Minimizing and learning energy functions for side-chain prediction.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning

Background:

• Protein side-chain prediction is crucial for understanding protein folding.
• Current methods like ROSETTA achieve limited accuracy (72%) for buried residues.
• Improving prediction accuracy requires advancements in search methods and energy functions.

Purpose of the Study:

• To introduce a novel search method and energy function learning approach based on Tree Reweighted Belief Propagation (TRBP).
• To evaluate the effectiveness of TRBP in optimizing protein energy functions and improving side-chain prediction accuracy.
• To explore the synergy between machine learning and approximate inference for state-of-the-art predictions.

Main Methods:

• Developed a novel search method and energy function learning technique utilizing Tree Reweighted Belief Propagation (TRBP).
• Applied TRBP to optimize the ROSETTA energy function, achieving global optima for a majority of proteins.
• Integrated TRBP with Conditional Random Fields (CRF) for learning energy function weights from training data.

Main Results:

• TRBP successfully found global optima for 85% of proteins in a benchmark set within minutes.
• Learning new energy function weights using TRBP and CRF significantly improved side-chain prediction accuracy from 72% to 78%.
• The highest accuracy of 82.6% was achieved using an extended rotamer library and CRF-learned weights.

Conclusions:

• TRBP offers an efficient method for optimizing protein energy functions.
• Combining machine learning (CRF) with approximate inference (TRBP) substantially enhances protein side-chain prediction accuracy.
• The developed approach represents a significant advancement in the field of computational protein structure prediction.