Should ChimeraX structure prediction focus on Boltz-1, Chai-1, AlphaFold 3, ...?

[Tom Goddard]

There is a lot of competition making AlphaFold 3 clones. Boltz-1 from MIT seems to be the only one that is fully open source including training data. Chai-1 has open source inference. AlphaFold 3 is limited to non-commercial use. The following review suggests Boltz-1 may get the most developer support because it is open.

The ABCs of Alphafold 3, Boltz and Chai-1
​https://blog.booleanbiotech.com/alphafold3-boltz-chai1

The Boltz-1 and Chai-1 papers are on biorxiv

​https://www.biorxiv.org/content/10.1101/2024.11.19.624167v2
​https://www.biorxiv.org/content/10.1101/2024.10.10.615955v2

Boltz-1 and Chai-1 have github repositories

​https://github.com/jwohlwend/boltz
​https://github.com/chaidiscovery/chai-lab

ChimeraX support might include creating input files for prediction (possibly yaml) and analyzing output confidences scores specific to each program. ChimeraX could attempt to work with multiple programs. Being able to directly run predictions from ChimeraX would be useful.

[Tom Goddard]

So far my effort has been on figuring out how to run AlphaFold 3 from source code. I've written a few web pages about how to run it on the UCSF Wynton cluster and how to dock ligands, e.g. 1500 FDA approved active ingredients to viral targets:

https://www.rbvi.ucsf.edu/chimerax/data/macromethods-jan2025/af3_drugs.html

https://www.rbvi.ucsf.edu/chimerax/data/af3-drugs-dec2024/af3_drugs.html

https://www.rbvi.ucsf.edu/chimerax/data/af3-wynton-dec2024/alphafold3_wynton.html

I think it might be wise to switch focus to Boltz-1 from MIT because of its fully open source license.

[Tom Goddard]

I made a Boltz-1 prediction tool in ChimeraX that installs Boltz on the local computer (Mac, Windows, or Linux) and allows running predictions.

https://www.rbvi.ucsf.edu/chimerax/data/boltz-apr2025/boltz_help.html

The main problem is that it can't handle very big molecular complexes (< 1000 residues) because it runs out of memory. Another important problem is that running on only an Intel CPU is quite slow, taking 96 minutes for a 911 amino acid prediction. The same prediction takes 3 minutes on a Linux Nvidia 4090 system or 5 minutes on a Mac Studio.

While AlphaFold 3 is more memory efficient it only runs on Linux with Nvidia graphics as far as I understand. Running locally on the user's PC is a major feature of Boltz so investigating how to lower memory use and increase speed with Boltz seems the best current direction.

I'm going to make separate tickets for addressing specific Boltz issues.