Odd Dip in FSC Curve in 3D Flex

[Glenn-Masson]

I have a relatively small and flexible pair of domains.It is a homodimer, and despite trying to reduce the flexibility as much as possible, the structure still suffers from motion between the two domains. 3DVA and heterogeneous refinement have done quite a bit to improve this, but the density is still poor.

The 3D Flex Training, to my eyes, looks successful. I have three segments (66,65,64), and there is a very simple tree of 66>65, 66>64.

However, the 3D Flex Reconstruct looks odd:

Rigid looks ok (if somewhat poorer resolution than what I expected):

The Flex reconstruction however has an odd dip and “recovery” around 8A.

I have tried particle re-extraction, RBMC, and a few other things too, but I get the same dip. Any ideas? No symmetry has been imposed (despite it being a C2 symmetry). There are 172k particles.

For the training, I have 1 latent dim, 0.2 Rigidity, 0.075 Noise Injection SDEv, and 2 latent centring strength. I tried increasing the BFGS in the reconstruction up to 40 but got the same answer.

The flex map looks much better. I can see helical pitch and some bulky side chains, exactly where they are predicted to be.

Flex:

No-Flex:

Any ideas what the dip in the FSC is telling me?

Thanks!

[olibclarke]

Hi,

From the picture, helical pitch and sidechains are not obvious, and an improvement of ~6-3Å would be surprising. Do you have pictures showing more clearly the features you are talking about?

If I had to guess I would say this is overfitting of some kind. How much did you downsample for the training - what was the Nyquist?

Cheers

Oli

[Glenn-Masson]

Here is one helix; next to another, admittedly, it’s not beautiful.

The data is downsampled (800 px → 200 px); micrograph pix size 0.28A → 1.12A output.

Data Prep started at 200 px particles at 1.12 Å/px (a 224 Å box) and I didn’t crop before downsampling:

• training pixel size = 224 Å / 128 px = 1.75 Å/px

• training Nyquist = 2 × 1.75 = 3.5 Å