Overfitting (?) in local refinement

I am working on a ‘small’ (~85kDa) membrane protein with low snr particles (~270k). NU refinement doesn’t work well (~6-7A), however using local refinement excluding a flexible domain (~20kDa) appears to ‘work’ (excluded by mask, no particle subtraction). My issue is what I believe to be overfitting - while I see density improvements in the tm regions down to ~3.5A, the loop regions gets overfitted to a point when maps are not interpretable (refinement keeps goooooing until tm regions break down too, <3A).

I’ve varied mask dilation/padding (0-5 & 10-30), turned on gaussian priors (3/2). When I ‘restrict’ refinement sufficiently with these parameters (i.e. no overfitting) the refinement stops @ ~4-4.5A. My issue is that there is clearly high res information beyond this resolution in the tm regions…

Anyone had a similar issues before? Thoughts? Suggestions?


Are you enabling recentering of rotations & shifts during local refinement? If not, I would recommend trying that (as otherwise you are limited by the initial search range)

Thanks for your reply Oli!

I’m right now running a test with both on - I’ll post the results in a couple of hours

Hi Mika,

I am facing the same overfitting issue. Have you figured out this problem? How did you work around it?


Hi Chen

unfortunately not… We have collected more data and I’m still working on them, but the ‘issue’ is still there. I believe it goes back to it being a small particle, low SNR, and some flexibility in the (small) fiducial marker. I hope to find a solution and post it here… but if you get there first then please let me know!