unexpectedly small, yet non-zero sigma2 value when using --always_cc

[Phaips]

Hi,

My Class3D job in RELION5 failed with the above mentioned error: I guess this is related to issue #582 which has never been solved it looks like(?) All my Class3D (as well as Refine3D) jobs run perfectly fine! I usually run them something like: which relion_refine_mpi --continue Class3D/job249/run_it015_optimiser.star --o Class3D/job266/run --dont_combine_weights_via_disc --scratch_dir $TMPDIR --pool 32 --pad 2 --iter 20 --tau2_fudge 1 --particle_diameter 150 --solvent_mask mask_sphere_bin1_box128_r40_s16.mrc --oversampling 1 --healpix_order 2 --offset_range 4.5 --offset_step 3 --allow_coarser_sampling --j 12 --gpu "0" --pipeline_control Class3D/job266/

However, I wanted to test --always_cc which could be beneficial for subtomo averaging (to not use the Bayesian approach). But with this flag now both the Class3D and Refine3D jobs fail with this same error. Hope this helps the troubleshooting! Thanks :)

Cheers, Philippe

[biochem-fan]

--always_cc which could be beneficial for subtomo averaging

Why do you think so?

[Phaips]

Well that's a good question :D it's mostly for testing purposes and also curiosity. From what I heard a pure CC-based refinement seems to work better at least in some cases for particles embedded in a crowded environment (as seen for example in STOPGAP). Along these lines, collaborators have reported the --always_cc flag to help with their 3D classifications of subtomograms in RELION and we are exploring this option now :)

[biochem-fan]

a pure CC-based refinement seems to work better at least in some cases

This is theoretically unjustified but may happen empirically; I don't know.

But you should remember that RELION is designed for the MAP approach. The CC-based approach is not an intended use case and there may be numerical instabilities as you observed.

What about using the MAP approach but with a limited marginalization by a small --maxsig?