Multiscale conformational sampling of multidomain fusion proteins by a physics informed diffusion model.

Area of Science:

• Biophysics
• Computational Biology
• Drug Design

Background:

• Multidomain fusion proteins, like bispecific antibodies, require flexible linkers for efficacy.
• Characterizing protein conformational ensembles is vital for rational drug design.
• All-atom molecular dynamics (MD) is computationally expensive for large-scale protein motion simulation.

Purpose of the Study:

• To develop a rapid and accurate method for sampling protein dynamics in flexible multidomain architectures.
• To overcome limitations of generic diffusion models in capturing biophysical constraints for large-scale dynamics.
• To accelerate the rational design of fusion protein therapeutics.

Main Methods:

• Trained a multiscale diffusion framework using an Equivariant Graph Neural Network (EGNN) on microsecond MD trajectories.
• Employed a coarse-grained spatial graph model, condensing rigid domains and preserving linker resolution.
• Integrated biophysical rules into the training objective and inference process.

Main Results:

• Generated high-fidelity conformational ensembles that reproduce long-timescale MD thermodynamic distributions.
• Demonstrated a physics-informed approach for stable and scalable protein dynamics characterization.
• Achieved rapid multiscale characterization of flexible biologics.

Conclusions:

• The developed framework significantly accelerates the characterization of flexible biologics.
• This physics-informed deep learning approach enhances rational drug design for fusion protein therapeutics.
• The model offers a computationally efficient alternative to traditional MD for large-scale protein dynamics.