iQDeep

iQDeep provides residue-level error classification and probabilistic protein scoring using multiscale deep residual neural networks (ResNets) to assess protein structure prediction quality.

Key Features:

• Multiscale deep residual neural networks (ResNets): Uses multiscale ResNets to analyze structural models at multiple scales.
• Residue-level error classification: Performs residue-level error classifications to identify local inaccuracies in predicted structures.
• Probabilistic score aggregation: Probabilistically combines residue-level classifications into a cohesive global protein score.
• Adjustable error resolutions and GDT estimation: Adjusts error resolutions to estimate both standard and high-accuracy variants of the Global Distance Test (GDT) metric.
• Residue-wise breakdowns and feature agreement metrics: Produces residue-wise score breakdowns and reports agreements between sequence- and structural-level features.
• Benchmark validation: Performance validated against state-of-the-art methods in CASP12 and CASP13 benchmark assessments and blind evaluation in CASP14.

Scientific Applications:

• Protein model quality assessment: Enables self-assessment of predicted protein structures at residue and global levels.
• Benchmarking and method comparison: Supports comparative evaluation of prediction methods using CASP benchmarks (CASP12, CASP13, CASP14).
• Structural modeling support: Provides GDT-based metrics (standard and high-accuracy variants) to inform predictive modeling scenarios in structural biology.

Methodology:

Uses multiscale deep residual neural networks (ResNets) for residue-level error classification, probabilistically aggregates those classifications into a global protein score, adjusts error resolutions to estimate standard and high-accuracy Global Distance Test (GDT) variants, and was validated via benchmarking in CASP12, CASP13, and blind CASP14.