Update AlphaFold prediction to use version 2.3.0 which handles larger structures

[Tom Goddard]

The AlphaFold 2.3.0 release made December 13, 2022 has memory optimizations that allow predicting larger multimers. Also the network parameters have been update and are supposed to give more accurate predictions for large proteins. And the sequence databases have been updated.

We should update to use this if possible. Currently we use ColabFold, so it will require ColabFold to support the new AlphaFold 2.3.0. I made a ticket on the ColabFold github to ask for that update

​https://github.com/sokrypton/ColabFold/issues/345

[Tom Goddard]

It would be worth making docker and singularity images for AlphaFold 2.3.0 for minsky and wynton and test how much larger a prediction can be made and update the AlphaFold performance web page

https://www.rbvi.ucsf.edu/chimerax/data/alphafold-jan2022/afspeed.html

and also the instructions for running AlphaFold on Wynton.

https://www.rbvi.ucsf.edu/chimerax/data/singularity-apr2022/alphafold_wynton.html

To use the new AlphaFold on Wynton I would need the Wynton administrators to update the databases. I've made a request to Wynton support for that.

[Tom Goddard]

There is a ColabFold beta that uses AlphaFold 2.3.1 now. Sergey Ovchinnikov made that update and made a pull request to the Steinegger lab AlphaFold fork to update it that has not yet been merged. Once 2.3.1 becomes the production ColabFold I'll try to update ChimeraX to use it.

​https://github.com/sokrypton/ColabFold/issues/345

[Tom Goddard]

ColabFold appears to now use AlphaFold 2.3. Should update ChimeraX to use a current AlphaFold. Will require significant testing.

[Tom Goddard]

Done.

I updated about 3 weeks ago to use the newer ColabFold and no changes were needed. I was forced to do this update because Google Colab updated to Python 3.10 (unannounced) and broke ColabFold because it relied on an older tensorflow package that was not available in 3.10. Unfortunately the Python 3.10 update also broke OpenMM installation so energy minimization with AlphaFold is broken. Although ColabFold fixed this by updating to OpenMM 7.7.0 (from 7.5.1) it is still broken in ChimeraX because the new OpenMM changes the preferred text encoding to ascii from utf-8 breaking all Google Colab shell magic which ColabFold does not use but ChimeraX does. I had a work-around to fix the text encoding change caused by OpenMM that no longer works after the update from Python 3.9 / OpenMM 7.5.1 to Python 3.10 / OpenMM 7.7.0. There is a separate ticket to fix this problem #8975.