SynthQA

SynthQA evaluates the quality of predicted protein structures using a hierarchical machine-learning architecture that integrates multi-scale features, including energy scores and structural topology, to assess decoy quality.

Key Features:

• Single-model quality assessment: Assesses individual protein decoys without requiring multiple-model consensus.
• Hierarchical machine-learning architecture: Trains models in a hierarchical framework to capture relationships across feature scales.
• Multi-scale feature integration: Integrates diverse features derived from protein structures, including energy scores and structural topology.
• Inter-feature relationship modeling and feature generation: Models relationships between features and generates additional features to improve predictive accuracy.
• Decoy quality prediction and ranking: Predicts and ranks protein decoys to identify models closest to the native conformation.
• Training and validation on CASP datasets: Trained on CASP10–CASP12 and validated on 33 targets from CASP14.
• Comparative performance: Reportedly outperforms traditional machine-learning-based methods and each of the 14 individual features used in evaluation.

Scientific Applications:

• Protein model quality assessment: Evaluates the accuracy of predicted protein structures (decoys) relative to native conformations.
• Decoy selection in structure prediction workflows: Ranks and selects decoys to support identification of near-native models in structure prediction.
• Benchmarking QA methods using CASP datasets: Enables performance comparison using CASP10–CASP12 training and CASP14 validation sets.

Methodology:

Implements a hierarchical machine-learning architecture trained on multi-scale features (including energy scores and structural topology), models inter-feature relationships to generate additional features, and was trained on CASP10–CASP12 with validation on 33 CASP14 targets.