[QUESTION]: Recommended paramaters for molecule screen about molpal HOT 1 CLOSED

Hi @finlayiainmaclean,

Thanks for the question. k, window, and delta are all parameters that control the dynamic stopping criterion. Namely, if the difference between the average of the current top-k molecules does not exceed the rolling average of the top-k molecules from the window most recent iterations by at least some fraction delta, then MolPAL should stop exploration. Alternatively, if you would like to guarantee stopping by the time you've explored 2M molecules, then you can also set a maximum exploration budget, either as a fraction between 0-1 (2/13 in this case) or as an absolute number of molecules (2M, also in this case). For most of my use-cases, I usually set window to be sufficiently large so that I guarantee that I explore for the set amount of max-iterations or fully exhaust my budget. I.e., if window=10 but max-iter=5 then you can't determine a rolling average with which to trigger early stopping.

For your use case I might do the following:

k=100000
budget=0.02
window=10

You'll also need to decide how you would like to acquire molecules, i.e., init-size and --batch-size. We generally did a 1+5 split, where we acquired 1/6 of our total budget in our initialization batch and in each exploration batch. Practically this will look like --init-size=0.0033 --batch=size=0.0033 to get you to that. You could also spend less of your budget on initialization and put that towards exploration (e.g., --init-size=0.0025 --batch-size=0.0025.) In our cases this usually led to a higher enrichment factor ("sample efficiency",) but at the cost of more time spent on model (re)training and inference. I'd also generally recommend you use the --model mpn --conf-method mve --metric ucb as these were a generally robust hyperparameter set (c.f., figures 2/4 from the paper).

The last thing I'll say is that while we got great results on our test datasets from the paper, those are no guarantee of future quality. Every optimization is different, so you might find that your sample efficiencies are lower (or possibly higher) than those we showed in the paper. Hopefully, this was helpful, but let me know if you have any more questions!

from molpal.