Skip to content

Foldseek Structural Search

Overview

Foldseek is a structural similarity search tool that identifies proteins with similar 3D structures. The pipeline integrates Foldseek to search for structural homologs of both Boltzgen-designed and Protenix-refolded structures against large databases like AlphaFold or Swiss-Model.

What is Foldseek?

Foldseek uses a novel 3Di structural alphabet combined with traditional amino acid sequences to enable ultra-fast structural similarity searches. It's significantly faster than traditional structural alignment tools like TM-align while maintaining high sensitivity.

When to Use Foldseek

Enable Foldseek structural search when you want to:

  • Identify homologs: Find proteins with similar structures in large databases
  • Validate designs: Check if designed structures resemble known proteins
  • Discover function: Infer potential functions based on structural similarity
  • Assess novelty: Determine if designs are truly novel or similar to existing structures

Enabling Foldseek

nextflow run seqeralabs/nf-proteindesign \
    -profile docker \
    --input samplesheet.csv \
    --run_foldseek \
    --foldseek_database /path/to/database_dir \
    --foldseek_database_name afdb \
    --outdir results

Key Parameters

Required Parameters

Parameter Description
--run_foldseek Enable Foldseek structural search (default: false)
--foldseek_database Path to Foldseek database directory (required when Foldseek is enabled)
--foldseek_database_name Name of the database within the directory (default: afdb)

Search Parameters

Parameter Default Description
--foldseek_evalue 0.001 E-value threshold for reporting matches (lower = more stringent)
--foldseek_max_seqs 100 Maximum number of target sequences to report
--foldseek_sensitivity 9.5 Search sensitivity (1.0-9.5, higher = more sensitive but slower)
--foldseek_coverage 0.0 Minimum fraction of aligned residues (0.0-1.0)
--foldseek_alignment_type 2 0=3Di only, 1=TMalign (global), 2=3Di+AA (local, default)

Database Setup

Download and prepare the AlphaFold database:

# Download AlphaFold database (choose your version)
wget https://foldseek.steineggerlab.workers.dev/afdb-swissprot.tar.gz
tar xzf afdb-swissprot.tar.gz

# Use in pipeline - database files will be in the extracted directory
--foldseek_database /path/to/afdb-swissprot \
--foldseek_database_name afdb

Swiss-Model Database

Alternatively, use Swiss-Model structures:

# Download Swiss-Model database
wget https://foldseek.steineggerlab.workers.dev/swissprot.tar.gz
tar xzf swissprot.tar.gz

# Use in pipeline
--foldseek_database /path/to/swissprot \
--foldseek_database_name swissprot

Custom Database

Create a custom database from your PDB files:

# Create database from directory of PDB/CIF files
foldseek createdb /path/to/structures/ mydb

# Use in pipeline - specify your custom database name
--foldseek_database /path/to/ \
--foldseek_database_name mydb

What Structures Are Searched?

The pipeline runs Foldseek on:

  1. Boltzgen budget designs - All structures from intermediate_designs_inverse_folded/
  2. Protenix refolded structures - All structures predicted by Protenix (if enabled)

Each structure is searched independently, allowing comparison of: - Original Boltzgen designs - ProteinMPNN-optimized sequences refolded by Protenix

Output Files

For each design, Foldseek generates:

results/
└── sample_id/
    └── foldseek/
        ├── design_id_boltzgen/
        │   ├── aln.m8              # Alignment results in BLAST-like format
        │   ├── summary.tsv         # Summary of top hits
        │   └── alignment.html      # Detailed alignment visualization
        └── design_id_protenix/     # (if Protenix enabled)
            ├── aln.m8
            ├── summary.tsv
            └── alignment.html

Output Format

The summary.tsv file contains:

Column Description
query Query structure name
target Target structure identifier
evalue E-value (lower = more significant)
prob Probability score
score Alignment score
qlen Query length
tlen Target length
alnlen Alignment length
qstart, qend Query alignment boundaries
tstart, tend Target alignment boundaries
description Target protein description

Interpreting Results

E-value Interpretation

  • E < 1e-10: Very strong structural similarity
  • E < 1e-5: Strong structural similarity
  • E < 0.001: Moderate similarity (default threshold)
  • E < 0.01: Weak similarity
  • E > 0.1: Likely not significant

Example Analysis

# View top hits for a design
head results/sample1/foldseek/design1_boltzgen/summary.tsv

# Count significant hits (E < 1e-5)
awk '$3 < 1e-5' results/sample1/foldseek/design1_boltzgen/summary.tsv | wc -l

# Extract top hit details
head -n 2 results/sample1/foldseek/design1_boltzgen/summary.tsv

Integration with Other Analyses

Foldseek results are automatically integrated into the consolidated metrics report when both are enabled:

nextflow run seqeralabs/nf-proteindesign \
    --input samplesheet.csv \
    --run_foldseek \
    --foldseek_database /path/to/afdb \
    --run_consolidation \
    --outdir results

The consolidated report includes: - Best E-value for each design - Top matching protein name/description - Number of significant hits - Comparison across Boltzgen and Protenix structures

Performance Notes

  • GPU accelerated: Foldseek can utilize GPUs for faster searches
  • Memory usage: ~4-8 GB per search depending on database size
  • Search time: ~1-5 minutes per structure with AlphaFold database
  • Database size: AlphaFold database is ~200 GB

Troubleshooting

Database Not Found

ERROR: Foldseek database not found at /path/to/database

Solution: Ensure the database path is correct and accessible: 

ls -l /path/to/database/
# Should show database files like database.index, database.lookup, etc.

Out of Memory

ERROR: Foldseek ran out of memory

Solution: Reduce the number of results or increase memory allocation: 

--foldseek_max_seqs 50  # Reduce from default 100

No Significant Hits

If no hits are found: - Check E-value threshold (try relaxing to --foldseek_evalue 0.01) - Increase sensitivity (--foldseek_sensitivity 9.5) - Verify database is appropriate for your designs - Consider if designs are truly novel (no existing homologs)

References

See Also