Protein complex structure prediction

What is protein complex structure prediction?

Protein complex structure prediction is the computational process of building the three-dimensional arrangement and interfaces of two or more protein chains from their sequences and specified composition. It differs from docking because a complex predictor may predict partner folds and assembly jointly rather than begin with fixed monomers.

The calculation depends on more than individual fold quality. Define biologically relevant chains and stoichiometry, then ask whether interface confidence, model recurrence, known residues, and cellular context support the predicted interface.

A plausible high-confidence interface is not proof of an in-vivo interaction. Incorrect paralogs, missing cofactors, alternative oligomeric states, and familiarity with training evidence can yield neat but misleading assemblies.

When to use protein complex structure prediction

Use complex prediction when the molecular partners and stoichiometry are defined:

• Two or more protein chains have evidence of interaction.
• You want competing interface hypotheses for mutagenesis.
• A native complex can benchmark a prediction protocol.

Strengths and limitations

Joint modeling can place partners in an explicit interface and return model-specific confidence for the chains and their assembly. Independent predictors can show whether a contact region recurs across methods.

Incorrect stoichiometry or missing binding partners can produce convincing but irrelevant interfaces. High monomer confidence does not validate chain placement, and all models can share training bias. The predicted complex must fit known cellular localization and biochemical evidence.

Complex-model interpretation

Multimer predictors place chains using learned geometric and evolutionary information. Chain-pair PAE, ipTM-like values, and interface-residue confidence are generally more relevant than monomer-wide confidence for an assembly.

Compare alternate stoichiometries only when biologically plausible. Larger assemblies gain contacts simply because more chains were supplied, so use co-purification, crosslinking, mutagenesis, or cryo-EM to assess a decisive claim.

Applications of protein complex structure prediction

Complex predictors can propose how known protein partners assemble when isolated chain structures do not specify an interface. Candidate complexes can guide cross-linking, mutagenesis and structural experiments.

They are less informative when stoichiometry, chain identity or interaction state is uncertain. A high-scoring interface formed by omitted domains or incompatible biological localization is not a plausible cellular assembly.

How to run protein complex structure prediction online

1. Define the assembly. Supply the intended chain sequences and stoichiometry, preserving identities and construct boundaries.
2. Run multimer predictors. Use AlphaFold2 Multimer on the same chain set.
3. Compare independent models. Review ESMFold2, Boltz-2 and Protenix outputs as separate complex hypotheses.
4. Inspect interfaces. Check chain placement, clashes, per-chain confidence and any model-specific interface metrics.
5. Validate the assembly. Compare with known cross-links, interaction data or a native complex where one exists.

How to interpret protein complex structure prediction results

High local confidence within each chain does not prove their relative placement. Read interface-specific confidence with its model label and look for agreement on contact regions rather than comparing raw values across predictors.

An interface model can guide experiments, but it does not confirm interaction, affinity, stoichiometry in cells or function.

How protein complex structure prediction works

Run four multimer-capable predictors from the same multi-record FASTA and compare their interface confidence and assembly geometry.

1. Define partners and stoichiometry. Create one FASTA record per chain and duplicate records only when the proposed stoichiometry requires it.
2. Prepare chain sequences. Run AlphaFold2 Multimer, ESMFold2, Boltz-2, and Protenix from the same input.
3. Run multimer models. Retain ranked models and every chain-aware confidence output.
4. Inspect interface confidence. Inspect interfaces, clashes, buried surface, PAE blocks, and agreement across model families.
5. Compare biological support. Prioritize assemblies that agree with biochemical and cellular evidence, then export models and provenance.

Inputs and outputs

Check formats before running, then inspect and download the result from every workflow step.

Inputs

Research input

FASTAPDBmmCIF

A multi-record protein FASTA containing each intended chain with the correct multiplicity.

Outputs

Prediction and review outputs

PDBmmCIFCSVJSON

Ranked complex models, chain-aware confidence metrics, PAE or interface scores, and downloadable structures.