Proteins and structures · What you can do here
Predict a protein structure from its sequence
Predict or retrieve three-dimensional protein structures from sequence, evaluate per-residue pLDDT and inter-domain PAE confidence metrics, and identify disordered regions or experimental caveats.
What this involves
Obtaining a reliable protein structure begins with checking established repositories before committing to de novo folding. Most targets of interest have empirical entries in the PDB or precomputed models in AlphaFold DB. Querying UniProtKB to resolve names into accessions allows you to retrieve experimental coordinates or models built with full multiple-sequence alignments, prioritising validated structures over fresh single-sequence predictions.
When an engineered sequence, synthetic construct, or novel isoform requires prediction, you can generate coordinates directly using ESMFold or AlphaFold2. De novo single-sequence folding requires proper sequence preparation: targets should be inspected for signal peptides, low-complexity disordered regions, and multi-domain linkers so that long chains can be partitioned appropriately prior to submission.
Interpreting structural output requires analysing confidence scores rather than inspecting atomic coordinates alone. Per-residue pLDDT scores indicate whether side-chain rotamers (>90) or merely backbone topologies (70–90) can be trusted, while predicted aligned error (PAE) confirms whether relative domain orientations are biologically meaningful or arbitrary.
How it goes
- 1Clean and characterise the sequenceInspect and clean input strings with analyze_sequence or resolve accessions in UniProtKB using search_proteins, identifying signal peptides, low-complexity regions, and stripping stop codons.
- 2Look up pre-existing structuresCall lookup_structure against the PDB and AlphaFold DB to retrieve experimental crystallographic entries or precomputed alignment-derived models before running de novo predictions.
- 3Fold uncharacterised sequencesFor novel constructs or unlisted isoforms, generate 3D coordinates using fold_protein (running ESMFold for sequences up to 400 residues) or run AlphaFold2 via run_model.
- 4Segment and interpret pLDDT confidencePartition the chain into pLDDT bands to distinguish reliable side chains (>90) and secondary structures (70–90) from regions of predicted intrinsic disorder (<50).
- 5Assess domain orientation with PAEExamine predicted aligned error across domain interfaces to determine if relative domain packing is reliable before measuring inter-domain distances or docking.
Start it here
Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.
What to watch for
- Never evaluate a structure using mean pLDDT alone, as a high-confidence domain can obscure an intrinsically disordered tail.
- Do not fold a point mutant to observe structural effects, because the structural difference between mutant and wild-type models will be smaller than the prediction error.
- High inter-domain predicted aligned error (above roughly 15 Å) indicates that relative domain arrangement is unverified and distances between domains must not be measured.
- Regions with pLDDT below 50 reflect predicted intrinsic disorder rather than a failed run, and they should not be refolded.
- In experimental entries, missing coordinates indicate disorder in the crystal rather than sequence deletion, meaning distances must never be calculated across coordinate gaps.
What it runs
Models
- ESMFoldMITPredicts all-atom 3D protein structures directly from single amino acid sequences without alignments in seconds.
- AlphaFold2Apache 2.0 (code)Predicts all-atom 3D protein structures from sequence data using multiple sequence alignments.
- Boltz-2MITPredicts 3D complex structures and binding affinities for protein-ligand pairs in a single forward pass.
- AlphaFold DB LookupCC BY 4.0Retrieve pre-computed AlphaFold protein structures and confidence scores across 247 million UniProt entries.
- PDB Structure FetchCC0Retrieves experimentally determined protein coordinates and metadata for any PDB entry.
Databases it can query
nf-core pipelines
Questions
Do I need to install anything to predict a protein structure from its sequence?
No. The research agent does this work on a Linux machine of its own — it installs what the job needs, keeps the files, and reads the results back to you. Long runs continue if you close the page.
When should I predict a structure instead of looking it up?
Use fold_protein only when lookup_structure yields no results, such as for designed sequences, user-pasted constructs, or uncatalogued isoforms. If an entry exists in the PDB or AlphaFold DB, use it instead, as precomputed models incorporate full multiple-sequence alignments.
How are sequences longer than 400 residues handled by ESMFold?
The fold_protein tool enforces a 400-residue limit for ESMFold. Multi-domain proteins exceeding this length must be split at their linkers and folded as discrete domains, or predicted using AlphaFold2 pipelines.
What does a pLDDT score below 50 mean for my model?
A score below 50 is generally a positive prediction of intrinsic disorder, often enriched in proline, serine, and glycine, rather than an algorithmic failure. These segments should be reported as disordered rather than resubmitted for folding.
Can I determine if a protein adopts an open or closed state from a predicted model?
No, a predicted model represents a single static conformation rather than an ensemble or populated equilibrium state. If inter-domain PAE is elevated, the relative orientation of domains is arbitrary and cannot verify conformational states.
Written from the method the research agent follows for this work and the documentation of the tools listed above. Statements citing a figure that does not appear there are removed before publication. Corollary is a non-commercial research demonstration; results are not for clinical use.