Thursday, 15 April 2010

PhD Progress

What does this mean? Well, to determine the minimum energy conformer - shape - of this molecule through exhaustive search would take 177,000 energy calculations. Each of which takes about an hour for one node on the National Computational Infrastructure network.

In fact, if we used all the nodes, all 140 TFlops, it would take about 4 months.A Teraflop is a trillion floating point operations per second.

We managed to find a way of doing it in less than 400 energy calculations (each taking quite a few Teraflops in its own right BTW). I'll have to check, but I think this is the molecule where we found a lower energy conformer than the one expected from X-ray crystallography of the solid state.

If so... we've found something no-one else in the world knows.

My work has been on formulating a systematic way of "tuning" genetic algorithms so their performance is good on a particular class of problems. "Tuning" has always been an art, rather than a science. My hypothesis is that for certain classes of classes of problems, this can be systematised.

The first class of problems I've worked on, along with a post-doctoral computational chemist, is determining the minimal-energy conformers (shapes) of molecules. This class of problems is firmly within the class of classes I'm interested in. We have good reason to believe that the harder the problem within a given class is, the better the relative performance becomes - but I wasn't expecting an improvement so great. I think it's an anomaly, but only further experimentation will determine that.

Now I have to consider whether to keep on doing more and more in this area, or to try to apply the same technique to other problem domains - my original intention. I'm no Chemist, but it looks like we may have attained the "Holy Grail of computational chemistry" here, as one real chemist put it.

Since the beginning of February, my co-author and I have produced three papers, one short paper and a poster for international conferences, and have given three presentations to the Research School of Chemistry and School of Computer Science at the Australian National University. This graphic is from one of them.

So if my blogging has been a bit lightweight recently - please forgive me. I've been a busy little Zoe. And my co-author, who actually made that graphic, has been just as frenetic in his activity. He's good. Really good. He should be - he's a former student of mine.

One thing he said to me when we first started collaborating though gives me an uneasy feeling:
I've been waiting for you, ZoBie-Wan. We meet again, at last. The circle is now complete. When I left you, I was but the learner; now *I* am the master.


Francis Turner said...

Now that is very very kewl.

I must be missing something but this looks like the sort of result that drug companies and the like would be willing to pay rather a lot of money for.

If I were you I'd be writing the patent NOW

Kiriel du Papillon said...

Wow... exciting stuff!

Sandra Louise said...

As a life long computer geek, I find this sort of thing fascinating!

I doubt that I could follow any of your algorithms but I am truly impressed that you have managed to improve computational analysis to this magnitude.

I was just happy to be able to devise Mandelbrot and Julia set calculations on my C64 that would complete in under an hour for minor magnifications.

If I could be so inquisitive, what language(s) are you using for your development? Are they mainstream style, like C++, or are they more math based like SPSS or are you developing native code?

And am I in the wrong ball park to think that this process could be applied to the protein folding problem?

Just geeking out a bit.

And to have found something that others have not is a wonderful achievement. You should be very proud!

How soon can we start referring to you as Dr. Brain?


Anonymous said...

Ahh, good old protein folding. I have built ye many a cluster.

One of the profs here is attempting to cut down the dimensionality of the solution space via singular value decomposition to make the search for local minima less of an ordeal.

Interesting stuff.

Anonymous said...

Links to your papers, Zoe? Or just a citation? Thanks.

Charlene :)