Sampling a transition that happens once in a billion steps

Direct simulation will not reach it, and the usual enhanced-sampling machinery wants a reaction coordinate I do not have — the transition is collective and every one-dimensional coordinate I have tried gives a barrier in the wrong place.

I am aware of the methods that claim not to need one. My experience is that they need one implicitly, in the choice of the basis, and I would like to hear from somebody who has made one of them work on a system where the coordinate genuinely was not known in advance.