Improved Protein Real-Valued Distance Prediction Using Deep Residual Dense Network (DRDN).

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Three-dimensional protein structure prediction remains a significant challenge in bioinformatics.
• Accurate real-valued distance prediction is crucial for determining unique protein structures.

Purpose of the Study:

• To propose a novel deep residual dense network (DRDN) for predicting protein real-valued distances.
• To leverage features from protein sequences and homologous sequences for enhanced prediction accuracy.

Main Methods:

• Utilized a deep residual dense network (DRDN) for real-valued distance prediction.
• Extracted features from query protein sequences and retrieved multi-aligned homologous sequences from five databases.
• Employed DeepMSA, HHblits, and HITS_PR_HHblits for homologous sequence retrieval.

Main Results:

• Achieved evaluation metric scores of 3.89 (Absolute Error), 0.23 (Relative Error), 0.45 (PDA), and 0.63 (PDT).
• Contact maps computed based on predicted distances showed high precision, with top L/5 long-range contact prediction precision of 0.834 on CASP13.
• Achieved an average precision of 0.847 for top-L/5 contact prediction on CASP14.

Conclusions:

• The proposed DRDN method demonstrates significant improvements in real-valued distance prediction for 3D protein structure.
• The method's performance in contact map prediction surpasses several existing state-of-the-art tools on CASP datasets.
• This approach offers a promising advancement for accurate and efficient protein structure modeling.