Rosetta and the journey to predict proteins’ structures, 20 years on
Interesting the conclusion:

The return of Rosetta to the top of the competition (CASP) will require extensive improvements for contact map prediction that are likely to rely on deep learning technology. If Rosetta choses that route, the formidable processing power that Rosetta@Home provides could be the key of its success: instead of generating tremendous number of decoys, it could provide the means of training the most complex machine learning based predictors.

---

And yesterday....
"De novo protein design by deep network hallucination" (preprint)

---

If Rosetta choses that route, the formidable processing power that Rosetta@Home provides could be the key of its success: instead of generating tremendous number of decoys, it could provide the means of training the most complex machine learning based predictors.

I have thought for some time that distributed computing is ideally suited for training the AI machines, since it can be done off-line from the actual predictions, and latency is not an issue. The first project I know of to use it is Cosmology, and then GPUGrid had a CPU "quantum chemistry" project for a while. I think QuChemPedIA uses it in some form also.

And it would seem that Rosetta could start with a ground-up design, maximizing the use of extensions and parallel processing. It would be a whole new ball game for us.

---

And it would seem that Rosetta could start with a ground-up design, maximizing the use of extensions and parallel processing. It would be a whole new ball game for us.

Rosetta can do both: ML and decoys!

P.S There is a new ML player: MLC@Home!!

---

And it would seem that Rosetta could start with a ground-up design, maximizing the use of extensions and parallel processing. It would be a whole new ball game for us.

Rosetta can do both: ML and decoys!

P.S There is a new ML player: MLC@Home!!

I just got one it thinks it will take around 7+ hours and has a 2 day deadline.

And it would seem that Rosetta could start with a ground-up design, maximizing the use of extensions and parallel processing. It would be a whole new ball game for us.

Rosetta can do both: ML and decoys!

P.S There is a new ML player: MLC@Home!!

I just got one it thinks it will take around 7+ hours and has a 2 day deadline.

After running a dozen or so the time is now down to just under 1 hour still with the 2 day deadline and I get 260 credits per task.