alphafold-database_skill

Overview

This skill provides programmatic access to AlphaFold DB's 200M+ AI-predicted protein structures. Use it to retrieve models by UniProt accession, download coordinate files (mmCIF/PDB), and extract confidence metrics (pLDDT, PAE) for downstream analysis. It is designed for structural biology, drug discovery, proteome-wide studies, and pipeline integration.

How this skill works

The skill queries AlphaFold REST endpoints and the public Google Cloud dataset to locate predictions for a given UniProt ID or species. It can download model coordinate files, per-residue confidence JSON, and PAE matrices, or use Biopython helpers to parse mmCIF files and embed pLDDT into B-factors. For large-scale work it leverages gsutil and BigQuery metadata to filter and batch-download proteomes or high-confidence subsets.

When to use it

• Retrieve a predicted structure by UniProt accession for modeling or visualization
• Download mmCIF/PDB files and per-residue pLDDT scores for docking or pocket analysis
• Assess domain arrangement confidence using PAE before comparative modeling
• Bulk-download proteomes or filter entries via BigQuery for large-scale structural screens
• Integrate AlphaFold predictions into automated pipelines for protein engineering or variant modeling

Best practices

• Always check pLDDT per-residue scores and treat regions <50 with caution
• Use PAE heatmaps to evaluate relative domain positioning before interpreting inter-domain contacts
• Prefer the latest model version (v4) and note fragment numbers for large proteins
• Cache downloaded files and respect API rate limits when batching requests
• For proteome-scale tasks use gsutil and BigQuery to filter by organism and confidence metrics

Example use cases

• Download AF-P00520 model_v4.cif and extract pLDDT to identify high-confidence binding-site residues
• Query BigQuery for human proteins with fractionPlddtVeryHigh > 0.8 and batch-download high-quality proteome models
• Visualize PAE for a multi-domain protein to decide whether relative domain positions are reliable for docking
• Process a list of UniProt accessions with Biopython to compute average pLDDT and produce a quality summary table
• Retrieve model coordinates and convert pLDDT into B-factors for downstream MD preparation or docking tools