relion 4 helical reconstruction worse than 3.0.8

[zhenwuliaoa]

I am sorry to let you know that for some reason I found relion 4's helical reconstruction is not as efficient as 3.0.8. And I don't know what exactly happened between these two versions.

I am handling a helical structure that is very big. It is a 24-mer per rise and the box is set as big as 400 to include everything. Since my old computer (relion 3.0.8, GPU: RTX2080, CUDA: 11) could not do such computation and I tried bin 2, it gave me around 6A resolution after 3D class (not even been refined) of around 500k particles. However, when I redo the same caculation with the same parameters in a new PC without rescale (relion 4 which is impossible to get relion 3 installed, GPU: RTX4090, CUDA: 12), the 3D class gave me worse volume than bin 2.

Now I am using single particles to do this job. I do not know why newer version could not perform better than 3.0.8.

Environment:

• OS: Ubuntu 22.04 LTS
• RELION version RELION-4.0.2-commit-f2e59d
• Memory: 525 GB
• GPU: GTX 4090

Dataset:

• Box size: 400 px
• Pixel size: 1.1 Å/px
• Number of particles: 500,000
• Description: A filamentous protein of 100kD (monomer), 24-mer per turn, 36 nm-diameter, spring-like structure.

[biochem-fan]

Since my old computer (relion 3.0.8, GPU: RTX2080, CUDA: 11) could not do such computation

Without padding, 8 GB of VRAM should have been enough for a 400 px box.

it gave me around 6A resolution after 3D class (not even been refined)

Resolution from Class3D is not based on gold standard FSC and should not be trusted.

the 3D class gave me worse volume than bin 2

To claim RELION 4 is worse than 3, you have to compare the results with the same input.
What happens if you use the binned particles in RELION 4?

[scheres]

Also, please note all runs have a stochastic component, so a claim one version is worse than the other should be backed up with multiple replicates.