FragHMMent

FragHMMent predicts residue-residue contacts from protein sequences to improve ab initio protein structure prediction by emphasizing long-range interactions.

Key Features:

• HMM-Based Approach: Employs a Hidden Markov Model to predict residue-residue contacts from protein sequences.
• Training Inputs: Trains on homologous sequence data, predicted secondary structures, and local descriptors of protein structure.
• Local Neighborhood Library: Uses a comprehensive library of local neighborhoods that capture recurring structural entities and short-, medium-, and long-range interactions to reassemble cores of proteins in the Protein Data Bank (PDB).
• Performance on Diverse Test Sets: Validated on an external set of 606 protein domains with no significant sequence similarity to the training data and on 151 domains with SCOP folds absent from training.
• Accuracy Metrics: Achieves 22.8% accuracy for long-range interactions in novel fold targets and an average accuracy of 28.6% across long-, medium-, and short-range contacts for the top 0.2 x L (where L is sequence length) predictions.

Scientific Applications:

• Ab initio protein structure prediction: Improves modeling of proteins lacking suitable structural templates by providing predicted long-range contacts.
• Structural biology: Enables more accurate structural models to support interpretation of protein architecture and function.
• Drug discovery: Facilitates generation of structural hypotheses for target proteins to inform drug design efforts.
• Evolutionary analysis: Supports exploration of evolutionary relationships among proteins through inferred contact patterns.

Methodology:

FragHMMent applies a Hidden Markov Model trained on homologous sequence alignments, predicted secondary structures, and local structural descriptors derived from a library of local neighborhoods representing short-, medium-, and long-range interactions.