Seder

Seder applies a machine learning-based scoring function to evaluate protein decoys and predict a fitness score quantifying similarity to native protein structures.

Key Features:

• Knowledge-Based Function: Constructs a knowledge-based scoring function from features describing the structure and sequence of entire protein chains.
• Residue-Residue Distances: Captures spatial relationships between amino acid residues.
• Residue-Solvent Distances: Accounts for interactions between residues and surrounding solvent molecules.
• Pairwise Knowledge-Based Potentials: Uses statistical pairwise potentials derived from known protein structures.
• Four-Body Potential: Incorporates a four-body potential to model collective interactions among four residues.
• Fitness Score Prediction: Predicts a fitness score that quantifies how closely a model resembles the native protein structure.
• Machine Learning Integration: Trains a back-propagating neural network on features derived from native structures and decoys to predict fitness scores.
• Implementation: Implemented using the GEneral Neural Network program.

Scientific Applications:

• Protein structure prediction: Scores and ranks protein decoys to aid selection of models that resemble native structures.
• Benchmarking (CASP10): Evaluated in CASP10, where it ranked third overall among approximately 200 groups for top model prediction and second for freely modeled hard targets, and was the only method consistently ranked within the top three across all targets and hard targets.

Methodology:

Constructs a knowledge-based function from chain-level structural and sequence features (residue-residue distances, residue-solvent distances, pairwise knowledge-based potentials, four-body potential) and trains a back-propagating neural network using the GEneral Neural Network program on features derived from native structures and decoys to predict a fitness score.