Assessment of comparative modeling in CASP2

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• The 1996 Critical Assessment of Structure Prediction (CASP2) evaluated comparative modeling submissions.
• Nine of twelve target protein structures were experimentally determined (8 by X-ray crystallography, 1 by NMR spectroscopy).
• Target difficulty varied widely, assessed by sequence identity to known parent structures (20%–85%).

Purpose of the Study:

• To assess the accuracy of protein structure prediction models submitted to CASP2.
• To evaluate the impact of sequence identity and alignment accuracy on model quality.
• To compare modeling approaches and their effectiveness.

Main Methods:

• Analysis of 62 models (89 coordinate sets) submitted by 19 groups.
• Comparison of predicted structures with experimentally determined target structures.
• Assessment of C-alpha root-mean-square deviations (RMSDs) and side-chain placement accuracy.
• Evaluation of sequence alignment accuracy and its correlation with model errors.

Main Results:

• Model quality generally reflected parental sequence similarity; high similarity yielded C-alpha RMSDs < 1 Å.
• Even at 26% sequence identity, best models showed C-alpha deviations of only 2.2 Å.
• Accuracy decreased significantly for targets below 25% sequence identity, primarily due to alignment errors.
• Geometry of models improved compared to CASP1; side-chain placement correlated with backbone accuracy.
• Refinement methods offered minor improvements; no single modeling method was consistently superior.

Conclusions:

• Sequence alignment accuracy is the primary determinant of protein model accuracy, especially for distantly related targets.
• Manual adjustments and expertise in alignment are crucial for accurate modeling, particularly with low sequence identity.
• While automated methods perform well for easy targets, human expertise remains vital for challenging predictions.