DeepDist

DeepDist predicts real-value inter-residue distances in proteins using a residual convolutional network within a multi-task deep learning framework to improve distance accuracy for tertiary structure modeling.

Key Features:

• Real-value distance prediction: Directly predicts real-value inter-residue distances instead of only classifying distances into discrete intervals.
• Residual convolutional network architecture: Employs a novel residual convolutional network architecture for distance prediction.
• Multi-task learning: Simultaneously performs regression of real-value distances and multi-class distance classification within a multi-task deep learning framework.
• Benchmark evaluation: Evaluated on 43 challenging domains from CASP13.
• MSE performance: Achieved an average mean square error (MSE) of 0.896 Ų for real-value distance predictions when excluding distances ≥16 Å, compared to 1.003 Ų from multi-class distance prediction.
• Binary contact conversion and precision: When converted to binary contacts using an 8 Å threshold, multi-class distance predictions yielded 79.3% precision for top L/5 contacts and 66.1% precision for top L/2 contacts.
• Impact on structure modeling: The dual-task predictions improve binary contact precision and provide direct inputs for tertiary structure reconstruction due to lower MSE of real-value distances.

Scientific Applications:

• Protein tertiary structure reconstruction: Provides real-value inter-residue distances that can be used to reconstruct or model protein tertiary structures.
• Binary contact prediction: Enables binary contact prediction by converting distance predictions using an 8 Å threshold for contact assignment.
• Method benchmarking: Supports benchmarking and performance assessment of distance-prediction methods on CASP13 domains.

Methodology:

Uses a residual convolutional network within a multi-task deep learning framework to simultaneously regress real-value inter-residue distances and classify distances into multiple intervals; reported evaluations include MSE calculations excluding distances ≥16 Å and conversion to binary contacts at an 8 Å threshold.