Protein Sequence Similarity Search

Protein Sequence Similarity Search

Overview

Two complementary search engines, each wrapped as a submit-poll-download-parse pipeline:

• MMseqs2 (ColabFold) — fast, default. Searches UniRef (+ optional MGnify environmental). Best for quickly building an MSA-grade hit set.
• EBI BLAST (NCBI BLAST hosted at EBI) — slower but more exhaustive search across multiple UniProt/UniRef/PDB databases. Use when you need BLAST-specific scoring (e.g., comparing against a published BLAST result) or when MMseqs2 returns few hits.

Both tools save the full hit list to a JSON file on disk and return only a top-10 markdown preview in the response, keeping the agent's context window small.

Project Tools (VenusFactory2)

When to Use This Skill

• The user gives a sequence and asks "find similar proteins" / "what's this protein's family" / "find homologs"
• You need to build an MSA from a single seed sequence: run MMseqs2 → write hits → download_clustalo_msa_by_fasta
• You need orthologs in a specific clade: BLAST with database=uniprotkb_human (or _bacteria, _viruses, etc.)
• You need PDB hits to seed structure-based analysis: BLAST with database=pdb

Which Engine to Pick

Supported BLAST Databases

uniprotkb uniprotkb_swissprot uniprotkb_swissprotsv uniprotkb_reference_proteomes uniprotkb_trembl uniprotkb_refprotswissprot uniprotkb_archaea uniprotkb_arthropoda uniprotkb_bacteria uniprotkb_complete_microbial_proteomes uniprotkb_eukaryota uniprotkb_fungi uniprotkb_human uniprotkb_mammals uniprotkb_nematoda uniprotkb_rodents uniprotkb_vertebrates uniprotkb_viridiplantae uniprotkb_viruses uniprotkb_enzyme uniprotkb_covid19 uniref100 uniref90 uniref50 pdb

Output JSON Schema (file_info.file_path)

jsonCopy code
{
  "hits": [
    {"target_id": "...", "q_cov": 78.3, "e_value": 1.2e-50, "identity": 0.45, "aln_len": 240, ...},
    ...
  ],
  "metadata": {"engine": "MMseqs2 (ColabFold)", "ticket_id": "...", "query_length": 256, "hit_count": 178, ...}
}

Rate Limiting

• ColabFold MMseqs2: 2 req/s, ~1-3 min wall clock per job. Polls every 10 s.
• EBI BLAST: 2 req/s submit; polls every 30 s. Jobs take 2-10 min.
• Both timeout at 15 min by default — increase timeout_secs for very large queries.

Common Mistakes

• Passing a FASTA file with multiple records but expecting all to be searched: only the first record is used. Loop in the agent if you need multiple queries.
• Skipping out_dir: required. Tools error with ValidationError: empty out_dir.
• Setting database to an unsupported value: returns ValidationError with the full allowed list — pick from that list.
• Treating MMseqs2 hits as the final answer when count is 0: this is a soft failure; fall back to BLAST.

References

• ColabFold MMseqs2 API
• EBI BLAST web service and terms of use
• Adapted from google-deepmind/science-skills:skills/protein_sequence_similarity_search/