There's an article on the BBC website about Alphafold's performance on CASP14. It doesn't mention rosetta - is rosetta still used for the final settling of the model?

https://www.bbc.co.uk/news/science-environment-55133972

---

Here is a slightly different take on it, though you may need a subscription.
https://www.telegraph.co.uk/news/2020/11/30/google-ai-researchers-crack-50-year-old-protein-folding-problem/

But I can make a prediction too: It will change how we do business here, though they may not tell us much for a while.

---

David Baker, the director of the Institute for Protein Design at the University of Washington, who has been using similar computer technology to design anti-coronavirus drugs, said DeepMind’s methods could accelerate that work.

“We were able to design coronavirus-neutralizing proteins in several months,” he said. “But our goal is to do this kind of thing in a couple of weeks.”

https://www.nytimes.com/2020/11/30/technology/deepmind-ai-protein-folding.html

---

And Nature:

https://www.nature.com/articles/d41586-020-03348-4

---

From AlphaFold blog:
AlphaFold2 and CASP14

---

Just read all the articles, things are really starting to move!

Be interesting to see what comes out of the CASP14 Conference.

In the past BakerLab and AlphaFold had some contacts...
AlphaFold visit IPD
Working together will be great!

---

“This really is a big deal” — David Baker on today's #AlphaFold news

---

From Foldit blog:

It is not yet clear how DeepMind's methods will perform with protein design, but the Rosetta community has already incorporated AlphaFold's methods from CASP 13 and we plan to integrate these into Foldit soon.

Lastly, we want to congratulate all of you as DeepMind's co-founder and CEO revealed that Foldit was an inspiration for AlphaFold! Today they gave the keynote talk at the CASP meeting, with an entire slide about Foldit.

Keep up the great folding and keep inspiring everyone!

---

It is not yet clear how DeepMind's methods will perform with protein design, but the Rosetta community has already incorporated AlphaFold's methods from CASP 13 and we plan to integrate these into Foldit soon.

But where does Rosetta@home fit into the picture? I was under the impression that we helped them out with modeling the folding process to improve its accuracy. That would seem to put us in competition (or in the line of fire) with AlphaFold.

I am perfectly comfortable with being redundant here if the science progresses to that point. That would be a great success, but it would be helpful if someone at Rosetta told us a little about it.

---

But where does Rosetta@home fit into the picture?

Even if Rosetta’s current modelling approach ends up being obsoleted by AlphaFold’s, that won’t remove the need for computational resources to refine the models further and run other experiments using them. So I don’t think we’re redundant quite yet…

I am perfectly comfortable with being redundant here if the science progresses to that point. That would be a great success, but it would be helpful if someone at Rosetta told us a little about it.

Rosetta makes eterogeneous simulations, not only folding...
And refinement of simulations is also important!

---

I don’t think we’re redundant quite yet…

That said: it’s eminently foreseeable that Rosetta@home gets complemented by AlphaFold@home if DeepMind can’t (or don’t want to) fold their new methods into the core Rosetta software.

I believe Alphafold uses Rosetta for the final refinement - is that confirmed somewhere?

And I guess the options are (based on my limited understanding!):

1. Alphafold source will be released, in which case either the training/classifier work (on GPU) could be done in house or through BOINC. This might only need to be done once, in which case in house/cloud would make sense, but if it's to be repeated regularly then maybe BOINC world be ideal.

2. trrosetta will be built on using
the Aplhafold details released during CASP14 and in the paper that's yet to be released, and then it's the same as above.

Then I guess there's:
* a huge number of proteins to work through whose 3D structure is yet to be determined, especially for larger proteins. No doubt some will prove problematic from Alphafold too.

* Protein complexes.

* Protein design.

So I think the future will require far more processing power rather than less, but maybe there will be a lull before then whilst there's a code and coffee frenzy. Or maybe Rosetta is still really useful in the interim as it's the best thing most researchers have access to?

Also, I expect there will be some inertia to overcome due to existing grants/funding, and new funding might be needed before much work can be done in a new direction.

---

Also, this might be interesting. I haven't listened to it yet though so can't confirm... Too late here so it's one for tomorrow.

https://a16z.com/2020/12/06/16mins-deepmind-alphafold-casp-protein-folding-deep-learning-ai-bio/

---

Very interesting. They interviewed Vijay Pande (creator of Folding@home), and so quite the expert. If I understood it correctly, he estimated they used 100 to 200 GPUs to train Alphafold. That could easily be done in-house.
So if they need us, it will probably be for something else.

---

1. Alphafold source will be released, in which case either the training/classifier work (on GPU) could be done in house or through BOINC.

1 - Not so sure they will release their code. At the end of CASP13 they released only results and parts of documentation/code.
2 - During CASP the simulations are about "little" proteins. The problems are with big proteins and interactions between different proteins.
So, use of GPU will be more important with these problems.

2. trrosetta will be built on using
the Aplhafold details released during CASP14 and in the paper that's yet to be released, and then it's the same as above.

As above, we are not sure of what Alphafold team will release...

So I think the future will require far more processing power rather than less,

That's for sure!!

Also, I expect there will be some inertia to overcome due to existing grants/funding, and new funding might be needed before much work can be done in a new direction.

Rosetta@Home has historical problem in his code (in a recent publication they said that HALF of the Rosetta code is redundant!!!). They have to clean up heavely and start with these new features.

---

Just to be clear, my understanding is that GPUs are much quicker for the AI training, but once that is done, the modelling is better suited to CPUs, presumably due to the need for memory per process/model and the difficulty of multithreading that process.

---

Just to be clear, my understanding is that GPUs are much quicker for the AI training, but once that is done, the modelling is better suited to CPUs, presumably due to the need for memory per process/model and the difficulty of multithreading that process.

I am sure you know much more about that than I do. However, anything they send out to crunchers must withstand the high latency (usually days), rather than a few microseconds with an in-house machine.
I think the AI training would allow for that latency, except that they probably don't need us there. So can we effectively do the modelling, even if they are willing to send it out?

---

Directly from the December Update of the Microbiome Immunity Project (which uses Rosetta) on World Community Grid:

New technology announced

A few weeks ago, the researchers participated in a biennial protein structure predication conference (CASP14). This year, they heard about a brand new technology that may have a major impact on the protein structure prediction problem and the study of protein shapes in general. It is very exciting to see such progress in the field which is going to benefit all of us in the long run.

Over the coming months, the researchers will learn more about this technology, if and how to use it, and how computationally feasible it is at the scale they are working on. They'll let us know if/when the Microbiome Immunity Project would embrace it.

obviously they are referring to DeepFold ;) perhaps it could be possible for us to crunch such workunits sometime in the future

Source: https://www.worldcommunitygrid.org/about_us/viewNewsArticle.do?articleId=672[/url]