Dear Eike,

The relion_refine program implements an expectation-maximisation 
algorithm, which is a local optimiser. It will therefore converge to the 
nearest local minimum. This means that if your starting model is 
incorrect, you may end up with an incorrect structure! (The same is true 
for most refinement programs out there.) The initial and/or final 
structure merely 'making sense' is probably not enough to guarantee a 
correct structure. Relion-2.1-beta now also implements a stochastic 
gradient descent algorithm, which was originally proposed for cryo-EM 
initial model generation by Marcus Brubaker in Toronto. Algorithms that 
include some stochasticity (like SGD) have the potential to reach the 
global minimum, but this is not guaranteed either. You could try to 
generate multiple initial models using SGD and see what comes out is 
consistent, although this also does not guarantee a correct solution.

When you have high-resolution data, the appearance of protein-like 
features like alpha-helices, beta-strands and ultimately amino acid side 
chains provide some internal validation for the correctness of the 
structure. However, this is not available at the resolutions you 
mention. My main suggestion would therefore be to collect a tilt pair 
data set on your sample, and then perform a tilt-pair validation (see 
papers by Richard Henderson et al). If your structure passes this test, 
then it is likely to be correct.

HTH,
Sjors



On 06/28/2017 04:39 AM, Schulz, Eike-Christian wrote:
> Dear all,
>
> I am working on noisy, low-resolution data with mainly amplitude contrast and I am concerned about model bias in my final solution.
>
> In my current workflow, I am generating 2D class averages and initial models in EMAN2.1 for subsequent 3D auto refinement in RELION2, omitting the 3D classification as I have not seen a sensible outcome at my resolution (< 20 Å).
>
> However, albeit the initial model appears to make physical sense in terms of its shape the solutions seem to be quiet biased by the input model. I am applying low-pass filters between 50 and 100 Å, with similar results.
>
> • How susceptible is relion 3D-refinement to model bias (no CTF correction)?
>
>   o Is there a rule of thumb, how much to filter an initial model (size of the particle, expected resolution etc.)?
>   
>   o Is there a way to display the lowpass filtered model relion generates (relion_image_handler ?)?
>
>   o Are there other strategies to avoid model bias besides filtering the input model ?
>
>
> With best regards,
>
> Eike
>
>
>
>
>

-- 
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres