adaptyvbio/foldseek
> (1) Finding similar structures in PDB/AFDB databases, (2) Structural homology search, (3) Database queries by 3D structure, (4) Finding remote homologs not detected by sequence, (5) Clustering structures by similarity. For sequence similarity, use uniprot BLAST. For structure prediction, use chai or boltz.
npx skills add https://github.com/adaptyvbio/protein-design-skills --skill foldseek| Requirement | Minimum | Recommended |
|-------------|---------|-------------|
| Python | 3.8+ | 3.10 |
| RAM | 8GB | 16GB |
| Disk | 10GB | 50GB (for local databases) |
Note: Foldseek can run locally or via web server. No GPU required.
# Upload structure to web server
curl -X POST "https://search.foldseek.com/api/ticket" \
-F "[email protected]" \
-F "database[]=afdb50" \
-F "database[]=pdb100"
# Install Foldseek
conda install -c conda-forge -c bioconda foldseek
# Search PDB
foldseek easy-search query.pdb /path/to/pdb100 results.m8 tmp/
# Search AlphaFold DB
foldseek easy-search query.pdb /path/to/afdb50 results.m8 tmp/
import subprocess
import pandas as pd
def foldseek_search(query_pdb, database, output="results.m8"):
"""Run Foldseek search."""
subprocess.run([
"foldseek", "easy-search",
query_pdb, database, output, "tmp/",
"--format-output", "query,target,pident,alnlen,evalue,bits"
])
return pd.read_csv(output, sep="\t",
names=["query", "target", "pident", "alnlen", "evalue", "bits"])
| Parameter | Default | Description |
|-----------|---------|-------------|
| --min-seq-id | 0.0 | Minimum sequence identity |
| -e | 0.001 | E-value threshold |
| --alignment-type | 2 | 0=3Di, 1=TM, 2=3Di+AA |
| --max-seqs | 1000 | Max hits to pass through prefilter; reducing this affects sensitivity |
| Database | Description | Size |
|----------|-------------|------|
| pdb100 | PDB chains | ~340K structures |
| afdb50 | AlphaFold DB clustered at 50% sequence identity | ~53M structures |
| swissprot | SwissProt structures | ~540K structures |
| cath50 | CATH domains | ~50K domains |
# results.m8 (tabular)
query target pident alnlen evalue bits
query 1abc_A 85.2 120 1e-45 180.5
query 2def_B 72.1 115 1e-32 145.2
$ foldseek easy-search query.pdb pdb100 results.m8 tmp/
# results.m8 columns: query target pident alnlen mismatch gapopen qstart qend tstart tend evalue bits
query 1abc_A 85.2 120 ... 1e-45 180.5
query 2def_B 72.1 115 ... 1e-32 145.2
Hit identities and E-values above are placeholders; foldseek does not print the
[INFO] lines shown by some other tools.
Should I use Foldseek?
│
├─ What are you searching?
│ ├─ By 3D structure → Foldseek ✓
│ ├─ By sequence → Use BLAST (uniprot skill)
│ └─ Both → Run both, compare results
│
└─ What do you need?
├─ Find structural homologs → Foldseek ✓
├─ Remote homolog detection → Foldseek ✓
├─ Structural clustering → Foldseek ✓
└─ Functional annotation → Cross-reference with UniProt
# Compare your design to PDB
foldseek easy-search design.pdb pdb100 similar_natural.m8 tmp/
# Ensure design is novel (low similarity to known)
foldseek easy-search design.pdb afdb50 novelty.m8 tmp/
# Novel if: top hit identity < 30%
# Find scaffolds for motif grafting
foldseek easy-search motif.pdb pdb100 scaffolds.m8 tmp/ \
--min-seq-id 0.0 -e 10
wc -l results.m8 # Number of hits
No hits: Lower e-value threshold, try larger database
Too many hits: Increase min-seq-id threshold
Slow search: Use smaller database
| Error | Cause | Fix |
|-------|-------|-----|
| Database not found | Wrong path | Check database location |
| Invalid PDB | Malformed structure | Validate PDB format |
| Out of memory | Large database | Use more RAM or web server |
Next: Download hits with pdb skill → use for scaffold design.
Take adaptyvbio/foldseek from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.