This skill retrieves precomputed models from the AlphaFold Protein Structure Database (AFDB) by UniProt accession and, more importantly, reads their confidence files correctly. An AFDB entry is a single AlphaFold2 prediction of one UniProt sequence modeled as an isolated chain - it has no ligands, ions, cofactors, PTMs, quaternary assembly, or alternative conformations, so it answers "what fold does this sequence adopt on its own?" not "what is the functional state in its complex?" The two confidence signals answer different questions: pLDDT is per-residue local confidence (stored in the B-factor column, opposite polarity to thermal motion), while PAE is per-residue-pair confidence that governs inter-domain placement. Low-pLDDT stretches are usually intrinsically disordered regions, not errors. The skill guides fetching via REST metadata (so version suffixes are discovered, not hard-coded), banding pLDDT, segmenting domains by PAE, and deciding when to trust AFDB versus run a fresh prediction.
pip install biopython numpy requests matplotlib
Tell your AI agent what you want to do: - "Download the AlphaFold model for UniProt P04637 and report mean pLDDT" - "Which regions of this AlphaFold prediction are low confidence, and are they disordered?" - "Read the PAE matrix and tell me if the two domains are confidently placed relative to each other" - "Is this AlphaFold model suitable for docking, or should I run my own prediction?"
"Retrieve the AlphaFold model for UniProt ID P04637 using the current database version"
"This is a 3200-residue human protein - get all AlphaFold fragments for it"
"Fetch the AlphaFold entry and its PAE file for accession Q9Y6K9"
"Report the per-residue pLDDT for this model and classify each residue into confidence bands"
"There is a long low-pLDDT stretch here - is it a disordered region or a modeling error?"
"The mean pLDDT is 88 - can I trust the distance between the N- and C-terminal domains?"
"Load the PAE matrix and identify the confidently predicted domains"
"Are the relative orientations of these two domains reliable, or should I treat them as independent rigid bodies?"
"Segment this multidomain model into blocks I can place independently for molecular replacement"
"Is this AFDB monomer usable as a molecular-replacement search model?"
"I need the ligand-bound state of this enzyme - can AFDB give me that?"
"Should I use the AFDB entry or run AlphaFold3 for this protein-protein complex?"
The prediction metadata endpoint returns the current file URLs, so code should read them rather than hard-code a version token:
GET https://alphafold.ebi.ac.uk/api/prediction/{UNIPROT_ACCESSION}
The response is a JSON list; each entry carries cifUrl, pdbUrl, bcifUrl, paeDocUrl (the PAE JSON), and paeImageUrl. Files follow the pattern AF-{accession}-F{fragment}-model_v{N}.cif where the version token is v6 as of 2025 and F1 is the only fragment for most proteins.
| pLDDT | Band | Meaning | AFDB color |
|---|---|---|---|
| 90-100 | Very high | Backbone and well-oriented side chains | Blue |
| 70-90 | Confident | Backbone generally correct | Cyan |
| 50-70 | Low | Backbone uncertain, cautionary zone | Yellow |
| < 50 | Very low | Placeholder ribbon, frequently an IDR | Orange |