Work in progress July 25 - September 19, 2024

Tom Goddard
September 19, 2024
Group meeting

Topics

• Demonstration of similar structure ligands, backbone traces and clustering
• Finding ligands bound to similar structures
• MMseqs2 sequence searches using RCSB web service
• Saving Foldseek and MMseqs2 search results
• AlphaFold 3 predictions setup with JSON
• AlphaFold 3 code release

Demonstration of similar structure ligands, backbone traces and clustering

• Look at hundreds of structures similar to one structure of interest to derive information.
• Example: Nipah virus attachment glycoprotein G, PDB 8XPS.
• 40% fatality rate in humnas, several outbreaks in Africa and Southeast Asia.
• open 8xps
• menu Tools / Structure Analysis / Similar Structures
• Search using Foldseek using Seoul National University server from Martin Steinegger lab.
• 194 similar structures found.
• Note about 2/3 have sequence identity less than 10%, undetectable by sequence searches.
• Press Traces button to show backbones of hits.
• Press Clusters button to cluster structures by C-alpha atom positions.

Finding ligands bound to similar structures

Nipah G protein tan Influenza neuramidase blue

Influenza ligands on Nipah G protein

Zanamivir anti-viral ligand from PDB 2CML chain A.

Flu 2CML white superimposed on Nipah tan, backbone matches.

• Nipah G protein binds to target cell surfaces.
• Influenza virus neuramidase has about 8-10% sequence similarity.
• Click on 1inh_A table row and press Open to superimpose flu structure.
• Similar fold but differs in many surface loops.
• Click species column header to order by species, and select all Influenza rows.
• Under Options click ligands for selected rows only.
• Press Ligands button to copy ligands from flu structures to Nipah G protein.
• Says it will fetch 67 structures from PDB.
• Takes about 20 seconds if structures fetched or 5 seconds if cached.
• Hide solvent.
• Hover over largest ligand (ZMR from 2CML chain A), FDA approved drug Relenza for flu treatment.
• Select 2cml_A line from table and Open to see backbone approximately matches.

MMseqs2 sequence searches using RCSB web service

• Similar sequences to Nipah 8xps can be searched using MMseqs2 (Many against many sequence searching) instead of Foldseek.
• This new capability uses the RCSB search server.
• Finds 158 similar chains in PDB, sequence identity as low as 20%.
• Backbone traces, clustering and ligands all work with MMseqs2 search results.
• BLAST search results from RBVI web service can also be used, but much slower than MMseqs2 for 200 million structure AlphaFold database.

Saving Foldseek and MMseqs2 search results

• Similar sequence search results appear in File History.
• This allows reopening results in later sessions instead of searching again.
• Results are saved in ~/Downloads/ChimeraX in Foldseek, MMseqs2 and BLAST subdirectories a ".sms" (similar structures) JSON file specific to ChimeraX.

AlphaFold 3 predictions setup with JSON

[
    {"name": "CNAG_01613-CNAG_06362",
     "sequences": [
         {"proteinChain": {"sequence": "MPLQSS...", "count": 1}},
         {"proteinChain": {"sequence": "MLSVPE...", "count": 1}}
     ],
     "modelSeeds": ["1"]},

     {"name": "CNAG_01613-CCH1_CNAG_01208",
      "sequences": [
	  {"proteinChain": {"sequence": "MPLQSS...", "count": 1}},
	  {"proteinChain": {"sequence": "MSHFRN...", "count": 1}}
      ],
     "modelSeeds": ["1"]},

     ...
]

• The alphafold interfaces command can now output JSON for running AlphaFold 3 predictions on Google's server.

• I used this to run all-by-all monomer and dimer predictions of the Cryptococcus neoformans sodium leak channel 5-mer complex for Hiten Madhani.
• With AlphaFold 2 those predictions took about 5 days.
• With AlphaFold 3 it took 21 minutes.
• A 4700 amino acid dimer that took 65 hours with AlphaFold 2 on Wynton, took 12 minutes with AlphaFold 3 on Google's server!

AlphaFold 3 code release

• Google said in May 2024 the would release AlphaFold 3 code within 6 months (mid-November 2024).
• Three Open Source clones of AlphaFold 3 are making progress The Race to Reproduce AlphaFold3.
• AlphaFold 3 will likely need an MMseqs2 server to produce the input multiple sequence alignment, one of the reasons we upgraded the Plato cluster memory.