Bok
12-16-2005, 09:44 AM
Posted by David Baker on the Rosetta forums..
We have results now for nine proteins. For two of these (2reb and 1dtj) all of the lowest energy structures are extremely close to the correct structure. For most of the remaining proteins, there are one or two quirks which prevent the low rmsd structures from achieving very low energies (like a puzzle where the last piece won't fit). Our goal, of course, is to be able to compute the structures of all of these proteins as accurately as we can for 2reb and 1dtj. For remaining proteins, it is clear that if we cheat and guide the runs close to the correct structure, they fall into the energy minimum surrounding the native structure, but we just haven't been getting close enough in the unbiased runs (this is the "wall" that people have commented on). From analysis of the results thus far, we have made a number of improvements to the method, that from our limited in house tests suggest should significantly improve the results on all the proteins.
You will be seeing two new things in the next several weeks. First, I will be testing the improvements mentioned above systematically on a set of eleven proteins to get a clear idea of how effective each improvement is individually, and to see what challenges still remain. Second, you will start seeing jobs submitted by some of the very talented other scientists in our research group. David Kim is just developing an internal queuing system that will make the BOINC resource available to other group members (up to now I've been rather unfairly monopolizing it!). Each scientist will be posting explanations of what they are doing and what they are learning. Also, David will get to start using BOINC for his own scientific research--he has had time for nothing but building the infrastructure to support the project for the past several months, but with more people in the group contributing and much in place already he will very deservedly get to use this terrific resource for his research on structure prediction.
and
I forgot to mention that we think we have found a possible contributor to the "stuck at 1%" problem. In the next few days we will be sending out a new version of the code with this problem fixed and some additional new features we would like to test.
Bok
We have results now for nine proteins. For two of these (2reb and 1dtj) all of the lowest energy structures are extremely close to the correct structure. For most of the remaining proteins, there are one or two quirks which prevent the low rmsd structures from achieving very low energies (like a puzzle where the last piece won't fit). Our goal, of course, is to be able to compute the structures of all of these proteins as accurately as we can for 2reb and 1dtj. For remaining proteins, it is clear that if we cheat and guide the runs close to the correct structure, they fall into the energy minimum surrounding the native structure, but we just haven't been getting close enough in the unbiased runs (this is the "wall" that people have commented on). From analysis of the results thus far, we have made a number of improvements to the method, that from our limited in house tests suggest should significantly improve the results on all the proteins.
You will be seeing two new things in the next several weeks. First, I will be testing the improvements mentioned above systematically on a set of eleven proteins to get a clear idea of how effective each improvement is individually, and to see what challenges still remain. Second, you will start seeing jobs submitted by some of the very talented other scientists in our research group. David Kim is just developing an internal queuing system that will make the BOINC resource available to other group members (up to now I've been rather unfairly monopolizing it!). Each scientist will be posting explanations of what they are doing and what they are learning. Also, David will get to start using BOINC for his own scientific research--he has had time for nothing but building the infrastructure to support the project for the past several months, but with more people in the group contributing and much in place already he will very deservedly get to use this terrific resource for his research on structure prediction.
and
I forgot to mention that we think we have found a possible contributor to the "stuck at 1%" problem. In the next few days we will be sending out a new version of the code with this problem fixed and some additional new features we would like to test.
Bok