View
215
Download
0
Tags:
Embed Size (px)
Citation preview
Hybrid Protein Model Quality Assessment
Jianlin Cheng
Computer Science Department & Informatics InstituteUniversity of Missouri, Columbia, MO, USA
MULTICOM-CLUSTER: Automated Hybrid Quality Assessment
Sequence-Based Prediction:•Secondary Structure•Solvent Accessibility•Contact Map
Model
Matching Scores
Support VectorMachine
Predicted GDT-TS score
ModelEvaluator
MULTICOM: Human Hybrid Quality Assessment
• Download CASP8 QA predictions
• Calculate average predicted quality score of each model
• Rank models by average scores
Meta Analysis
Server Results of Global QualityPredictor Average Per-
Target Correlation
Overall Correlation
Average Loss (GDT-TS)
Initial-Ranking 0.73 0.76 7.3
Hybrid-Refinement
0.88 0.89 6.2
Improvement of correlation and loss is about 15%
Predictor Average Per-Target Correlation
Overall Correlation
Average Loss (GDT-TS)
Initial-Ranking 0.79 0.84 5.2
Hybrid-Refinement
0.89 0.91 4.7
Human Results of Global Quality
Improvement of correlation and loss is about 10%
Conclusions• Single-model approach can put good, but not
always the best models at the top• Score refinement by structure comparison can
improve both ranking and correlation• Better initial ranking leads to better final ranking• A simple average is a very effective meta QA
method• Structure comparison with reference models is a
hybrid, semi-clustering approach