← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-alignment-structural-alignment / references / usage-guide.md

Structural Alignment - Usage Guide

Overview

This skill covers backbone-aware structural alignment: pairwise (TM-align, US-align, Bio.PDB.Superimposer), database search at scale (Foldseek 3Di), structural multiple alignment (Foldmason, PROMALS3D, T-Coffee Expresso), and pLM embedding aligners (TM-Vec, vcMSA). Use it when sequence identity drops below ~25% (twilight zone), when remote-homology detection is required, or when fold-similarity quantification is the goal.

Prerequisites

# CLI tools
conda install -c bioconda foldseek tmalign usalign foldmason
# OR direct downloads
# Foldseek: https://github.com/steineggerlab/foldseek
# TM-align: https://zhanggroup.org/TM-align/
# US-align: https://zhanggroup.org/US-align/
# Foldmason: https://github.com/steineggerlab/foldmason

# Python
pip install biopython

# Optional: pLM aligners (heavier installs; GPU strongly recommended)
pip install fair-esm  # ESM-2 embeddings
# vcMSA: https://github.com/clairemcwhite/vcmsa
# DEDAL: https://github.com/google-research/google-research/tree/master/dedal
# TM-Vec: https://github.com/valentynbez/tmvec (active fork; original at tymor22/tm-vec is in limited maintenance)

For Foldseek searches against AlphaFoldDB, download the database:

foldseek databases Alphafold/UniProt /path/to/afdb tmp/
foldseek databases PDB100 /path/to/pdb tmp/

Quick Start

Tell your AI agent what you want to do: - "Compute the TM-score between these two PDB structures" - "Search this AlphaFold model against AlphaFoldDB to find structural homologs" - "Build a structural MSA from these 50 PDB structures with Foldmason" - "These proteins share <15% identity; align them via predicted structures instead of sequence"

Example Prompts

Pairwise Structural Alignment

"Align these two crystal structures with TM-align and report the TM-score and superposition"

"I have two homologs at 18% sequence identity. Compute structural alignment with US-align and tell me whether they share a fold."

"Compare apo and holo conformations of the same protein and quantify the conformational change."

Database Search at Scale

"Search my predicted structure against AlphaFoldDB to find evolutionary distant homologs"

"I have a metagenomic protein with no PDB hit. Predict its structure with ESMFold and search the ESM Atlas."

"Cluster all PDB structures sharing TM-score > 0.5 with my query."

"Search this antibody-antigen complex against AFDB-Multimer with Foldseek-Multimer and rank by complex TM-score"

Structural Multiple Alignment

"Build an MSA from these 30 PDB structures using Foldmason"

"Generate a structure-aware MSA combining sequence and PDB templates with PROMALS3D"

"I need a structural MSA for tree inference from a remote homologue family"

Hybrid Sequence-Structure

"Run T-Coffee Expresso to use PDB templates wherever they exist for my sequences"

"Improve my sequence MSA by adding structural information from AlphaFold predictions"

pLM Aligners

"These proteins are 12% identical; align them with a protein language model aligner"

"Use TM-Vec to predict TM-scores between sequences without computing structure"

What the Agent Will Do

  1. Assess whether structural alignment is the right tool (sequence identity, structure availability, downstream goal)
  2. Choose pairwise vs database-search vs MSA based on dataset
  3. Select tool: TM-align/US-align (pairwise), Foldseek (search), Foldmason (MSA), PROMALS3D (hybrid)
  4. Run with appropriate flags and output formats
  5. Interpret TM-score, RMSD, and LDDT in context (TM > 0.5 = same fold; TM > 0.8 = equivalent topology)
  6. Optionally derive a sequence MSA from the structural MSA for downstream phylogenetics

Tips