Learning probabilistic models of hydrogen bond stability from molecular dynamics simulation trajectories

Area of Science:

• Computational Biology
• Biophysics
• Structural Biology

Background:

• Hydrogen bonds (H-bonds) are crucial for protein structure formation and stability.
• H-bond dynamics are essential during protein conformational changes.
• Predicting H-bond stability solely by energy may be insufficient.

Purpose of the Study:

• To develop and evaluate protein-independent probabilistic models for H-bond stability prediction.
• To assess the performance of inductive learning methods in modeling H-bond stability.
• To compare the predictive power of these models against energy-based approaches.

Main Methods:

• Utilized inductive learning to train models on molecular dynamics (MD) simulation trajectories.
• Incorporated 32 attributes describing H-bonds and their local environments.
• Employed regression trees to model the probability of H-bond presence over time.

Main Results:

• Trained models using 6 MD trajectories, analyzing over 4000 distinct H-bonds.
• Achieved approximately 20% improvement in H-bond stability prediction compared to energy-only models.
• Successfully identified a high fraction of the least stable H-bonds, with 80% accuracy in identifying the bottom 10%.

Conclusions:

• Inductive learning provides effective protein-independent models for H-bond stability.
• Probabilistic models significantly outperform H-bond energy alone in predicting stability.
• These models enhance understanding of protein dynamics and structure-function relationships.