CGRig: a rigid-body protein model with residue-level interaction sites for long-time and large-scale protein assembly simulation

The paper introduces CGRig, a rigid-body protein model that integrates residue-level interaction sites with overdamped Langevin dynamics to enable efficient, large-scale simulations of protein assembly while preserving essential structural specificity and achieving performance comparable to all-atom accuracy.

The Gist

Imagine you are trying to watch a movie of a massive crowd of people (proteins) dancing, bumping into each other, and forming dance couples in a huge ballroom.

The Problem:
In the world of computer simulations, scientists usually try to watch every single atom in every person's body to see how they move. This is like trying to film the movie by tracking every single hair, fingernail, and speck of dust on every dancer. It's incredibly accurate, but it's so slow and computationally expensive that you can only watch a few seconds of the movie, or maybe just a tiny corner of the ballroom. You can't see the whole dance floor or the long-term patterns of how the crowd forms.

On the other end of the spectrum, some scientists simplify the dancers into just floating balls. This is fast! You can watch the whole ballroom for hours. But there's a catch: balls are round. Real people have arms, legs, and specific hands that need to grab specific partners. If you turn everyone into a ball, they lose their ability to "recognize" each other. They can't form the specific, complex dance couples that real proteins do.

The Solution: CGRig
This paper introduces a new tool called CGRig. Think of CGRig as a clever compromise.

Instead of tracking every atom (too slow) or turning proteins into featureless balls (too dumb), CGRig treats each protein as a rigid, solid Lego brick.

• The Brick: The protein is one solid piece that doesn't wiggle internally. This lets the computer take huge "steps" in time, speeding up the simulation massively.
• The Stickers: But here's the magic: The authors put "stickers" on specific spots of the Lego brick. These stickers represent the chemical "hands" of the protein (amino acids).

How it Works (The Analogy):
Imagine you have a box of rigid Lego bricks. Each brick has specific colored stickers on it.

1. The Shape Matters: The bricks aren't perfect spheres; they are shaped like the actual proteins. This means they bump into each other in realistic ways.
2. The Stickers: If a red sticker on Brick A matches a blue sticker on Brick B, they snap together. If they don't match, they bounce off.
3. The Physics: The computer uses a special math formula (Langevin dynamics) that accounts for how the shape of the brick affects how it spins and slides through the "water" (the solvent). It knows that a long, thin brick spins differently than a round one.

What They Found:
The researchers tested this new "Lego" system in three ways:

1. The Solo Dance: They watched a single protein floating alone. CGRig predicted exactly how fast it would spin and drift, matching real-world experiments perfectly. This proved the "shape physics" was correct.
2. The Dance Couple: They tried to make two proteins find each other and hold hands.
   • Old "ball" models failed; the proteins just bounced off or stuck together in the wrong way.
   • Old "atom-by-atom" models were too slow to watch the whole process.
   • CGRig succeeded: The proteins drifted around, found each other, and locked into the exact correct pose, just like in real life. It even calculated how fast they found each other, which matched other high-end simulations.
3. The Massive Ballroom: They simulated 1,000+ of these Lego bricks trying to build a tower (like microtubules in our cells).
   • Speed: It was incredibly fast. They could simulate 17 microseconds of time in just one day of computer time. For comparison, the old "atom-by-atom" method might take a supercomputer a year to do that much.
   • Result: The bricks spontaneously organized themselves into the correct structures, showing that the model captures the "chemistry" of the stickers even at this massive scale.

Why This Matters:
CGRig is like upgrading from a slow-motion, high-definition camera that can only film one person, to a fast, wide-angle drone that can film a whole stadium of people dancing, while still seeing who is holding hands with whom.

This allows scientists to finally simulate large-scale biological events—like how viruses assemble, how cells build their internal skeletons, or how drugs might clump together—over long periods of time, without losing the crucial details of how the molecules actually fit together. It bridges the gap between "too slow to be useful" and "too simple to be accurate."

Technical Summary

1. Problem Statement

Molecular Dynamics (MD) simulations are essential for understanding biomolecular functions but face a fundamental trade-off between spatiotemporal scale and structural resolution:

• All-Atom (AA) MD: Offers high resolution but is computationally expensive, limiting simulations to microsecond timescales and small systems (single proteins or small complexes).
• Coarse-Grained (CG) Models: Reduce computational cost by mapping multiple atoms to single particles. However, extreme simplifications (e.g., representing proteins as single spherical particles) discard critical structural information, such as molecular shape anisotropy and residue-specific interaction details. This makes them inadequate for studying specific protein-protein recognition and assembly mechanisms.

The core challenge is to develop a model that bridges this gap: enabling millisecond-to-second scale simulations of large systems while retaining residue-level specificity and anisotropic shape information necessary for accurate assembly modeling.

2. Methodology

The authors introduce CGRig, a novel coarse-grained framework that treats each protein as a single rigid body with embedded residue-level interaction sites.

