An improved method to detect correct protein folds using partial clustering.

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein structure prediction

Background:

• Structure-based clustering is vital for selecting correct protein folds from predicted decoys.
• Traditional clustering methods struggle with large decoy sets due to poor runtime performance.

Purpose of the Study:

• To develop a more efficient partial clustering approach combined with an improved scoring scheme.
• To enhance both the speed and performance of protein decoy selection methods.

Main Methods:

• Proposed a novel scheme for rapid, incomplete clustering of protein decoys.
• Detected structurally similar decoys using C(α) RMSD or GDT-TS scores, extracting representatives without full cluster assignment.
• Integrated the clustering strategy with various scoring functions.

Main Results:

• The HS-Forest method demonstrated improved correct fold detection rates on Rosetta and I-TASSER decoy sets.
• Outperformed two recent clustering methods, Durandal and Calibur-lite, in accuracy.
• Achieved up to 22 times faster performance, enabling handling of much larger decoy sets.

Conclusions:

• HS-Forest offers superior correct-fold selection by avoiding exhaustive clustering.
• The method's enhanced speed, efficiency, and performance aid researchers in handling larger decoy sets.
• Facilitates significant improvements in ab initio protein structure prediction.