GENESIS CGDYN: large-scale coarse-grained MD simulation with dynamic load balancing for heterogeneous biomolecular systems

Abstract

Residue-level coarse-grained (CG) molecular dynamics (MD) simulation is widely used to investigate slow biological processes that involve multiple proteins, nucleic acids, and their complexes. Biomolecules in a large simulation system are distributed non-uniformly, limiting computational efficiency with conventional methods. Here, we develop a hierarchical domain decomposition scheme with dynamic load balancing for heterogeneous biomolecular systems to keep computational efficiency even after drastic changes in particle distribution. These schemes are applied to the dynamics of intrinsically disordered protein (IDP) droplets. During the fusion of two droplets, we find that the changes in droplet shape correlate with the mixing of IDP chains. Additionally, we simulate large systems with multiple IDP droplets, achieving simulation sizes comparable to those observed in microscopy. In our MD simulations, we directly observe Ostwald ripening, a phenomenon where small droplets dissolve and their molecules redeposit into larger droplets. These methods have been implemented in CGDYN of the GENESIS software, offering a tool for investigating mesoscopic biological processes using the residue-level CG models.

Fig. 1. Domain decomposition scheme in CGDYN.

a The domain decomposition algorithm with the cell-based kd-tree method. b Communication in sending the coordinates from MPI rank 5 to other processes (upper) and receiving the coordinates from other processes to MPI rank 5 (lower). c Updates of subdomains during MD simulations of Hero11 and TDP-43-LCD. The first column displays snapshots of molecular structures at specific MD steps. The remaining three columns on the right illustrate the expansion of subdomains across three dimensions and on the second layer of a 4 × 4 × 4 = 64 domain decomposition. Source data of (c) are provided as a Source Data file.

Fig. 2. Benchmark results of CG MD simulations with three algorithms: ATDYN, SPDYN-like and CGDYN.

a The two benchmark systems with different densities. b Benchmark performance ($\times {10}^6$ steps/day). ATDYN can be used with a small number of nodes, but the parallel efficiency is low, and the performance is saturated from 16 nodes. The SPDYN-like algorithm performs better using many nodes but has a higher dependency on particle density. CGDYN performs better than MD simulations based on the ATDYN and SPDYN-like algorithms, showing a weak dependency on particle densities. Source data are provided as a Source Data file.

Fig. 3. CG MD simulation of the fusion process of two TDP-43-LCD droplets.

The system consists of 1000 chains of TDP-43-LCD. In the initial structure, two separate droplets were put close to each other. The system was simulated at 280 K for ${10}^8$ steps. Snapshots of simulated structures at $t=0$ (a), $t=2\times {10}^7$ steps (c), and $1\times {10}^8$ steps (e). Chains from the two droplets in the initial structure are colored red and blue, respectively. Time-averaged distributions of TDP-43 chains along the $z$ axis during $0<t<0.4\times {10}^7$ (b), $1.8\times {10}^7<t<2.2\times {10}^7$ (d), and $9.6\times {10}^7<t<10.0\times {10}^7$ (f) steps, respectively. g Density of TDP-43 particles along the $z$ axis as a function of simulation time. h Time series of average chain-chain distances $D_{I,J}$ (upper) and shape coordinate $\eta$ (lower, $\eta =3$ indicates perfectly symmetrical sphere and larger $\eta$ values indicate deviations from spherical symmetry). i Droplet mixing (depicted by coordinate $m=\sqrt{D_{\mathrm{1,1}}{D}_{\mathrm{2,2}}}/D_{\mathrm{1,2}}$) against shape change ($\eta$). Detailed definitions of $D_{I,J}$, $\eta$, and $m$ are in the Methods section. Source data are provided as a Source Data file.

Fig. 4. CG MD simulations of the dynamics of multiple TDP-43-LCD droplets at two different densities.

a Snapshots of the high-density system taken at five different times during the simulation. b similar to (a) but for the low-density system. Colors of protein chains in (a) and (b) are according to the DBSCAN clustering results: chains in droplets are colored blue or red, and chains classified as “noise” are colored white. c Number of droplets ($n_d$, upper) and sizes of droplets ($s_d$, lower) as functions of simulation time. In the lower plot, special markers are used for the two droplets in the high-density system (a) after $4.25\times {10}^8$ steps: the markers have yellow edges, with the face color being red for the larger droplet and blue for the smaller droplet, respectively. For the other dots and lines, yellow represents the high-density system, while green represents the low-density system. Source data are provided as a Source Data file.