of PSFM as an input feature contributed to enhanced performance.
• The method outperformed other popular contact prediction techniques.

Conclusions:

• The combined ResNet and DenseNet framework offers a more effective approach to protein contact prediction.
• The integration of PSFM represents a valuable addition to input features for contact prediction models.
• This advancement contributes to more accurate protein structure modeling, particularly in the absence of homologous templates.