GitHub | GitLab mirror | Preprint

All-Atom Reconstruction¶

CABS-flex simulations are performed in a coarse-grained representation (C-alpha trace). To make the results useful for further analysis (such as docking, binding site identification, or high-resolution refinement), the representative models are reconstructed to all-atom representation.

Default Reconstruction (cg2all)¶

⬆ Back to top

By default, CABS-flex standalone 3 uses cg2all, a deep-learning-based tool for ultra-fast and accurate all-atom reconstruction from coarse-grained models. This integration ensures that the resulting ensembles are of high structural quality and suitable for immediate downstream applications.

The reconstruction process using cg2all typically handles both the backbone and side-chain atoms, providing a complete structural model.

For more information, see the cg2all publication (Structure, 2024).

Alternative Reconstruction (Modeller)¶

⬆ Back to top

While cg2all is the default, CABS-flex also supports the traditional reconstruction method using Modeller. This process involves: 1. Using the C-alpha positions as constraints. 2. Optimizing the placement of backbone and side-chain atoms based on spatial restraints.

To use Modeller, you must have it installed and a valid license key configured. See Installation for details.

Cyclic and Disulfide Reconstruction (Two-Stage Hybrid)¶

⬆ Back to top

When modeling cyclic peptides (using --backbone-cyclization) or structures with disulfide bridges (using --disulfide-bonds), CABS-flex employs a specialized hybrid reconstruction procedure to ensure that both coordinate accuracy and chemical topology are perfectly preserved.

Depending on your selected --aa-method, the pipeline behaves as follows:

1. Hybrid cg2all + Modeller Pipeline (Default)¶

When --aa-method cg2all is used with cyclization or disulfide constraints, CABS-flex automatically triggers a two-stage hybrid reconstruction:

• Stage 1: All-Atom Generation (cg2all): The C-alpha/coarse-grained trace is reconstructed into an initial all-atom model using cg2all's deep-learning model to ensure high-quality sidechain and backbone placements.
• Stage 2: Topological Patching (Modeller): CABS-flex immediately feeds the cg2all all-atom structure into Modeller (via ca2all with only_cyclization=True and iterations=1). In this step:
  • Modeller applies the official chemical topology patches (e.g., the DISU patch for disulfide bridges or peptide backbone links).
  • It strips any excess atoms (like the HG hydrogens in Cysteines forming a disulfide).
  • It performs a single, rapid iteration to physically "seal" and refine the geometry of the newly formed covalent bonds without disturbing the rest of the high-quality cg2all prediction.

2. Modeller-Only Pipeline¶

When --aa-method modeller is selected, Modeller handles the entire all-atom reconstruction from scratch in a single stage, applying the standard cyclization/disulfide patches throughout its default optimization cycles.

Reconstruction Options¶

⬆ Back to top

You can control the reconstruction process using several command-line flags:

Basic Control¶

• -A, --aa-rebuild [MODE]: Enable all-atom reconstruction (default: on). The mode defines which structures to rebuild:

  • M: rebuild Medoids (model_?.pdb) sequentially (default).
  • C: rebuild Clusters (cluster_?.pdb) in parallel.
  • T: rebuild Trajectories (replica_?.pdb) in batch DCD.
  • A: rebuild ALL of the above.

• --aa-method <ARG>: Set the method for all-atom reconstruction. Options are cg2all (default) or modeller.

• --aa-minimize [BOOL]: Enable vacuum energy minimization of reconstructed all-atom structures using OpenMM with the modern Amber19 forcefield. Single-model medoids are minimized by default, while multi-model trajectories/replica ensembles are minimized only if explicitly requested. Pass false to completely disable.

cg2all Specific Options¶

• --cg2all-representation <REP>: Coarse-grained representation passed to cg2all. calpha for CA-only or calpha-sc for CA plus side-chain pseudoatoms (default: calpha-sc).
• --cg2all-env-prefix <DIR>: Path to the isolated environment containing cg2all dependencies.

Modeller Specific Options¶

• -m, --modeller-iterations <NUM>: Set the number of iterations for the reconstruction procedure in Modeller (default: 3).

Quality of Reconstruction¶

⬆ Back to top

The conversion of Coarse-Grained (CG) models back to all-atom descriptions is a critical step. Recent integration with deep learning-based methods, specifically cg2all (Structure, 2024), has shown significant improvements over traditional spatial restraint methods.

