CHpredict

CHpredict predicts weak hydrogen bonding interactions within proteins, specifically Cα-H...O and Cα-H...π interactions, to characterize their contributions to protein stability and function.

Key Features:

• Interaction Prediction: Identifies residues involved in Cα-H...O interactions between backbone Cα atoms and backbone oxygen atoms, and Cα-H...π interactions between backbone Cα atoms and aromatic side-chain π systems.
• Algorithms: Applies artificial neural networks including feed-forward neural networks (FNN) and recurrent neural networks (RNN) for prediction.
• Training Dataset: Trained and validated on a non-homologous dataset of 2298 protein chains with less than 25% sequence identity.
• Performance Dependence: Prediction accuracy varies with donor–acceptor sequence separation (delta D-A), with RNNs outperforming FNNs across tested distances.
• Reported Sensitivity: Maximum sensitivities achieved with RNNs are 51.2% for Cα-H...O interactions at delta D-A = 4 and 82.1% for Cα-H...π interactions at delta D-A = 3.
• Secondary Structure Input: Incorporation of PSIPRED-predicted secondary structure improves RNN performance by 1–3% for both interaction types.

Scientific Applications:

• Structural Analysis: Characterizing protein architecture by mapping weak hydrogen bonds that contribute to stability and function.
• Protein Folding and Stability: Providing insights into folding pathways and stability determinants through predicted weak interactions.
• Drug Design and Molecular Engineering: Informing drug design and protein engineering by identifying weak interactions that influence protein conformation.

Methodology:

Uses artificial neural networks (FNN and RNN) trained and validated on a non-homologous set of 2298 protein chains (<25% sequence identity), evaluates performance as a function of donor–acceptor sequence separation (delta D-A), and incorporates PSIPRED-predicted secondary structure to improve RNN performance by 1–3%.