FunHunt: model selection based on energy landscape characteristics.

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Protein folding and binding are often modeled as a search for minimum energy states.
• RosettaDock generates low-energy conformations, but distinguishing native from non-native ensembles is challenging, especially with backbone flexibility.
• Understanding the energy landscape near native orientations is crucial for accurate protein complex modeling.

Purpose of the Study:

• To develop a machine learning algorithm (FunHunt) to differentiate native protein complex orientations from other low-energy ensembles.
• To assess FunHunt's accuracy in identifying correct protein complex structures.
• To gain insights into the characteristics of native protein interfaces.

Main Methods:

• Application of a machine learning classifier, FunHunt, to analyze ensembles of low-energy protein conformations.
• Testing FunHunt on a dataset of 52 protein complexes and 12 recent CAPRI targets.
• Evaluating FunHunt's ability to select near-native orientations from models generated by different algorithms.

Main Results:

• FunHunt correctly identified the native orientation for 50 out of 52 protein complexes and all 12 CAPRI targets.
• The classifier demonstrated general applicability by selecting near-native orientations from models not generated by RosettaDock.
• Analysis revealed a significantly larger energy decrease towards near-native orientations compared to others.

Conclusions:

• FunHunt effectively distinguishes native protein complex orientations by analyzing the local energy landscape.
• The observed greater energy decrease towards native interfaces suggests a reason for their stability.
• The FunHunt approach offers a versatile method for protein model selection and can be extended to other tasks like interface design and specificity prediction.