DeepUMQA

DeepUMQA evaluates residue-level protein model quality using Ultrafast Shape Recognition (USR) and deep residual neural networks to improve structural assessment accuracy.

Key Features:

• Ultrafast Shape Recognition (USR): USR describes residue-level topological relationships by calculating the first moment of residue distance sets.
• Deep residual neural network integration: A deep residual neural network integrates USR with 1D, 2D, and voxelization features for comprehensive residue characterization.
• Voxelization and topological complementarity: USR supplements voxelization features to capture detailed residue-level topological information missing from voxel-only representations.
• Performance and validation: Experimental evaluations on CASP13, CASP14, and CAMEO blind tests show improved assessment accuracy and competitive performance relative to ProQ2, ProQ3, ProteinGCN, and the ModFOLD series.

Scientific Applications:

• Protein structure prediction and validation: Provides residue-level quality scores to support evaluation and selection of predicted protein models.
• Model refinement and functional interpretation: Aids refinement of structural models and interpretation of protein function, contributing to contexts such as drug design.

Methodology:

USR features are computed by the first moment of residue distance sets; USR is combined with 1D, 2D, and voxelization features and input to a deep residual neural network; performance was evaluated on CASP13, CASP14, and CAMEO blind-test datasets.