RFQAmodel

RFQAmodel assesses protein structure prediction model quality to identify correct folds among models generated by template-free prediction protocols by applying a Random Forest classifier that integrates existing quality assessment scores and two predicted contact map alignment scores.

Key Features:

• Quality Assessment Framework: A Random Forest-based classifier integrates existing quality assessment scores with two predicted contact map alignment scores to identify correct models, including cases where traditional methods underperform.
• Training and Testing: The classifier was trained on a large, structurally diverse and balanced dataset of protein domains and validated on an independent test set of 244 protein domains.
• Performance Metrics: When ranking models by RFQAmodel scores, the highest-ranking model had the correct fold for 67 targets (52 confirmed accurate), and the method correctly identified incorrect models for 59 targets.
• Comparative Advantage: The method excels at pinpointing correct models when only a few among many are accurate and showed consistent performance on CASP12 and CASP13 free-modelling model sets.
• Iterative Modeling Protocol: Supports an iterative generate-and-evaluate approach in which models are repeatedly generated and assessed with RFQAmodel until a high-confidence correct model is identified.

Scientific Applications:

• De novo structure prediction model selection: Prioritizes and ranks template-free (de novo) protein models to identify correct folds for downstream analysis.
• Resource prioritization for challenging targets: Enables iterative focusing of computational effort on difficult targets by identifying when additional modeling is warranted.
• Structural biology and structural genomics: Improves reliability of predicted structures used in structural biology studies and structural genomics projects.

Methodology:

RFQAmodel uses a Random Forest classifier that integrates existing quality assessment scores with two predicted contact map alignment scores; it was trained on a large, structurally diverse, balanced dataset of protein domains, validated on an independent test set of 244 domains, and applied to rank models (including CASP12 and CASP13 free-modelling sets) with support for iterative generate-and-evaluate cycles.