Elastic network model of learned maintained contacts to predict protein motion

Area of Science:

• Computational Biology
• Protein Dynamics
• Machine Learning Applications

Background:

• Traditional elastic network models (ENMs) assume static protein contact topology.
• This assumption limits their ability to simulate localized functional motions with significant topological changes.
• Existing ENMs are best suited for highly collective protein movements.

Purpose of the Study:

• To introduce lmcENM, a novel elastic network model incorporating machine learning.
• To enable accurate prediction of protein motion, including localized movements with changing contact topology.
• To improve the general applicability of elastic network models.

Main Methods:

• Developed lmcENM, a machine learning-enhanced elastic network model.
• Implemented a machine learning approach to differentiate between breaking and maintained protein contacts.
• Validated lmcENM against classical ENM and three variants on diverse protein datasets.

Main Results:

• lmcENM accurately captures functional protein transitions unexplained by classical ENM.
• The model preserves the computational simplicity of traditional ENMs.
• Effectiveness demonstrated across a wide range of proteins and motion types.
• Accurate prediction of deformation-invariant contact topology enhances ENM applicability.

Conclusions:

• Predicting dynamic contact topology is crucial for advancing ENM capabilities.
• A combination of features, potentially protein-specific, is relevant for accurate topology prediction.
• lmcENM offers a promising, simple, yet powerful approach to protein dynamics simulation.