Boltz Multimer¶
Predicts the 3D structure of a protein complex (binder + target) using Boltz multimer modeling. It takes MSA inputs for both partners, optionally constrains the interface via residue indices, and returns ranked PDB structures with confidence metrics. Deprecated: Use Boltz2 Multimer for enhanced features and future support.
Usage¶
Use this node when you have separate MSA results (A3M) for a binder protein and a single target protein and want to generate complexed structures. Typical workflow: perform MSA searches for binder and target, feed both A3Ms here, optionally provide interface residues (target_specs) to guide binding, then consume the returned PDBs and confidence scores for downstream selection or analysis.
Inputs¶
| Field | Required | Type | Description | Example |
|---|---|---|---|---|
| binder_a3m | True | A3M | MSA search results for the binder protein. Provide as a dictionary {sequence_id: a3m_content}. Each binder sequence will be paired with the single target to form a complex. | {"binder_seq_1": ">seq\nAAAA...\n>...\n"} |
| target_a3m | True | A3M | MSA search results for the target protein. Must contain exactly one entry: {target_id: a3m_content}. | {"target_seq": ">seq\nBBBB...\n>...\n"} |
| recycling_steps | True | INT | Number of recycling iterations used during prediction. Higher values can improve accuracy but increase compute time. | 10 |
| diffusion_samples | True | INT | How many diverse structure samples to generate via diffusion. More samples increase diversity and runtime. | 5 |
| target_specs | True | STRING | Comma-separated residue indices on the target chain to define or bias the binding interface. Leave empty for no constraints. | 1,2,3,45,76 |
| seed | True | INT | Base random seed for reproducibility. If multiple binder entries are provided, each subsequent one uses an incremented seed. | 42 |
Outputs¶
| Field | Type | Description | Example |
|---|---|---|---|
| folding.pdb | PDB | Ranked complex structures. Returned as a dictionary mapping identifiers to PDB text. Keys combine the binder sequence ID and a model rank identifier. | {"binder_seq_1_rank_1": "ATOM ...\nEND"} |
| confidence_scores.json | JSON | Confidence and quality metrics for each generated structure (e.g., per-model scores). Keys align with the structure identifiers. | {"binder_seq_1_rank_1": {"score": 0.78, "metrics": {"plddt": 72.3}}} |
Important Notes¶
- Deprecated: This node is deprecated in favor of Boltz2 Multimer, which adds constraints, templates, multiple entities, and affinity prediction.
- Single target required: Only one target A3M is supported. Supplying multiple targets will cause an error.
- Input format: Both binder_a3m and target_a3m should be dictionaries of A3M contents keyed by sequence IDs.
- Interface specification: target_specs must be a comma-separated list of integers (residue indices on the target). Leave empty to run unconstrained.
- Runtime vs quality: Increasing recycling_steps and diffusion_samples improves thoroughness and diversity but increases runtime.
- Seeding behavior: The seed is incremented per binder entry to ensure unique sampling across multiple binders.
Troubleshooting¶
- Multiple targets error: If you see an error about unsupported multiple target A3Ms, ensure target_a3m contains exactly one entry.
- Invalid target_specs: If parsing fails, verify target_specs is either empty or a comma-separated list of integers (e.g., 5,12,29).
- Missing outputs or empty results: Check that A3M inputs are non-empty and properly keyed. Ensure binder and target sequences are valid and not excessively short.
- Long runtimes: Reduce recycling_steps and diffusion_samples to shorten execution time.
- Identifier mismatches: If downstream steps cannot match outputs, confirm that you are using the returned keys from folding.pdb and confidence_scores.json.
- Low confidence models: Increase diffusion_samples to explore more structures, or refine constraints via target_specs if appropriate.