Optimizing Parameters for Integrating 3rd Party Docking Software #82
-
|
Hi All, I would like to integrate a 3rd party docking software, which takes approximately 4 seconds per molecule to dock. What are the best parameters, such as the number of molecules per epoch, batch size, learning rate, etc., that I can adopt for this scenario to optimize and expedite the process? Regards, |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
Hi, many thanks for your interest in REINVENT and welcome to the community! There is probably a bit of a discussion to be had what "optimize and expedite" means and if this should really be the goal. Fast learning may mean that you lock yourself into a chemical space too early. Having said that, in practice, you would want to minimize the docking calls as much as possible, of course. A typical approach would be to use staged learning. In the first stage(s) you would use rather cheap scoring components e.g. to fix the chemistry with say QED or other scores that are important to you. In a follow-up stage(s) you would then phase in the expensive omponent(s). The idea here obviously being that you would carry out docking only on more reaonable (as per your definition) compounds. Staged learning may also help in cases where scoring components "oppose" each other. The next level could be active learning but I believe we do not offer those workflows yet. Changing hyperparameters like batch size or learning rates may be useful too but would require a lot of experimentation to find the ones best suited to your problem. Also keep in mind that e.g. batch size has a direct influence on the optimizer. Many thanks, |
Beta Was this translation helpful? Give feedback.
-
|
Hi Hannes |
Beta Was this translation helpful? Give feedback.
Hi,
many thanks for your interest in REINVENT and welcome to the community!
There is probably a bit of a discussion to be had what "optimize and expedite" means and if this should really be the goal. Fast learning may mean that you lock yourself into a chemical space too early. Having said that, in practice, you would want to minimize the docking calls as much as possible, of course.
A typical approach would be to use staged learning. In the first stage(s) you would use rather cheap scoring components e.g. to fix the chemistry with say QED or other scores that are important to you. In a follow-up stage(s) you would then phase in the expensive omponent(s). The idea here obviously being tha…