A. Rigid-Body Dynamics & Friction

• Rigid Body Approximation: Proteins are treated as rigid entities, eliminating high-frequency internal vibrations (bond stretching/angle bending), which allows for significantly larger integration timesteps.
• Overdamped Langevin Dynamics: The translational and rotational motions are governed by the overdamped Langevin equation.
• Full Friction Matrix: Unlike simplified models that assume spherical symmetry, CGRig utilizes a 6×6 full friction matrix ($\Xi$) derived from the protein's specific 3D shape (calculated via US-SOMO). This matrix captures:
  • Anisotropic translational and rotational diffusion.
  • Translation-rotation coupling, which is crucial for accurately modeling the hydrodynamics of non-spherical proteins.

B. Interaction Potential (NELVEX)

The model employs a custom potential energy function called NELVEX (Native contact, Electrostatics, and Volume Exclusion), defined at the residue level (C$\alpha$ positions):

1. Gō-like Native Contact Potential ($u_{NC}$):
   • Based on a reference structure (from AA MD).
   • Force-Matching Optimization: Interaction strengths ($H_{ij}$) are not uniform; they are optimized by minimizing the difference between forces in the CG model and reference AA MD trajectories. This allows for both attractive and repulsive forces between native contacts, essential for stability.
2. Debye–Hückel Electrostatics ($u_{ele}$): Models long-range electrostatic interactions between charged residues with a cutoff of 35 Å.
3. Volume Exclusion ($u_{vex}$): A cosine-based repulsive potential for non-native pairs to prevent steric clashes.

C. Implementation

• Implemented as a plugin for LAMMPS.
• Supports CPU multithreading (OpenMP) and GPU offloading (Kokkos).

3. Key Contributions

1. Hybrid Resolution Model: Successfully combines the computational efficiency of rigid-body dynamics with the chemical specificity of residue-level interactions.
2. Hydrodynamic Accuracy: The use of a full 6×6 friction matrix allows for the accurate reproduction of anisotropic diffusion and translation-rotation coupling, which spherical approximations fail to capture.
3. Advanced Parameterization: The NELVEX potential uses a force-matching scheme to derive specific interaction strengths, including necessary repulsive terms within native contacts, overcoming the limitations of standard Gō-like models.
4. Scalability: Demonstrates the ability to simulate systems with >1,000 molecules over microsecond-to-millisecond timescales.

4. Results & Validation

A. Diffusion Validation (Ubiquitin)

• Simulations of isolated ubiquitin showed that CGRig with the full friction matrix accurately reproduced translational ($D_{trans}$) and rotational ($D_{rot}$) diffusion coefficients matching theoretical (US-SOMO) and experimental values.
• Simplified spherical or ellipsoidal approximations failed to capture rotational anisotropy and translation-rotation coupling.

B. Complex Stability (Protein Dimers)

• Tested on 11 protein dimer complexes.
• Comparison: Standard CG potentials (HPS-Urry, KH, Mpipi) caused rapid dissociation of the complexes.
• CGRig Performance: The NELVEX potential maintained native complex structures with high stability (low RMSD).
• Force-Matching Importance: Using a uniform interaction strength (standard Gō-like) resulted in lower stability compared to the force-matched NELVEX parameters. Restricting interactions to only attractive forces also led to dissociation, proving the necessity of the optimized repulsive/attractive balance.

C. Association Kinetics (Barnase-Barstar)

• Simulated association from a dissociated state (35 Å separation).
• Outcome: Proteins spontaneously formed the native complex. The dominant encounter cluster had an RMSD < 1 Å to the native structure.
• Rate Constant: The calculated association rate ($k_{on} \approx 1.56 \times 10^9$ M$^{-1}$s$^{-1}$) agreed well with previous AA MD studies for encounter complex formation, though it was higher than experimental values (likely due to the omission of dehydration/induced-fit steps in the rigid-body model).

D. Large-Scale Assembly (Tubulin)

• Simulated the self-assembly of 16 tubulin dimers.
• Result: The system successfully reproduced the formation of tetramers and longitudinal oligomers, consistent with known biological mechanisms and previous studies. This demonstrated the model's ability to handle anisotropic, specific interfaces in large systems.

E. Performance

• Throughput: Achieved 17.8 $\mu$s/day for a system of 1,024 tubulin subunits using a single GPU.
• Timestep: Stable simulations were possible with $\Delta t = 0.5$ ps (up to 1.0 ps for some systems), a 50–100x increase over typical AA MD timesteps (2 fs).

5. Significance

CGRig represents a significant advancement in computational biophysics by providing a computationally efficient framework that does not sacrifice the structural specificity required to study protein self-assembly.

• Bridging Scales: It enables the simulation of large-scale biomolecular phenomena (e.g., microtubule nucleation, viral capsid assembly) that are currently inaccessible to AA MD and too structurally complex for standard spherical CG models.
• Mechanistic Insight: By retaining residue-level anisotropy and specific interaction potentials, it allows researchers to investigate the mechanistic details of molecular recognition and assembly pathways.
• Future Outlook: While currently limited to rigid bodies (lacking intrinsic disorder flexibility), the framework lays the groundwork for future multi-resolution approaches combining rigid domains with flexible particle-based representations.