CGBack

Coarse-grained (CG) molecular dynamics buys reach at the cost of resolution: by collapsing groups of atoms into single interaction beads, CG force fields make million-particle systems and microsecond trajectories tractable, but they discard exactly the atomic detail — hydrogen bonds, side-chain rotamers, stereochemistry — that most downstream analysis depends on. Recovering that detail, a step called backmapping, is an underdetermined inverse problem: many all-atom configurations map onto the same bead positions, and a reconstruction that is geometrically plausible but sterically or stereochemically wrong will destabilize any atomistic simulation started from it.

CGBack, from the Computational Biophysics Research Team at RIKEN R-CCS and the Theoretical Molecular Science Laboratory at the RIKEN Pioneering Research Institute, attacks this with a denoising diffusion probabilistic model coupled to an SE(3)-equivariant graph neural network. Rather than treating a whole chain at once, it factorizes the problem per residue: each training and inference instance is a local graph built around one target Cα, and the diffusion model denoises randomly initialized heavy-atom coordinates into place while the surrounding context stays fixed. That locality is what lets the same fixed checkpoint run on a single small protein and on a densely packed protein condensate without retraining.

The framework pairs generation with an explicit refinement stage that detects and corrects ring penetrations, chirality inversions, and steric clashes, with optional OpenMM energy minimization under Cα positional restraints. It is distributed as a pip-installable package, cgback, in three pretrained sizes.

#Key Features

• Per-residue diffusion with local context: Each residue is reconstructed from a graph of neighboring Cα atoms within an 8 Å cutoff, making cost linear in system size rather than quadratic in chain length.
• SE(3)-equivariant ClofNet backbone: Switching from an E(3)-equivariant GNN to a ClofNet-based architecture nearly eliminated the mirror-image chirality artifacts that reflection-invariant networks produce.
• Integrated refinement pipeline: A post-generation pass removes ring penetrations and corrects chirality outright, with optional OpenMM minimization to clear remaining clashes and hand off simulation-ready coordinates.
• Three checkpoint sizes: S, M, and L differ only in GNN depth (2, 3, and 4 layers), trading runtime against reconstruction quality for high-throughput versus high-accuracy workflows.
• Condensate-scale throughput: Applied to TAR DNA-binding protein 43 (TDP-43) low-complexity-domain droplets of increasing radius, with both sampling and refinement time scaling linearly past a million residues.

#Technical Details

CGBack was trained on the 100% thinning of the CASP12 subset of SidechainNet, an extension of ProteinNet carrying high-resolution atomic coordinates. After filtering to residues with complete heavy-atom coordinates and both sequence neighbors present, the training set comprised 21,649,653 residues, each a single datapoint. Node features encode atom type (36 categories), residue type (20 amino acids), connectivity role (previous, central, next, or unconnected), and 3D coordinates; coordinates are centered on the target Cα with no global alignment. All three released models use a 8 Å neighborhood cutoff, 20 diffusion steps, and a 128-dimensional embedding, trained for 100 epochs with Adam.

On the SidechainNet CASP12 test split (39 proteins, 10 generated structures each), the L model reaches a bond score of 99.86%, clash 2.28%, ring penetration 1.11%, and chirality 0.013% before refinement; the M model scores 99.84%, 2.68%, 1.27%, and 0.036%. After refinement, M reports 0.00% clashes, 0.00% ring penetration, and 0.000% chirality errors. Against baselines on the same split, the deterministic cg2all achieves the best raw bond score (99.98%) and PULCHRA 98.60%, while the flow-matching generative method FlowBack shows 3.687% chirality errors with its Euler solver, reduced to 0.229% with its chirality-corrected solver.

#Applications

CGBack targets multiscale simulation workflows where CG sampling identifies representative conformations or rare-event transitions and atomistic detail must then be restored for energetic analysis, ligand placement, or a follow-on all-atom MD run. Because it handles multi-chain and densely packed systems, it is directly usable on biomolecular condensates and liquid–liquid phase separation models that rule-based backmapping tools struggle with, and its linear scaling makes it practical for initializing large atomistic simulations from mesoscale CG configurations.

#Impact

Released under an MIT license with pretrained weights shipped in the PyPI package, CGBack lowers the barrier to the CG-to-all-atom transition that has long been a manual, tool-specific step in multiscale modeling, and its refinement stage addresses the stereochemical failure modes that make generative backmapping risky in practice. The scope is deliberately narrow: the released models handle proteins under a Cα-only CG representation, with lipids, nucleic acids, glycans, and other mapping schemes such as Martini named as future work. The diffusion process itself is not equivariant even though the network is, which is why a small residual chirality error survives sampling and has to be repaired by refinement.