Recent work Oct 24 - Dec 10, 2025

Tom Goddard
December 11, 2025
Group meeting

Topics

• BoltzGen protein design
• Low memory Boltz for larger structure predictions
• ChimeraX help chatbot NotebookLM
• Deep mutational scan coloring GUI
• SLAC cryoEM workshop animation tutorial
• Lab holiday card
• Web page explaining Claude AI use with ChimeraX
• Structure prediction for Biophysics Macromethods course
• ChimeraX privacy and security wiki page

• BoltzGen designs proteins that bind to specified proteins, peptides, nucleic acids or small molecules.
• It creates new protein sequences optimized for tight binding, predicts the bound structure.
• BoltzGen github
• Preprint BoltzGen: Toward Universal Binder Design (38 authors including Bill DeGrado at UCSF).

• I ran 1000 protein designs 80-140 residues (5 hours) for a binder to ganglioside activator protein PDB 1G13.
• I ran 100 protein designs 120-140 residues (24 minutes) for a binder to disordered amyloid beta, Alzheimer's plaque protein, 42 residue peptide.
• Authors recommend running 10,000-60,000 designs before experimental verification.

10 BoltzGen designed proteins (blue to red) binding GM2 activator (tan)	5 BoltzGen designed proteins binding amyloid beta (red)
Scores design_to_target_iptm 0.63 0.57 0.51 0.46 0.41 ... min_design_to_target_pae 2.9 3.1 4.4 5.3 5.2 ... interaction_pae 10.3 10.2 14.0 15.2 16.3 ... design_ptm 0.81 0.78 0.86 0.82 0.79 ... num_prot_tokens 256 244 291 296 296 ... ...	Scores design_to_target_iptm 0.62 0.64 0.60 0.60 0.54 min_design_to_target_pae 1.2 2.2 2.1 2.0 2.6 interaction_pae 10.5 8.8 10.5 11.8 11.6 design_ptm 0.77 0.77 0.74 0.72 0.72 num_prot_tokens 168 176 176 162 180 ...

• Tom Litfin forked and modified Boltz (github) to optimize memory allowing much larger structures to be predicted.
• Preprint LMI4Boltz: Optimising VRAM utilisation to predict large macromolecular complexes with consumer grade hardware.
• Increases the maximum number of residues on Nvidia 4090 (24 GB) from 1300 to 2500.
• Reduces computation time on Windows with Nvidia 3070 by a factor of 2 on large structures (> 900 residues).
• On Mac and systems without Nvidia graphics, no improvement, since main optimization is to move seldom-used tensors from GPU memory to CPU memory. Mac and non-Nvidia are already using CPU memory.
• Not yet added to ChimeraX.

Boltz prediction times in minutes. Tokens is number of standard polymer residues plus ligand atoms.

• Linux Nvidia 4090 - VRAM 24 GB, cpu i9-13900K (24 cores) 64 GB (DDR5 5200MHz), minsky.cgl.ucsf.edu
• Windows Nvidia 3070 - Windows 11, VRAM 8 GB, PCIe 4.0, cpu i7-12700K (12 cores) 64 GB DDR5 4000MHz, vizvault.cgl.ucsf.edu

ChimeraX Chat Help web browser interface.

• Matthias Vorlander made a ChimeraX Help chatbot using Google NotebookLM in August 2025 which gave good quality answers in our testing, but used limited ChimeraX input sources.
• Tom Terwilliger of the Phenix team made a Phenix Help chatbot with NotebookLM, and Nigel Moriarty said they were happy with its answers.

• I made a ChimeraX Help NotebookLM using our extensive documentation including:
  1. ChimeraX User's Guide (286K words, 266 html files)
  2. ChimeraX tutorials and presentations (192K words, 140 html files)
  3. ChimeraX mailing list archives 2020-2025 (1105K words, 9000 messages)
  4. ChimeraX Programmer's Guide (195K words, 78 html files)
  5. ChimeraX Recipes example code (22K words, 59 markdown files)
• Free NotebookLM allows only 50 input sources (files), each at most 500K words and paid versions allow only 300 input sources.
• I merged the above source into 8 input files in Markdown format which NotebookLM is supposed to handle best. Mailing list was divided into 3 files for 2020-2025 to meet size limits.
• I converted relative links in User's Guide to absolute links to the web site.

• I put a Chat Help link in the ChimeraX documentation index.
• Using NotebookLM chat requires signing in to Google.
• NotebookLM can make podcasts (audio), reports on specific topics, "mind maps", videos and flashcards covering specific topics.
• I had it make a summary podcast without specifying a topic (12 minutes) which was frigheningly slick, and included some wrong information.

• Willow Coyote-Maestas in UCSF Bioengineering and Therapeutic Sciences suggested adding structure coloring controls in ChimeraX using deep mutational scanning data.
• Residue colors use a red-blue based on the mean or sum of assay scores for each mutation of that residue.
• I added the coloring controls to mutation score scatter plots.
• Also added list of previously defined colorings. Often Willow's lab has dozens of colorings for different assays (e.g. opioid receptor response to different drugs) and different subsets of mutations.

• Different colorsing will be listed in a table with more detailed description of the coloring parameters.
• Ability to save the residues scores to .csv files.
• Improved ordering of 20 amino acids in the mutation heatmap residue labels.
• Willow is working on a paper describing the ChimeraX mutation score capabilities.
• Willow and I will teach 2 days of deep mutational scan analysis with ChimeraX in the Biophysics Macromethods course.

• I attended a 3 day CryoEM Modeling and Structure Validation Workshop, hosted by Wah Chiu at the Stanford Linear Accelerator, October 28-30, 2025.
• Spent several hours talking with Tristan Croll (ISOLDE) and Alexis Rohou (Claude AI with ChimeraX), also with Greg Pintilie, Mike Schmid, Jane Richardson, Wah Chiu, and many others.
• I presented a new ChimeraX animation tutorial using recent voltage-gated potassium channel structures. using commands and scenes.
• Audience was lively with dozens of questions. 75 minutes. 50 in-person participants and 140 online via Zoom.
• Recording of tutorial.
• Video of animation created at end of tutorial.
• Tristan Croll gave an insightful ISOLDE talk and demonstration right before my tutorial.

• Had Claude AI follow Elaine's recipe to create RBVI holiday card flotillin image.
• Layed out and printed cards.

Front

Back

• Created a web page explaining how to setup Claude AI to control ChimeraX.

Claude started ChimeraX and issued about 20 ChimeraX commands and described the steps it took in about 5 minutes with no intervention.

Claude Desktop App on Mac

ChimeraX 1.11 release candidate

• 4 days of ML structure prediction training for Macromolecular Methods course January 8-20, 2026.
• Will have 3 team project (each team 3 students) covering
  1. Ligand binding predictions with Boltz 2.
  2. Protein-protein interaction discovery with Boltz 2.
  3. Protein engineering with BoltzGen.
• Other parts of course are cryoEM taught by Dan Southworth, NMR taught by John Gross, and Deep mutational scans taught by Willow Coyote-Maestas.
• Organized by John Gross.

• Created web page describing ChimeraX data privacy and security issues.
• This is primarily for pharmaceutical companies and other license holders to better understand when data they access in ChimeraX may be sent to off-site servers.
• It is on the ChimeraX Wiki so anyone in our lab can add to it or edit it.