Boltz ligand binding predictions and Animation examples

Tom Goddard
May 23 - August 14, 2025
Group meeting

Topics

• Boltz structure prediction evolution in ChimeraX
• Batch ligand predictions for drug discovery
• Boltz ligand prediction demo
• When are ligand predictions accurate?
• Prediction consistency with different random seeds
• Future ChimeraX Boltz improvements
• Animation examples
• Animation tips

Boltz structure prediction evolution in ChimeraX

• Boltz is a structure prediction program derived from AlphaFold 3 with similar accuracy to AF3 predicting assemblies of proteins, DNA, RNA, ligands, ions, solvent.
• ChimeraX installs and runs it on local Mac, Windows or Linux computer.
• ChimeraX Boltz tool web page, and User Guide documentation.
• July 11, 2025: Boltz 2.1 with ligand affinity used by ChimeraX in daily builds, how-to-use video posted on YouTube and BlueSky and X.
• June 26, 2025: Boltz 1.0 used by ChimeraX 1.10 release with 16-bit floating point precision I added for handling larger structures, and steering potentials to improve geometry.
• April 18, 2025: Boltz 0.4.1 added to ChimeraX. How-to-use video on Youtube and BlueSky and X.

Next ChimeraX Boltz improvement:
batch ligand predictions for drug discovery

• Demonstrate how to setup and run batch prediction.
• ChimeraX can setup and run batch ligand predictions against protein targets using Boltz 2.2.
• 839 FDA compounds run against HIV protease dimer took 3 hours on Nvidia 4090 (Linux), 24 hours on Mac M2 Ultra.
• Plan to release in ChimeraX daily build in coming week and post example video.

Boltz ligand prediction demo

• HIV protease dimer (198 amino acids) docking 51 anti-viral compounds.
• MacBook Pro M1 Max took 2 hours 11 minutes (GPU averaging 30 watts), Linux Nvidia 4090 took 14 minutes (10x faster, GPU averaging 330 watts).
• 12 known hiv protease inhibitors had affinity rankings 1,2,3,4,5,6,8,9,13,18,26 and ipTM rankings 1,2,3,4,5,6,7,8,10,11,15,16.
• HIV protease inhibitors with poorer affinity have high ipTM scores, so combinings scores is more accurate.

When are ligand predictions accurate?

• bioRxiv preprint Have protein-ligand co-folding methods moved beyond memorisation?
• Shows AlphaFold 3, Boltz 2, Chai-1 and Protenix all give ligand accuracy strictly proportional to how often similar ligands and pockets appeared in their training data.
• Experimental ligand binding training data is not extensive enough to generalize to novel ligands an pockets.
• Current methods do seem to summarize well ligand binding seen in training data which is still valuable.

Prediction consistency with different random seeds

The Mac and Linux runs of HIV protease with 51 ligands used different random seeds (based on time) since no seed was set. These plots compare the ipTM, log binding affinity, and binding probability for the two runs for the 51 ligands. All dots should be exactly on the diagonal if both runs gave identical results.

Future ChimeraX Boltz improvements

• User interfaces to specify post-translational modifications (e.g. phosphorylated residues), binding pockets, distance constraints, templates, ligands given as atomic structure instead of SMILES or CCD code.
• Improve memory efficiency of Boltz to allow running larger predictions (> 1000 amino acids).
• Allow local MSA calculation using mmseqs.
• ChimeraX Boltz improvements ticket

Animation Examples

Look at example animations we would like to be able to make in an Animation GUI.

These are better than average movies from 2024 Science and Nature papers found by Elaine to help Tom Goddard find an example to use for a cryoEM animation tutorial in October 2025 at Stanford.