[ Co-Folding a Complex | Conformations | Confidence Model | Weights Update | Run on CPU | Local Boltz Diffusion | Reference ]

The Boltz tools in ICM provide AI-based prediction of biomolecular complexes using the Boltz-2 diffusion model. Boltz-2 can be used to predict protein structures and biomolecular complexes, including protein-ligand complexes.

For more information about Boltz-2 see:

Introducing Boltz-2: Toward Accurate and Efficient Binding Affinity Prediction

In ICM, Boltz can be used for co-folding, where one or more protein sequences and an optional ligand are modeled together, or for Local Boltz Diffusion docking, where an existing receptor structure is used as the starting point for locally flexible docking.

The Boltz co-folding tools are available from:

Homology -> Co-Folding Boltz Model

Note: At least one protein sequence must first be loaded into ICM. If no sequence is loaded, the Co-Folding Boltz Model menu option will be greyed out.

[ Multiple Sequences ]

Co-folding predicts a molecular complex directly from the supplied protein sequence information and ligand structure.

One or multiple protein sequences can be included in the same prediction. This enables Boltz to model multi-chain protein complexes as well as protein-ligand complexes.

First load the required protein sequence or sequences into ICM and select:

Homology -> Co-Folding Boltz Model

The Perform co-folding using Boltz dialog contains the following options:

• Sequence - Select the protein sequence or sequences to include in the prediction. Multiple sequences can be used when modeling multi-chain protein complexes.
• Ligand - Select the ligand to include in the co-folding calculation. The ligand can be supplied as a chemical structure, for example from a SMILES representation loaded into ICM.
• N conformations - Specifies the number of independent Boltz diffusion predictions to generate.
• Run Confidence Model - Calculates pLDDT confidence values for the resulting prediction.
• Force Weights Update - Forces ICM to download or refresh the Boltz model weights before running the calculation.
• Run on CPU - Runs the Boltz calculation on the CPU rather than using GPU acceleration.

Click OK to start the prediction.

The resulting model contains the predicted protein structure or protein complex together with the predicted ligand pose when a ligand is included.

Boltz co-folding is not restricted to a single protein chain.

Multiple protein sequences can be supplied in the same calculation, allowing the relative arrangement of the chains to be predicted simultaneously.

A ligand can also be included so that protein-protein interactions and ligand binding are considered together during the co-folding calculation.

The workflow can therefore be used for:

• Single-chain protein-ligand complexes
• Multi-chain protein complexes
• Multi-chain protein-ligand complexes

[ Diffusion Stack ]

The N conformations setting controls the number of independent Boltz diffusion predictions that are generated.

When N conformations is greater than 0, Boltz performs the specified number of independent diffusion runs. The resulting ICM stack contains the final minimized conformation from each independent run.

For example:

N conformations = 5

produces five independently generated and minimized predictions.

The conformations in this stack are independent predictions and should not be played or interpreted as a trajectory. Each frame represents the final result of a separate diffusion run rather than a successive stage of the same calculation.

Differences in protein orientation, ligand orientation or overall structure between frames are therefore expected.

The diffusion and denoising process itself can be visualized by setting:

N conformations = 0

In this mode, instead of storing independent final predictions, ICM stores the intermediate structures from a single diffusion run.

The resulting stack contains approximately 200 diffusion steps showing the structure progressively changing from the initial noisy state toward the final denoised prediction.

Unlike a stack generated with N conformations greater than 0, this stack can be played to visualize the Boltz diffusion and denoising process.

• N conformations > 0 - The stack contains independent final minimized conformations from multiple diffusion runs. These structures are unrelated and should not be interpreted as a trajectory.
• N conformations = 0 - The stack contains approximately 200 successive diffusion steps from a single run and shows the progressive denoising process.

Enable Run Confidence Model to calculate pLDDT (Predicted Local Distance Difference Test) confidence values for the predicted structure.

The confidence values are calculated:

• Per residue for proteins
• Per atom for ligands

The pLDDT values are scaled from 0 to 1. Higher values indicate that the model has greater confidence in the predicted local structure and atomic placement.

The values are stored in the atom B-factor field of the resulting structure.

To visualize the confidence values, use the representation controls in the Display tab. For example, click and hold the Ribbon representation button and select:

Color by B-Factor

This provides a convenient way to identify regions of high and low confidence in the predicted complex.

The Force Weights Update option forces ICM to download or refresh the Boltz model weights used for the calculation.

This can be useful when updated model weights have become available or when the locally installed weights need to be refreshed.

Under normal circumstances it is not necessary to enable this option for every prediction.

Boltz calculations normally take advantage of GPU acceleration.

Enable Run on CPU to perform the calculation using the CPU instead. This option can be used on systems where a compatible GPU is not available, although calculations will generally be substantially faster using GPU acceleration.

[ LBD Command ]

Local Boltz Diffusion docking (_LBDdock) uses the Boltz diffusion model within the standard ICM docking framework.

Unlike full Boltz co-folding, which predicts the complex from sequence and molecular information, Local Boltz Diffusion starts from an existing receptor structure and diffuses from this structure at a lower noise magnitude.

It therefore provides a form of locally fully flexible receptor docking, allowing both the ligand and the surrounding receptor structure to adapt during docking while retaining information from the starting receptor structure.

Local Boltz Diffusion has several advantages:

• Locally fully flexible receptor docking - Both the ligand and the surrounding receptor region can adapt during the diffusion calculation.
• Reduced receptor size - ICM crops the receptor to the region surrounding the docking box, significantly reducing memory requirements and accelerating the calculation.
• Fewer diffusion steps - Local Boltz Diffusion requires approximately half the number of diffusion steps used for full co-folding.
• Ligand localization - The ligand is maintained within the region of interest defined by the docking project.
• ICM docking integration - The method uses an existing standard ICM docking project as input.
• Browsable results - The resulting structures can be browsed and analyzed similarly to a conventional ICM docking hitlist.

The basic command-line syntax is:

icm _LBDdock <dockingproj> input=<ligands.sdf> nconf=<confs_per_lig> [ output=<out.icb> ]

where:

• dockingproj - An existing ICM docking project defining the receptor and docking region.
• input - Specifies the ligand input file in SDF format.
• nconf - Specifies the number of conformations generated for each ligand.
• output - Optionally specifies the output ICB file.

Local Boltz Diffusion is useful when substantial binding-site adaptation or induced fit may be important while retaining an experimentally determined or previously modeled receptor structure as the starting point.

Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V. R., Getz, N., Portnoi, T., Roy, J., Stark, H., Kwabi-Addo, D., Beaini, D., Jaakkola, T., & Barzilay, R. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv. doi:10.1101/2025.06.14.659707 `