Hi,
I notice that in the new ab initio refinement job, the random initializations are different for each half map. Is there a reason to do this, as opposed to using the same initialization for both, given that both are initialized with distinct particle sets? I’m wondering if starting from different random initializations may lead to half-maps that diverge too much.
Actually that brings up another question - in the log, it is stated that ab initio reconstruction is initialized with “random initial densities”. Are they actually random synthetic densities, or generated from the input particles (I think the latter)?
Cheers
Oli
2 Likes
Hi @olibclarke ,
Thanks for the question 
We kept the initializations separate in order to keep the cross-talk across the half-maps as minimal as possible, and it ended up staying that way after testing, even though it is not necessary (e.g. refinement already uses a common initializer). In a subsequent version, we will be using the same initializations for each half map.
In practice we have observed the issue you’re describing about the half-maps occasionally diverging from each other. We think that in addition to using different initializations, this could be happening because stochastic gradient descent gives ab-initio more liberty to change the density from iteration to iteration than than homo refinement can. (SGD “bounces around” more than the EM algorithm does). The problem of keeping the half-maps “in sync” and preventing divergence may have to be addressed both in the initialization (keeping them the same) and most likely by some extra regularization step during the iterations.
Re: last question – the densities are generated by reconstructing a tiny batch of the particles with random poses applied (so not fully synthetic densities – they come from the data).
Making the HR-HAIR workflow fully gold-standard is definitely of interest to us & the community – glad as always to be able to build on the insights your group recently found!
Michael
2 Likes
