Inter-Residue Distance Prediction From Duet Deep Learning Models.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Accurate prediction of residue distances is vital for protein structure reconstruction, interaction prediction, and design.
• Predicting fine-grained distances between residues with long sequence separations remains a significant challenge in bioinformatics.

Purpose of the Study:

• To introduce DuetDis, a novel method for protein inter-residue distance prediction.
• To enhance prediction accuracy by fusing features from diverse genomic and metagenomic databases.
• To improve robustness and reliability compared to existing methods.

Main Methods:

• DuetDis utilizes duet feature sets and a deep residual network with squeeze-and-excitation (SE) for prediction.
• The method integrates features extracted directly or indirectly from whole-genome/metagenomic databases.
• Ensembling models trained on different feature sets minimizes information loss.

Main Results:

• Ensembling different feature sets demonstrably improves prediction performance.
• High-quality multiple sequence alignment (MSA) significantly boosts prediction accuracy during both training and testing.
• DuetDis outperforms 11 peer methods in overall prediction accuracy and reliability.

Conclusions:

• DuetDis offers a more accurate and reliable approach to protein inter-residue distance prediction.
• The method shows enhanced robustness, particularly against shallow multiple sequence alignments.
• Feature fusion and high-quality MSA are critical for advancing residue distance prediction accuracy.