Recent work Sep 18 - Oct 23, 2025

Tom Goddard
October 23, 2025
Group meeting

Correlation of ipTM interface quality scores for 2 Boltz runs with reduced or full length MSAs.

Topics

• Structure prediction
• Deep mutational scans
• Scenes and Animations
• Claude AI controlling ChimeraX

• Tried running boltz 4600 mycoplasma dimers on Mac. mmseqs MSA calc crashes on Mac.
• Fixed boltz MSA processing code and made 2 pull requests (#581, #583). Updated ChimeraX Boltz.
• Reran 4600 dimers on linux with fixed boltz msa processing, updated web page of results.
• Boltz dimer predictions of a single structure give somewhat different ipTM scores from run to run. Should predict 5 structures and take best score.

• Apple released SimpleFold protein monomer structure prediction September 24, 2025.
• arXiv preprint says it avoids explicit pair-wise residue representations, and time-consuming triangular attention layers to make it faster and less memory intensive.
• Uses Mac neural engine.
• I tested SimpleFold on my Mac using largest 3B (3 billion) parameter model on 2200 amino acid 9hma, 15 minutes.
• Not currently useful, low accuracy, only slightly faster than Boltz, many installation/run glitches, proteins only.

human separase PDB 9hma (cryoEM) 2200 amino acids

SimpleFold prediction 3B model wrong fold unphysical clashing helices

AlphaFold 3 prediction mostly correct

• Planning for 4 days of ML structure prediction training for Macromolecular Methods course January 8-20, 2026.
• Organized by John Gross.

• Updated Alphafold 3 on minsky.
• Studied how to use mmseqs MSA calculation instead of hmmer with AF3.

• Updated ChimeraX to use Alphafold database version 6, released September 2025.
• Version 6 has isoform predictions, we don't currently support fetching those.

• Looking at all single mutations of human opioid receptor, from Willow Coyote-Maestas and Matt Howard.
• For each mutation measure receptor activity when binding 11 drugs and surface expression.
• For overview of data I made a deep mutational trace plot using ChimeraX mutation scores capabilities.
• Surface expression normalization in plot filters out effect of reduced membrane receptor numbers.

• Improved mutation score association with structures.
• Use Needleman-Wunsch run locally instead of web service.
• Use sequence derived from mutation data instead of requiring user to specify a reference sequence.

• Added scene support for volume models including per-vertex coloring and masking.
• Made 2d label and arrow GUI.

• Alexis Rohou, director of cryoEM at Genentech, made videos showing Claude AI controlling ChimeraX using Model Context Protocol.
• Zach added a ChimeraX mcp_server bundle to daily build to allow others to use this.
• I tested Claude AI with ChimeraX on opioid receptor + fentanyl and potassium channel example systems.
• YouTube video describing opioid receptor Claude successes and failures. Video showing Claude log.

ChimeraX controlled by Claude coloring by bfactor.

Claude Desktop requesting it fix the bfactor coloring to not use the AlphaFold palette.

Claude successes and foibles

• Claude succeeded in many somewhat challenging ChimeraX tasks and struggled with others.
• How can we give Claude better command output so it succeeds more often?
• Survey the mistakes Claude makes.

Surprising successes

1. Claude recognized and explained that two copies of the opioid receptor in 8ef5 have different hydrogen bonds due to very small differences in poses.
2. Claude suggested and tried out alternate rotamers for a mutation to avoid clashes.
3. Claude tried the minimize command and realized my ChimeraX version was too hold and did not have it. Then when I installed the new version it reopened, deleted extra chains, mutated and minimized the structure with only being asked to get me back to where we were.
4. Energy minimized, then checked to make sure mutated residue clash was resolved.
5. Claude suggested after minimizing comparing the pre and post-minimized structures. It frequently makes sensible suggestion about the next step in the analysis.
6. Ran Boltz predictions through ChimeraX, understood that it would run in background.

Cases where Claude needed hints

About half the time Claude does not realize it made a mistake, and half the time it realizes it can't do something and asks the user to help.

1. Tried to hide, close, cartoon with atomspec #1.1-1.5 for submodels 1-5. This gives no error because it is a legal spec for 3-level models. Right syntax #1.1-5.
2. Could not read Boltz affinity score from Log because it apparently has no way to read the Log other than return values of commands. Claude asked me to tell it what the Log said.
3. Could not find the proper residue range when asked to select potassium channel voltage sensor helix S4. ChimeraX Uniprot annotations shows it, and Uniprot web page shows it.
4. Did not realize its Boltz predictions of opioid receptor and fentanyl did not include fentanyl. I told it the prediction had no fentanyl but eventually had to tell it to use the boltz command "ligands" option.
5. Color by bfactor added "palette alphafold" for PDB 9my3 which is not a predicted structure. Structure was solid blue. I had to check the command and tell it not to use alphafold coloring.
6. Claude could not figure out that 9my3 and 9my4 have different chain orders. I told it chains needed to be renamed to make a morph work but it insisted they were ordered fine even though morph sends molecules flying. Claude mistakenly used the "rename" command to rename models after I told it what chains needed swapping. Then I told it to use the changechains command.