DynMoCo: a Novel AI Framework to Reveal Modular Substructures of Protein From Molecular Dynamics

Area of Science:

• Biophysics
• Computational Biology
• Structural Biology

Background:

• Protein function is intrinsically linked to their dynamic structural changes.
• Static protein structures offer limited mechanistic insights.
• Molecular dynamics (MD) simulations provide atomistic detail but generate complex, high-dimensional data.
• Traditional analysis methods often miss localized, functionally critical protein motions.

Purpose of the Study:

• To develop a novel deep learning framework, DynMoCo, for analyzing protein dynamics from MD simulations.
• To identify and track dynamic communities of residues or atoms exhibiting coherent motion or functional coupling.
• To provide an interpretable method for understanding complex biomolecular system dynamics.

Main Methods:

• DynMoCo integrates graph convolutional networks and recurrent models for end-to-end dynamic community detection on molecular graphs.
• Proteins are modeled as time-evolving graphs, enabling community detection inspired by social network science.
• The framework identifies spatially grounded substructures and tracks their temporal evolution, incorporating structural knowledge for physical meaningfulness.

Main Results:

• DynMoCo successfully identifies modular substructures within protein domains during simulations.
• The method characterizes conformational rearrangements in response to external forces, demonstrated on integrin systems.
• Analysis of MD data is transformed into interpretable representations of modular dynamics.

Conclusions:

• Proteins function through dynamic, locally coordinated motions that are challenging to analyze from MD data.
• DynMoCo offers a novel deep learning approach to identify and track functionally relevant dynamic communities in proteins.
• This tool enhances the discovery of mechanistic insights into how molecular motions drive biological functions.