DeepDist: real-value inter-residue distance prediction with deep residual convolutional network.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Deep Learning

Background:

• Deep learning has significantly advanced inter-residue contact/distance prediction, crucial for ab initio protein structure prediction.
• Current methods often classify distances into intervals, requiring conversion to real values for tertiary structure prediction.
• This conversion step can introduce inaccuracies and limitations in downstream applications.

Purpose of the Study:

• To develop a multi-task deep learning model (DeepDist) for direct real-value inter-residue distance prediction.
• To evaluate DeepDist's performance against existing methods in both real-value and multi-class distance prediction tasks.
• To explore the utility of simultaneously predicting real-value and multi-class distance maps.

Main Methods:

• Developed DeepDist, a novel multi-task deep learning predictor utilizing residual convolutional networks.
• Trained and tested DeepDist on 43 CASP13 hard domains.
• Compared DeepDist's real-value and multi-class predictions against each other and state-of-the-art methods.

Main Results:

• DeepDist achieved comparable performance in both real-value and multi-class distance prediction tasks.
• The average mean square error (MSE) for real-value distance prediction was 0.896 Ų, outperforming multi-class prediction (1.003 Ų).
• Multi-class predictions from DeepDist showed slightly higher precision for contact prediction (79.3% at L/5) compared to real-value predictions (78.6%).

Conclusions:

• DeepDist effectively predicts inter-residue distances and enhances binary contact prediction.
• Directly predicted real-value distances lead to better protein tertiary structure reconstruction due to lower MSE.
• Simultaneous prediction of real-value and multi-class distance maps yields superior performance compared to predicting real-value distances alone.