AlphaFold3 and ChimeraX

[okovalevskiy@…]

Dear ChimeraX team,

I hope you are keeping well. I would like to take this opportunity to thank
you for all the great work on supporting AlphaFold2 in the ChimeraX and
implementing excellent visualisation of the confidence metrics like your
integrated PAE viewer.

As the key community members, I wanted to share an exciting announcement
<https://dpmd.ai/alphafold3>. Today, Google DeepMind announced AlphaFold 3
via a paper <https://www.nature.com/articles/s41586-024-07487-w> in Nature.
AlphaFold 3 is the next generation of AlphaFold, developed by Google
Deepmind and Isomorphic Labs, which can predict the structures of nearly
every biomolecule type — with state-of-the-art accuracy.

We have two main pieces of news to share:

1) Where previously AlphaFold predicted the 3D structures of proteins,
AlphaFold
3 can model other biomolecules, including DNA, RNA, ligands, and ions,
model large, complex structures made of these biomolecules, and predict
biologically meaningful interactions between them. Which we hope will
unlock significant further research directions across the life sciences. If
you would like to read more about the model’s new capabilities please see
the paper <https://www.nature.com/articles/s41586-024-07487-w> in Nature.

2) Alongside this, we are launching AlphaFold Server
<http://alphafoldserver.com>. Powered by AlphaFold 3, the Server is an easy
to use, web-based tool for non-commercial researchers to access AlphaFold
3’s predictions for non-commercial research. AlphaFold Server makes the
majority of AlphaFold 3 capabilities available- with the exception of some
biomolecule capabilities, particularly those relevant to drug discovery.
It will be the most accurate tool for predicting how proteins interact with
the other molecular components of the cell. The server is free of charge
for users, which we hope will enable even more researchers to make use of
structures as they build their understanding of the processes underpinning
the living world.

Please note that the AlphaFold Server also produces JSON files carrying PAE
(they are available as part of zip file for download). However, our new
implementation may not be compatible with ChimeraX visualisation of the PAE
plot. The reason is that PAE plot for AlphaFold2 was based on amino acid
residues, and for AlphaFold3 PAE is based on tokens (one token per standard
amino acid residue or nucleotide, one token per atom of any chemically
modified residue, nucleotide, ligand or ion). I am absolutely happy to
discuss this difference in more detail; we can set up a video chat, if you
are interested.

We are excited to share this news with you, and welcome any feedback you
might have on this announcement or AlphaFold Server more broadly. Please
also feel free to share this with your colleagues and wider network.

All the best,

Oleg


-- 
[image: logo]
Oleg Kovalevskiy
Applied Scientist
okovalevskiy@deepmind.com
If you received this email in error, please do not forward it to anyone
else (it may contain confidential or privileged information), erase all
copies and attachments, and let me know that it has gone to the wrong
person.

[Tom Goddard]

Thanks Oleg. I was excited to read about the AlphaFold latest developments announced in your October 31, 2023 pdf, and will read the AlphaFold 3 paper and try AlphaFold Server. I will try to make ChimeraX handle AlphaFold 3 PAE files. We are making a ChimeraX 1.8 release in about a month have already started release candidates but maybe I can get it in that version.

ChimeraX can currently run AlphaFold 2 predictions on Google Colab. While web servers also do that it can be a little easier for researchers to setup jobs and analyze results using ChimeraX. I'd be interested in allowing ChimeraX to submit and get results for AlphaFold 3 jobs.

Does the AlphaFold 3 Server have a REST API?

Will it be possible in the future to run AlphaFold 3 jobs on our local computers at UCSF? I hope Alphabet's desire to monetize drug discovery does not throw a wrench in open science.

[okovalevskiy@…]

Hi Tom,

Thank you for your swift reply!

On the compatibility with AF3 JSONs - it is so nice to hear that you are
keen to implement this for the next version, we really appreciate this! And
I am sure that many ChimeraX users will enjoy it as well :)

I would suggest simply running a modelling job containing a PTM, nucleic
acid and a ligand and check the JSONs in the downloaded zip file.
The formats are documented in the FAQ ("How do I interpret all the outputs
in the downloaded JSON files?" section) and if you have any questions or
something does not work - please let me know, I am very happy to help.

On running AF3 jobs from ChimeraX - we aren’t able to prioritise
implementing API access at the moment, but understand it will be an
enabling feature and are certainly keeping it in mind as we plan future
enhancements and improvements to our beta product. We will definitely
contact you if/when the API will become available, but I can't promise
anything (e.g. time-wise) at this moment.

However, we have a feature that is somewhat close to the API. We have
implemented uploading of the JSON files that describe AFServer jobs; I can
imagine that such JSON files could be generated via software like ChimeraX,
and then modelling results could be imported back after download.
Please see FAQ section "Can I import job files into AlphaFold Server?" and
this GitHub link
<https://github.com/google-deepmind/alphafold/tree/main/server> for the
documentation. This is less convenient than API, but that's what we have at
this moment.

On running AF3 locally - unfortunately, at the moment we don't have plans
to support this. Again, we will definitely let you know if this changes in
the future.

Thank you for the continuous support of AlphaFold in ChimeraX!

All the best,
Oleg


On Wed, May 8, 2024 at 10:34 PM ChimeraX <ChimeraX-bugs-admin@cgl.ucsf.edu>
wrote:

>
>
>
>
>

[Tom Goddard]

Thanks. I saw Pushmeet Kohli's comment that the plan is to release AF3 for academic use on their computers in 6 months

​https://x.com/pushmeet/status/1790086453520691657

Looking forward to that.

I have fixed ChimeraX display of AlphaFold 3 predicted aligned error (including non-polymer atoms) as described in more detail in the following ticket

​https://www.rbvi.ucsf.edu/trac/ChimeraX/ticket/15144