Project 11715 (GPU, Core21) to FAH
Posted: Mon Jul 02, 2018 6:55 pm
Project 11715 now on FAH!
Stats:
11715:
# atoms: 12775
credit: 9858
k-factor: 0.75
timeout: 7d
deadline: 10d
precision: mixed
This project is generating a test set for what we hope will increase the power of F@h by orders of magnitude, i.e. adaptive sampling. The idea is that instead of running many trajectories for a long time and seeing what happened at the end, we will run hundreds of trajectories for just 1 to a few gens - then pick frames from those that have advanced most on a given goal function (e.g. traveled furthest from their starting structure and closest to a target structure) and re-seed the next hundred of trajectories etc. - there’s a good amount of literature showing this can achieve very significant speedups. It will also let us make the same conclusions with much less data - speed the whole process up. 11715 is a short designed beta-hairpin peptide - trpzip2, which has unfolded, folded, and mis-registered states - i.e. it’s a good model of what we have with our cancer proteins where multiple conformational states have to be connected. We’re collecting ‘standard’ reference data here, and then will try our adaptive machinery for the first time to try to reproduce the ‘standard’ results. Thanks folks!
Stats:
11715:
# atoms: 12775
credit: 9858
k-factor: 0.75
timeout: 7d
deadline: 10d
precision: mixed
This project is generating a test set for what we hope will increase the power of F@h by orders of magnitude, i.e. adaptive sampling. The idea is that instead of running many trajectories for a long time and seeing what happened at the end, we will run hundreds of trajectories for just 1 to a few gens - then pick frames from those that have advanced most on a given goal function (e.g. traveled furthest from their starting structure and closest to a target structure) and re-seed the next hundred of trajectories etc. - there’s a good amount of literature showing this can achieve very significant speedups. It will also let us make the same conclusions with much less data - speed the whole process up. 11715 is a short designed beta-hairpin peptide - trpzip2, which has unfolded, folded, and mis-registered states - i.e. it’s a good model of what we have with our cancer proteins where multiple conformational states have to be connected. We’re collecting ‘standard’ reference data here, and then will try our adaptive machinery for the first time to try to reproduce the ‘standard’ results. Thanks folks!