Figure: Comparison of all-atom reconstruction quality using Modeller vs. cg2all (Protein Science, 2024). The deep learning-based cg2all method yields models with better physical realism, fewer steric clashes, and improved secondary structure recovery.

Reconstruction comparison in the CABS-flex pipeline¶

To complement the previously published validation of cg2all, we performed a focused comparison in the context of the CABS-flex reconstruction pipeline. Six structurally diverse protein structures were first converted using the standard CABS coarse-graining procedure and then reconstructed to all-atom representation using cg2all, cg2all followed by OpenMM minimization, or Modeller. No CABS conformational sampling was performed in this comparison, allowing the effects of coarse-grained representation and all-atom reconstruction to be assessed separately from conformational sampling.

Across the tested examples, cg2all showed better or comparable backbone and side-chain reconstruction accuracy than Modeller, while Modeller occasionally produced substantially larger deviations. The most pronounced difference was observed for the disordered 2MUU target. Raw cg2all structures showed high clashscores, whereas subsequent OpenMM minimization reduced steric clashes to low levels with only modest changes in backbone geometry.

Table: Focused comparison of all-atom reconstruction procedures in the CABS-flex pipeline. Cα RMSD, side-chain (SC) RMSD, and clashscore are reported for cg2all, cg2all followed by OpenMM minimization, and Modeller. RMSD values are calculated relative to the corresponding input experimental structures. No CABS conformational sampling was performed.

OpenMM minimization reduced the clashscore from 137–216 for the raw cg2all reconstructions to 0.6–9.5, while increasing Cα RMSD by only approximately 0.34 Å on average. These results support the use of cg2all followed by OpenMM minimization as the default reconstruction procedure for standard CABS-flex outputs.

Peptide-modeling reconstruction comparison¶

We additionally examined whether the updated reconstruction procedure affects peptide-modeling accuracy relative to the MODELLER-based protocol used in our 2024 study. On five representative peptide examples, cg2all followed by OpenMM minimization produced essentially the same mean best-of-10 backbone accuracy as MODELLER (Cα RMSD 2.33 Å versus 2.30 Å). MODELLER gave somewhat lower mean side-chain RMSD (4.51 Å versus 5.09 Å for cg2all+OpenMM). Thus, the updated reconstruction does not systematically improve peptide-modeling accuracy on this small comparison set, but preserves comparable backbone accuracy while providing the current integrated reconstruction and minimization workflow.

Energy Minimization and Refinement (OpenMM)¶

⬆ Back to top

Following the all-atom reconstruction step, CABS-flex can perform an energy minimization to resolve steric clashes, improve clashscores, and relax local geometries. This refinement process is powered by the OpenMM library.

Refinement Protocol¶

• Force Field: The Amber19 force field (amber19-all.xml) is used to model the potential energy of the protein.
• Hydrogen Addition: Hydrogen atoms are automatically parameterized and added to the reconstructed heavy-atom structure.
• Environment: Minimization is performed in vacuum using the NoCutoff nonbonded method to correct local coordinate overlaps.
• Integrator: A Langevin integrator is initialized with the following parameters:
  • Temperature: 300 Kelvin
  • Friction Coefficient: 1.0 ps⁻¹
  • Time Step: 2.0 femtoseconds (0.002 ps)
• Platform: The minimization executes on the CPU platform to ensure stability and avoid driver-level or memory limit issues when multiple reconstruction workers run concurrently.
• Optimization Limit: The energy minimization runs for a maximum of 500 iterations.
• Hydrogen Removal: After minimization, all hydrogen atoms are stripped from the output PDB. OpenMM adds hydrogens internally for energy calculations, but they are removed from the final file to keep output consistent with the heavy-atom structures produced by cg2all.

Handling Single and Multi-Model Structures¶

• Single-Model Medoids: Single-model structures are minimized directly.
• Multi-Model Ensembles: For trajectory replica ensembles, CABS-flex processes the structures frame-by-frame, minimizing each model independently and writing them back sequentially into a multi-model PDB format.

This refinement step is enabled by default for single-model medoid reconstructions and can be toggled using the --aa-minimize flag.

← Flexibility Modes | ⬆ Back to top | Next: Protein Flexibility