Ar_NHPred

Ar_NHPred predicts aromatic-backbone NH interactions from protein sequences to identify aromatic residue–backbone NH contacts that influence protein structural stability and function.

Key Features:

• Neural Network Architecture: Two consecutive feed-forward neural networks, each with a single hidden layer, trained using the back-propagation algorithm.
• Evolutionary and Secondary Structure Inputs: Incorporates evolutionary information from multiple sequence alignments obtained via PSI-BLAST and secondary structure predictions from PSIPRED.
• Sequence-to-Structure Network: Predicts aromatic residues involved in Ar-NH interactions from aligned protein sequences.
• Structure-to-Structure Network: Refines interaction predictions by processing outputs of the first network together with predicted secondary structure.
• Position Prediction Network: Identifies the specific position of donor residues within potential interaction fragments.
• Performance Metrics: Validated by five-fold cross-validation with overall accuracy 70.1%, Matthews correlation coefficient (MCC) 0.20, an MCC improvement from 0.12 to 0.20 when including evolutionary and secondary structure information, and a 15.2% improvement over random prediction.
• Dataset Composition: Development and validation used 3,121 segments from 2,298 non-redundant protein structures with less than 25% sequence identity.

Scientific Applications:

• Protein Structure Prediction: Identification of Ar-NH interactions informs tertiary structure modeling and stability analysis.
• Functional Analysis: Mapping aromatic-backbone NH contacts aids interpretation of protein function and active-site or interaction-network roles relevant to drug design and enzyme engineering.
• Evolutionary Studies: Use of PSI-BLAST-derived evolutionary information enables comparative analysis of conserved aromatic-backbone NH motifs across species.

Methodology:

Two consecutive feed-forward neural networks (each with a single hidden layer) trained by back-propagation; inputs include multiple sequence alignments from PSI-BLAST and secondary structure predictions from PSIPRED; implemented component networks are Sequence-to-Structure, Structure-to-Structure, and Position Prediction networks; development and validation used 3,121 segments from 2,298 non-redundant protein structures (<25% sequence identity) and five-fold cross-validation for performance assessment.