Constructing All-Atom Training Simulations for Coarse-Grained Models

A coarse-grained (CG) force field is fit to the results of one or more all-atom (AA) simulations. These AA simulations are the training data, and they serve as the reference that the Coarse-Grained Force Field Builder Panel reproduces when it fits CG parameters. The guidance below describes current best practices for building AA training simulations. The methodology continues to evolve, so treat these as practical recommendations rather than fixed rules.

The goal of a training simulation is to sample the movement of the component molecules well enough to characterize their molecular-level association preferences. This requires exploring a wide variety of relative structural arrangements. Because thorough configurational sampling is computationally demanding, compromises in system size and composition are usually necessary. If the target application involves large-scale emergent structures, training on much smaller systems is usually sufficient. Focus on sampling local association preferences rather than reproducing macroscopic assemblies.

System composition

Include all components

The most reliable and straightforward approach is to include all relevant components in a single training system. You can also train new components incrementally against a pre-existing force field, using the Use force field option and menu in the Coarse-Grained Force Field Builder panel, but co-simulating all components is simpler.

Keep the system small

Keep the system as small as is practical. Systems that exhibit large-scale structure can usually be trained effectively with fewer than 100,000 atoms.

Use high concentrations

Training on dilute solutions is inefficient for capturing association preferences. The underlying interaction strengths are largely independent of concentration, so you can raise solute concentrations to improve sampling.

• Set the overall solute concentration to between 10% and 50% of the solution.
• If a minor component significantly affects system behavior, raise its concentration close to the levels used for the major components. This improves training efficiency for that component.

Fragment large molecules

Splitting molecules that are still significantly larger than most molecules in the system into smaller fragments can improve sampling efficiency. The valence interactions needed to rejoin the fragments are largely insensitive to the environment, and can be trained separately in a quick, dedicated simulation. Some examples:

• Homopolymers: fragment long PEG chains into segments of 6 monomers or fewer.
• Double-helical nucleic acids: fragment into segments of fewer than 15 monomers.
• Single-stranded nucleic acids: fragment into pieces containing all relevant combinations of nearby residues, each fewer than 25 monomers.

Generating trajectories

Simulation length and relaxation

Sample long enough for the system to relax, then continue sampling for a period significantly longer than the timescale of local molecular rearrangement.

• Systems that evolve quickly, such as simple mixtures of rapidly diffusing molecules, may need only tens of nanoseconds.
• Systems that evolve slowly may need microseconds or longer.

Combining independent simulations

Sampling from a single trajectory can be limited. A more effective strategy, especially when several GPUs are available, is to run multiple independent simulations and combine them into one aggregate trajectory for training.

• All independent simulations must have identical compositions and atom ordering. Use the Disordered System Builder Panel to generate the starting configurations with the same settings but a different Set random number seed option and text box value for each run.
• Run each simulation distinctly longer than its relaxation time, so that every run reaches and samples the steady state.

Verifying the steady state

Visual inspection is often enough to confirm that a system has relaxed. For a closer check, compute structural measures such as radial distribution functions over different time windows and compare them. If the sampling looks insufficient, continue the simulation and examine it again.

Processing trajectories for the Coarse-Grained Force Field Builder

Merging trajectories

To combine multiple trajectories that share the same composition into a single training trajectory:

1. Use trj_slice.py Command Help to remove the initial equilibration phase from each trajectory.
2. Combine the truncated, steady-state trajectories with the concatenation mode of trj_merge.py Command Help.

Setting the analysis fraction

The Coarse-Grained Force Field Builder analyzes only the end of each trajectory, controlled by the Analyze last N % of simulation text box. The same fraction is applied to both the AA and the CG trajectories, so the initial portion of each is skipped for relaxation.

• The default is to analyze the last 20% of each trajectory.
• For challenging systems, raise this to 50–75%, and extend the CG simulations used in the fitting iterations beyond the default length.

Because the analysis fraction skips the same portion of the AA trajectory, part of your AA training data is discarded. To keep that data, use trj_slice.py to extract a slice of your combined steady-state trajectory and prepend it to the training trajectory with the concatenation mode of trj_merge.py. Choose a slice that, combined with the AA training trajectory, corresponds to the skipped portion. The analyzed fraction then covers your actual training data.