StruCloze

Coarse-grained (CG) simulation is how the field reaches systems that all-atom molecular dynamics cannot — condensates, disordered proteins, large assemblies over long timescales. The price is atomic detail: a CG trajectory reduces each residue to one or a few beads, and recovering the missing heavy atoms afterward, a task called backmapping, is a prerequisite for downstream chemical analysis. A closely related problem appears in experimental structures, where flexible loops are simply unresolved.

StruCloze, developed by Hai-Feng Chen's group at Shanghai Jiao Tong University with Kresten Lindorff-Larsen at the University of Copenhagen, treats these as one problem: both are structure completion from a partial specification, so a single network predicting Cartesian coordinates can serve both. It was posted to bioRxiv in July 2025 and published in the Journal of Chemical Theory and Computation in May 2026.

What distinguishes StruCloze from earlier learned backmappers such as cg2all and GenZProt is scope. Those are trained against a single CG scheme and are effectively protein-only; their rigid-frame and internal-coordinate parameterizations transfer poorly to nucleotides. StruCloze pretrains once on residue center-of-mass representations and covers proteins, RNA, DNA, and protein–nucleic acid complexes, with a short fine-tune adapting it to each additional CG scheme.

#Key Features

• One model for backmapping and inpainting: The same architecture reconstructs all-atom coordinates from CG beads and fills in never-resolved residues, using a masked self-supervised fine-tuning stage for the latter.
• Nucleic acid support: Direct supervision on atomic coordinates lets the model reproduce ribose puckering, distinguishing C2′-endo from C3′-endo geometries in DNA and RNA, and recover native base-pairing hydrogen bonds — a capability absent from prior learned backmappers.
• Cross-representation transfer: One pretrained model handles Cα/C4′, center-of-mass, MARTINI, IsRNA1, IsRNA2, and CALVADOS-RNA inputs, with 40 epochs of fine-tuning closing the gap on coarser schemes.
• Deterministic and fast: Removing the diffusion sampling loop used by generative structure models makes inference a single forward pass, at the cost of near-identical outputs on repeat runs.
• No MSA dependency: Sequence enters only as a one-hot encoding, since the CG input already supplies a strong structural prior.

#Technical Details

The architecture adapts the structure module of AlphaFold 3 with four changes: the diffusion process is removed, multiple sequence alignments are replaced by one-hot sequence encodings, token-level transformer layers are cut from 24 to 4 in favor of additional atom-level transformers in both encoder and decoder, and the input "noise" is not Gaussian but reference CCD coordinates translated onto the CG bead positions. An atom-level module extracts local atomic context while a residue-level transformer captures global interactions, with cross-level broadcasting linking the two.

Training used 52,570 PDB entries released before September 2021 — 25,352 protein monomers, 10,504 multimers, 3,902 nucleic acid structures, and 12,812 complexes — clustered at 40% sequence identity and cropped to 256 residues. The three-stage pipeline runs COM pretraining (roughly 76 hours on four NVIDIA 4090D GPUs), a 40-epoch representation-specific fine-tune, and a masked fine-tune. Checkpoints for Cα, COM, MARTINI, and masked inference are released. On a held-out set of 8,066 PDB structures (7,310 protein, 78 nucleic acid, 678 hybrid), Cα backmapping reached 1.03 Å heavy-atom RMSD against cg2all's 1.06 Å, with a higher clash ratio (0.24% vs. 0.15%); COM inputs beat cg2all on both metrics, and nucleic acid and hybrid reconstructions stayed below 2 Å. Zero-shot transfer to IsRNA1 (0.81 Å) and IsRNA2 (0.82 Å) was statistically indistinguishable from COM performance, while the coarser CALVADOS-RNA required fine-tuning. Inpainting held below 1 Å RMSD at mask ratios up to 30%.

#Applications

The immediate use is post-processing CG trajectories for chemical analysis. The authors backmapped CALVADOS2 condensate simulations of two IDR systems and recovered the interaction chemistry distinguishing them — charge–charge E–R and E–K contacts driving the phase-separating system, aromatic contacts dominating the dilute-phase one — analysis impossible at bead resolution. On 1,390 ATLAS trajectories, pairwise RMSD within reconstructed ensembles matched the originals, so relative conformational relationships survive reconstruction. The inpainting mode serves structural biologists completing short unresolved segments in crystallographic and cryo-EM models.

#Impact

StruCloze extends learned backmapping to nucleic acids, an underexplored area, and demonstrates that CG-scheme transfer is achievable with brief fine-tuning rather than full retraining — useful as new CG force fields continue to appear. Its limits are explicit in the paper: predictions resemble ensemble-averaged, low-energy conformations rather than sampling dynamics, torsion distributions are narrower than the underlying MD, χ3 and χ4 angles in long Arg and Lys side chains are poorly recovered, highly disordered RNA is reconstructed with crystal-like ribose geometry, and inpainting degrades beyond roughly ten masked residues, where the model reverts to memorized structural patterns. The authors identify generative modeling as the route past these deterministic failure modes. Checkpoints and test data are distributed via Zenodo under CC BY 4.0; the GitHub repository carries no license file, leaving the code's reuse terms unspecified.