CASP7 Drawing Near...

[PY 222]

From Dr. Baker's journal:


I'm delighted to see that the rosetta@home throughput has been climbing recently! This is excellent timing as CASP7 is scheduled to start very soon. When it does start, there will be some changes in the screensaver as the true structure will not be known; we will have a question mark and the target number instead of the native structure.

As a warmup, we will be running targets from CASP6 on rosetta@home this week; some of these proteins are larger than what you are used to, but tests on ralph have not shown any problems. For the larger proteins we stop after the low resolution search because of the greater memory requirements of the high resolution search.

We had planned to release the new version of rosetta@home yesterday, but there have been some lingering issues with the watchdog and preemption that have taken a bit longer than we anticipated to resolve. The hope is to send out the new version tomorrow; if not we will wait until monday because many of you dislike weekend releases.

So, anybody interested in doing this for a while since the CASP trials are only held once every two years.

http://boinc.bakerlab.org/rosetta/forum_thread.php?id=1177#14723

[Digital Parasite]

Yes CASP shows the state of the art. In CASP 6 the Rosetta team didn't have their distributed computing setup and still performed much better than the competition out there. With the massive increase in CPUs and improved algorithm it should be interesting to see the results.

My clients have already downloaded some CASP6 targets. It would be great to push Rosetta extra hard during CASP 7 time so we can see just how much Rosetta has improved and compares to everyone else.

[PY 222]

Yes CASP shows the state of the art. In CASP 6 the Rosetta team didn't have their distributed computing setup and still performed much better than the competition out there. With the massive increase in CPUs and improved algorithm it should be interesting to see the results.

My clients have already downloaded some CASP6 targets. It would be great to push Rosetta extra hard during CASP 7 time so we can see just how much Rosetta has improved and compares to everyone else.

I remembered we were doing the CASP trials during the Distributed Folding days as well. Was it CASP5 or CASP6?

I too have gotten a few CASP6 workunits and I think this will generate quite a lot of interest amongst us. I know I am going in with everything starting next week to maybe get a good result.

[PY 222]

From Dr. David Baker's Journal


I just recieved the following email from the CASP organizers with the latest information on the CASP7 eperiment which begins on May 10:

Dear Predictors,

CASP7 will begin May 10 with a single target trial followed by two
targets on May 11.

Some additional notes on this yearsÕ process:

MODEL ACCURACY
We will pay special attention to model accuracy predictions. These can
be submitted for own predictions in the regular CASP format (PDB
B-factor field) on a per atom basis. These should be error estimates in
Angstroms. In addition an overall score for a given model can be
submitted as follows:
REMARK SCORE 0.7
in prediction header, where SCORE range is 0.0 to 1.0 (1.0 being a
perfect model). Independently, model accuracy predictions can be
submitted on server models, usually available within few days of target
release, or on your own models in the following format:
http://predictioncenter.org/casp7/doc/casp7-format.html#QA
The deadline for accuracy predictions on server models will be the
regular deadline for that target (typically 3 weeks).

In addition, to assess model evaluation methods on all CASP models
(i.e., including human expert predictions) we will be shorly collecting
software for a later assessment. Additional details will be announced
separately.

MODEL REFINEMENT
Special attention will also be paid to model refinement. In some cases
(CM/easy targets of less than 150 residues, preferably less than 100
residues), a single model submitted during the regular prediction window
will be selected for further refinement by others. Refinement window
will then open for additional 3 weeks. The usual refinement of own
models is still encouraged (using the unrefined and regular Ð for
refined models - model designations).

PREDICTION WINDOWS
Prediction windows will in general be shorter than in previous CASPs
(approximately 3 weeks). This is to adhere more closely to the
target structure release timelines adopted by crystallographers and to
minimize information leaks and subsequent target cancellations. However,
to allow assessment of methods requiring longer computation times, at
least some target deadlines will be extended. In such cases we will
still strongly encourage submitting models within the 3 week prediction
window. If information leak occurs after the initial three weeks but
before the assigned prediction deadline, evaluation of models will be
limited to those submitted within the 3 week window only.

PREDICTION OF FUNCTION
The format for function predictions is as follows:
http://predictioncenter.org/casp7/doc/casp7-format.html#FN
Additional targets will be made available for this category of
prediction (targets for which experimental structures may not be
forthcoming).

MODEL QUALITY FILTERS
Human expert predictions with severely unrealistic geometry will be
rejected outright. The criteria for this are as follows:
More than 5% of CAs taking part in clashes of less than 1.9 Angstrom.
OR
More than 25% of CAs taking part in clashes of less than 3.6 Angstrom.
CA-CA clashes below these percentage values as well as segmented
predictions with more than 4 chain breaks (CAs adjacent in sequence
separated by more than 5 Angstroms) will be flagged (warnings issued).
The model will be accepted, but it might be penalized in the assessment.
Missing loops or other deletions are acceptable.

Server predictions with clashes will be accepted in all cases, but
similarly to the human expert predictions will be issued warnings and
might be penalized in the assessment.

--
CASP7 organizers

Can't wait for this to start.

Also from his following post:



I was just looking at the results for the top teams:

Name Members Recent average credit Total Credit Country
1 XtremeSystems 246 311,043.65 25,052,975.81 International
2 Dutch Power Cows 1071 261,066.85 24,927,849.52 Netherlands
3 Free-DC 154 174,032.69 28,027,746.46 International

this is serious computing power; I think the recent credit numbers correspond to
3,100, 2,600 and 1,740 computers crunching more or less full time
for these top three teams which is fantastic!

It looks like the DPC have nearly caught up to XtremeSystems in total credit, but XtremeSystems is
moving ahead faster. Why does Free-DC have the most total credits but considerably less than the other two teams recently?

We will have to have prizes when CASP finishes at the end of July for the top overall team and the
top team during CASP. definitely a citation in the science paper on rosetta distributed contributing at the minimum!

So, I am calling... ALL HANDS ON DECK! BATTLESTATIONS PEOPLE.

This is no drill, I repeat, this is no drill.:rock: :rock:

[Digital Parasite]

Yes, prizes and mention in a journal woud be tres cool! I'm already maxed out as much as I can and will continue to go strong throughout CASP.

[LAURENU2]

So, I am calling... ALL HANDS ON DECK! BATTLESTATIONS PEOPLE
I will 2nd that motion
OK I will find room to buy a dozen more nodes and put them on line.

:Pokes: Come on Free-DC-ers Let show them what we are made of:thumbs:

[willebenn]

A question on the CASP test(s), will BakerLab flush all cached WU's on each users system when the CASP WU's are released ??? or should users flush???
Sure wouldn't want to miss jumping right on them when available.

[Thor]

Box up and running...

Lets smoke 'em!:smoking:


Florian

[PY 222]

A question on the CASP test(s), will BakerLab flush all cached WU's on each users system when the CASP WU's are released ??? or should users flush???
Sure wouldn't want to miss jumping right on them when available.

That is a good question.

I'll go check the forum and let you know.