I have done a little java programing and came across this programming strategy.

http://ai.bpa.arizona.edu/~mramsey/ga.html

http://www.aic.nrl.navy.mil/galist/src/

Maybe you can use it as one of the methods in Rosetta?

Thanx for a nice project.


Regards
Anders.

To the programmers:

One way of doing this:

If you have a protein chain. You can take the angle(amino-x to amino-x+1)and call this data one gene.

If you then take the best 1028 results, and save all the angles for each amino pair. This is the basic population.

a.
Next step is to pair them up, and mix the genes(angels for each amino-pair) makeing 16 new angel sets from each pair.(babies)

b.
Now comes testing witch ones are the best fit. This can be done in by "shaking" them in the rosetta programe and seeing how low an energy they fall to.

c.
When the results come back, keep the best 1024 results and let them be the next set of parents.

repeat a-c.

Just an example.

Regards
Anders

---

try it here:
http://www.aridolan.com/ga/gaa/gaa.html

me again :-)

---

A very good suggestion -- certainly worth looking into. Genetic algorithms have been some of the most successful in the past. They're especially good at optimization since each generation starts closer to the best solution.

It would clearly have to be run on a constantly evolving "population" since the program can't wait for all of one generation to finish before the next starts. However, I can't imagine that it would be insurmountable to re-weight the different starting points once a day or so.

It would require a bit more work on the server side, but that work could be as minimal as you like. For example, you could have the best results weighted to send more starting positions in those areas while the worst values would have their initial conditions sent less and less until they stopped altogether.

Clearly I don't understand the science behind proteins, and I'm not TOTALLY sure how much work is done in the "ab initio" phases. Running a genetic algorithm assumes that you ARE sending out different initial conditions in each WU. Keeping track of "good" results and sending more in those ranges could be very simple to code (again, it could be as simple or as complicated as you had time/computing power for) and it could greatly improve the efficiency of the overall search over time.

---

We recently had a very talented visiting graduate student in the lab who implemented a modified genetic algorithm in Rosetta, that could be easily adapted to boinc. I think someone is going to be testing it further in the lab before it's decided whether to use it in Rosetta@ home.

---

Thank you. I hope that you do.

It would be fun to join.

I will be keeping an eye on the progress on this thread.

---

Thank you. I hope that you do.

It would be fun to join.

I will be keeping an eye on the progress on this thread.

You won't have to join if they decide to test it with us. Just add an application to the project and create work units. Predictor@Home does this with at least two different kinds of work.

SETI@Home has been going in this direction for a long time with AstroPulse being the first of the next applications.

CPDN of course has its various models, with "slab" and "Sulfur" being the current crop (though they are not issuing new Slab models and it is not clear if they will restart issuing them or not). Down the pike is another new model ...

WCG, though not for windows, allows you to run one or two different projects with BOINC. The second is due "soon" for windows machines. Interestingly enough they are using Rosetta the application for the one project I am able to run for them ...

---