Recent work Dec 10, 2025 - March 4, 2026

Tom Goddard
March 5, 2025
Group meeting

Topics

• Biophysics structure prediction class
• Boltz team goes private
• OpenFold in ChimeraX
• Bad ligand predictions
• Deep mutational scan visualization
• Samsung, Acer, Sony 3D display improvements

• January-March 2026 UCSF Biophysics course Methods in Macomolecular Structure covers structure prediction, cryoEM and NMR, 9 students
• For 2 weeks we tried Boltz structure prediction and BoltzGen design and students did projects in 3 teams of 3.
• Class plan and projects, day 1 using Boltz and RunPod, day 2 designing opioid peptides with BoltzGen, day 3 work on projects, day 4 project presentations.

Project 1: Do ABC transporter and drug predictions predict drug efflux?

Project 2: Do transcription factor predictions recognize specific DNA sequences?

Project 3: Design peptide binders to Alzheimer's disease amyloid beta.

• Used RunPod.io commercial cloud virtual machines, Linux with Nvidia GPUs.
• RunPod problems:
  1. First day of class no GPUs available at datacenter where I had installed Boltz.
  2. UCSF IT security software on laptops blocked ChimeraX Boltz server on RunPod machines.
  3. Work around: Used ssh tunnels and students installed Boltz from scratch.

• Mysteriously Boltz github went silent, no commits, no issues answered starting September 2025.
• Boltz team from MIT starts company with $28 million of VC funding.
• Boltz company - "intelligent agents to accelerate drug discovery - from hit identification to lead optimization".
• Demonstration at Andreesen Horowitz VC firm in January by Boltz developers Gabriele Corso and Jeremy Wohlwend showed web interface like ViewDock and web interface to setup drug binding predictions. Software as a service.
• Main partner of Boltz Company is Pfizer.
• Only changes in Boltz public github in last 6 months has been a bot closing 150 issues as "not planned".

Boltz company web site.

MIT Boltz team who founded company

Boltz Community edition?

• March 2026 businessman Andrei Volgin made a Boltz Community fork to try to continue Boltz development.
• He has merged 20 pull Boltz requests including some of mine.
• Does not look like this will go far with the 3 MIT Boltz developers engaged in private development.

• OpenFold 3 preview in ChimeraX - installs on Mac, Windows and Linux similar to Boltz.
• Uses a ChimeraX OpenFold 3 fork that allows it to run on Mac and Windows without Nvidia CUDA and simplifies installation of kalign dependency.

OpenFold predicting 52 anti-viral ligands run against HIV protease.

Prediction ipTM confidence for ligands.

ChimeraX OpenFold Prediction Explorer

• To promote OpenFold in ChimeraX I made a new tool Prediction Explorer to show MSAs and templates.
• In the future I will make this also work with Boltz and AlphaFold 3 and AlphaFold 2 predictions.

Prediction Explorer tool showing template PDB structures for OpenFold prediction for chain B. Chain A white and red colored by MSA sequence identity.

OpenFold multiple sequence alignment shown in as a profile grid. Red diamonds highlight predicted sequence, green outline selected residue.

OpenFold Pros and Cons

• Long term open community development.
• Ligand affinity contributed by a collaborating company AQAffinity reimplements Boltz affinity.
• Successful track record with OpenFold 2 that reimplemented AlphaFold 2.

• Gatekeeping may reduce community involvement. OpenFold 3 beta repository is private, as are discussion channels. I have a meeting scheduled today with the OpenFold business manager to gain access (took 2 weeks so far).
• Not as far along as Boltz
  OpenFold github 77 forks, 44 issues, 87 pull requests, 564 stars, created Oct 2025
  Boltz github 763 forks, 477 issues, 168 pull requests, 3800 stars, created Nov 2024
• Stated aim is not too exciting "... aiming to be a bitwise reproduction of DeepMind's AlphaFold3" while Boltz has pioneered many useful novel features (ligand affinity, protein design, steering potentials, distance restraints).

• Tristan Croll suggested we write an article on bad ligand chemistry in structure predictions.y
• Bad ligand chemistry in half of predictions by AlphaFold 3, OpenFold 3 and Boltz 2.
• Predicted geometry disregards bond orders so number hydrogens would not be right.
• Prediction often flips chiral centers
• These mistakes may substantially degrade the quality of binding predictions.

Wrong bond orders and hydrogens. Cyclohexane (left) and AlphaFold 3 prediction (right) hydrogens added by ChimeraX

Wrong chirality. AlphaFold 3 often flips chiral centers. Orange flipped, green not flipped.

• Error rates of 30-50% in AlphaFold 3, OpenFold 3 and Boltz predictions of 52 anti-viral ligands bound to HIV protease.

Proposed fix for predicting ligands

• To fix the ligand bond order problem, prediction programs could use explicit hydrogens on ligands.
• Current prediction programs do not train with hydrogens.
• I tested if OpenFold could predict hydrogens by changing the code. Worked but it was not trained with hydrogens so knows nothing about their placement or effect on geometry.

Cyclohexane predicted by OpenFold with explicit hydrogens. Ring flattened. Colored by pLDDT.

Cyclohexane from PubChem.

• Chirality errors might be reduced by a training loss function that has a chiral center signed tetrahedral volume term.

Isomorphic Labs Drug Design Engine (IsoDDE aka "AlphaFold 4")

• Nature commentary on IsoDDE from Isomorphic labs.
• Better drug affinity especially for ligands/pockets not seen in training.
• Better antibody binding prediction.
• Implementation is secret.
• White paper shows benchmark plots.

• Continue to improve deep mutational scan visualization (ChimeraX tickets 19165, 19403) with Willow Maestas-Coyote.
• Exporting computed residue mutation scores as CSV.
• Optimized 20 amino acid layout in heatmap residue labels.
• Added normalization of mutation scores by surface expression in coloring controls.
• Added rename and delete controls for in mutation coloring history.
• Added and simplified color palettes for coloring structures by mutation scores.

• Samsung 3D display OpenXR support added by Andre Michaelis, Max Planck Institute of Biochemistry
• Andre is also adding a mouse pointer with correct depth in a new XR3D ChimeraX plugin on Toolshed.
• I added a web page describing 3D glasses-free display support in ChimeraX.