Structure prediction produces coordinate hypotheses from sequences and, for some methods, additional molecular context. Choose a method for the system you want to model, then evaluate the evidence at the residue, domain, and interface levels.
Define the structural question
Separate these problems before choosing a tool:
- Monomer folding: What fold does one protein construct adopt?
- Complex prediction: How might multiple molecular entities assemble?
- Antibody or nanobody modeling: What structure is plausible for a specific immune-receptor sequence?
- Peptide modeling: Which conformations are plausible for a flexible or cyclic peptide?
- Conformational sampling: Which alternative states deserve investigation?
If the question is how two existing structures bind, compare folding-based complex prediction with docking tools. The assumptions and outputs differ.
Choose a starting method
This table is a selection guide, not an accuracy ranking. Check each current service form for supported entities, modifications, restraints, and size limits.
| Research setup | Tools to investigate | Main decision |
|---|---|---|
| Protein sequence or protein assembly | ESMFold2, AlphaFold2 | Sequence-only versus alignment-based setup, assembly representation, and available confidence outputs |
| Proteins with ligands or nucleic acids | Boltz-2, Protenix, OpenFold3, Chai-1 | Whether the method accepts every relevant entity and constraint |
| Antibody, nanobody, or TCR structure | Immune Builder | Receptor-specific modeling versus prediction of a target-bound complex |
| Cyclic peptide | AfCycDesign or a co-folding service with appropriate cyclic input support | Explicit representation of backbone closure and the intended bound state |
| Alternative conformations | AlphaFlow and sampling options exposed by applicable models | Whether diversity addresses a defined structural question |
See the full structure-prediction catalog for additional methods. Compare a small number of appropriate methods rather than selecting a winner from incomparable confidence scores.
Prepare the full molecular context
Use the intended sequences, copy counts, chain boundaries, ligands, ions, and nucleotide identities. Include modifications only through supported representations. Check whether a ligand is specified by SMILES, a chemical component identifier, or an uploaded structure.
For alignment-based workflows, review MSA settings and supplied alignments. An MSA carries evolutionary context; it is not interchangeable with a structural template or an experimentally established interface.
Document the basis for each restraint. A hotspot or contact derived from an experiment is different from one chosen because it produces an attractive model. Confirm that selections match the input numbering.
See Preparing inputs for format and identity checks.
Configure a bounded prediction
Start with a tractable system and a defined sample budget. Record model settings, alignment choices, recycles or diffusion samples where exposed, seeds, and restraints.
More samples can explore uncertainty, but selecting the most attractive result from a large search does not independently validate it. Preserve the full candidate context and explain the selection criteria.
Review the credit estimate. Use pilot runs before expanding into multiple constructs, partners, mutations, or candidate libraries.
Interpret confidence and geometry
| Evidence | Useful interpretation | What it does not establish |
|---|---|---|
| Per-residue confidence, such as pLDDT | Local model certainty, with method-specific definitions | Correct binding, activity, or a unique solution conformation |
| Predicted aligned error, where provided | Uncertainty in relative positioning of regions | An experimental error measurement |
| pTM or related global metrics | Model confidence in the overall arrangement | Correct local chemistry throughout the model |
| ipTM or interface metrics | Confidence in relative chain arrangement | Measured affinity or proof of interaction |
| Agreement across methods | Consistency among computational hypotheses | Independent experimental confirmation |
Avoid one universal threshold across tools and molecular systems. Review the native metric definitions, low-confidence regions, inter-domain uncertainty, and interface geometry. Inspect clashes, ligand placement, missing partners, and implausible contacts in the viewer.
For a detailed example of metric interpretation, see Interpreting Chai-1 metrics. If that article’s example settings differ from the current service, follow the current form and outputs.
Why repeated predictions can differ
Boltz-2 and Chai-1 use stochastic sampling: the same amino-acid sequence can produce different coordinates across runs, and multiple samples within one run can differ too. Their sampling and confidence outputs are described in the Boltz prediction documentation and Chai-1 documentation. Submitting the sequence again under a different name is still a new prediction; the name itself does not change the protein’s biology.
First align the structures before comparing them. A different viewer orientation can make identical folds look different. Small changes in loops or side chains can also leave the overall fold effectively unchanged.
Larger differences can indicate uncertainty, especially in flexible regions, domain placement, or interfaces. Alternative conformations are another possible explanation, but variation alone does not demonstrate biologically relevant states or their populations. Check that sequences, copy counts, partners, MSAs, templates, restraints, and settings actually match. Record seeds where exposed; a fixed seed does not guarantee identical outputs across software versions and hardware.
Which prediction should I trust?
Within a run, start with the tool’s native ranking, then inspect confidence and geometry in the region relevant to your question. For a complex, review interface confidence and relative chain placement as well as local folding confidence.
Between Boltz-2 and Chai-1, there is no universally preferred winner for every system. Do not choose solely from the prettier image or compare raw ranking scores as if they were calibrated across models. Look for agreement across samples and methods, plausible geometry, and consistency with experimental evidence. Convergence is reassuring but can still reflect shared modeling biases. When predictions disagree substantially, preserve the alternatives and investigate the uncertain regions before using one downstream.
Decide what comes next
Advance a structure only if it is appropriate for the next question. Docking may require receptor and ligand preparation; simulation needs a compatible parameterized system; design needs correct protected regions and context.
Save predicted coordinates, confidence outputs, exact inputs, settings, and the job ID. A low-confidence segment may motivate a changed construct, new experimental evidence, or alternative modeling strategy rather than simply more samples.
For a practical first run, follow Run your first job. Explore protein structure prediction solutions for related workflows.