GitHub | GitLab mirror | Preprint

Modeling Workflow¶

CABS-flex standalone 3 provides three main modeling workflows: protein flexibility simulation, peptide modeling, and peptide–protein docking. Each workflow starts from different input data and is designed for a different modeling task.

In the protein flexibility workflow, the input is a protein structure. The simulation samples near-native conformational variability around the starting structure and generates representative models that can be used to analyze local and global flexibility.

In the peptide modeling workflow, the input is a peptide sequence, optionally with secondary-structure information or restraints. This workflow can be used for modeling linear peptides as well as cyclic or disulfide-constrained peptides.

In the peptide–protein docking workflow, the input is a receptor protein structure and a peptide sequence. The peptide is sampled flexibly around the receptor, either globally or with optional contact restraints guiding the search toward a selected binding region.

Despite these differences, all three workflows share a common modeling core. After input preparation, CABS-flex standalone 3 performs CABS coarse-grained simulation, followed by model selection using filtering, scoring, and clustering. Selected representative models or docking poses are then reconstructed to all-atom representation and can be further analyzed or visualized using the available output tools.

In the default mode, each workflow generates a coarse-grained trajectory containing 1000 to 10 000 models, depending on the workflow and settings. The trajectory is then filtered and clustered to select 10 representative outputs: protein conformations in the protein flexibility workflow, peptide structures in the peptide modeling workflow, or peptide–protein docking poses in the docking workflow. These representatives are reconstructed to all-atom representation, giving 10 final all-atom models or poses by default.

The number of generated models, selected representatives, sampling settings, restraint setup, reconstruction options, and analysis outputs can be modified by the user. In addition to final PDB models, CABS-flex standalone 3 provides outputs such as RMSF profiles, contact maps, scoring summaries, molecular visualization files, and generated reports.

The figure below shows the three main workflows and their shared modeling core in CABS-flex standalone 3.

Figure. CABS-flex standalone 3 modeling workflow. Protein flexibility, peptide modeling, and peptide–protein docking workflows use different input data but converge into a shared modeling core. The core workflow includes input preparation, CABS coarse-grained simulation, model selection by filtering, scoring and clustering, all-atom reconstruction, and analysis/visualization. In the default protein flexibility workflow, 10 representative all-atom models are generated from a clustered coarse-grained trajectory.

Computational Cost and Scaling with System Size¶

⬆ Back to top

CABS sampling is the main computational cost for systems in the size range tested here (92–2,400 residues). Across 32 CASP17 targets run with default settings on identical hardware, CABS sampling scaled approximately as t ∝ N^1.37 (R² = 0.96), while total runtime followed a similar trend (t ∝ N^1.13, R² = 0.94). Setup, cg2all reconstruction, and OpenMM minimization were comparatively inexpensive over this size range and showed weaker relationships with residue count. Their runtimes also depended substantially on the individual structure. Thus, for typical systems within this range, the CABS sampling and total-runtime trends provide the most useful practical estimates of computational cost.

These trends should not be extrapolated indefinitely. The largest system tested with CABS-flex to date contained 17,628 residues in 36 chains and required approximately 206.75 hours in total: about 193.5 hours for setup and CABS sampling and approximately 13.25 hours for all-atom reconstruction and subsequent minimization of the ten representative models. Within the all-atom post-processing stage, OpenMM minimization accounted for approximately 98% of the runtime, whereas cg2all reconstruction itself accounted for approximately 2%. The CABS simulation of this system required approximately 51.5 GB of RAM. This example shows that both runtime and memory requirements can become substantial for very large systems.

For very large systems, post-processing may impose additional limitations. In particular, contact-map generation has quadratic memory requirements and can become more memory-demanding than the simulation itself; it can therefore be disabled when necessary. Other analysis and output steps, including RMSF calculation, docking RMSD plots, B-factor and CSV output, and generation of visualization scripts, scale more favorably with system size and are less likely to become limiting. Notebook and HTML report generation may also become impractical for very large systems and has not been systematically tested at this scale. No cg2all-specific size limitation was encountered in our tests: reconstruction and OpenMM minimization completed successfully for all ten representative models of the 17,628-residue system.

Figure. CABS-flex runtime scaling with system size. Runtime of individual CABS-flex stages and total calculation time measured for 32 CASP17 targets ranging from 92 to 2,400 residues using default settings. Dashed lines show power-law fits (t ∝ N^α). CABS sampling dominates the computational cost over this range and shows the strongest relationship with system size.

← Installation | ⬆ Back to top | Next: CABS Model