How Can I reduce micelle-dominated alignment when applying HR-HAIR to a small GPCR?

[Beom95]

Hello everyone,

I previously posted about my dataset here:

Good 2D classes but no secondary structure in Ab-initio reconstruction of a small GPCR-BRIL construct

Hi everyone, I’m processing a relatively small GPCR in LMNG micells. The ICL3 region has been replaced with BRIL, so there is some additional protein density protruding from the micelle. The 2D classification results look reasonably good, and I can see particle features that appear to correspond to the protein rather than just the detergent micelle. However, once I proceed to Ab-initio, the secondary-structure feature of the TM domain are no longer apparent. In the Ab-initio and subsequent He…

Thank you very much for all the helpful suggestions on that thread.

Following the recommendations, I tried applying the HR-HAIR method to my dataset. However, even after HR-HAIR, the resulting reconstructions still appear to be dominated by the micelle density, while the GPCR/TM features remain relatively weak.

The image shown above are the results of Ab-Initio Reconstruction, using the following parameters:

• Number of Ab-initio classes: 3
• Maximum res(Å): 2.3
• Minimum res(Å): 5
• Fourier radius step: 0.005
• Center structures in real spaces: OFF
• Initial minibatch size: 300
• Final minibatch size: 1000

I followed the HR-HAIR strategy described by @olibclarke

https://www.biorxiv.org/content/10.1101/2025.09.08.674935v1

My understanding is that HR-HAIR is designed to reduce dependence on low-resolution features during initial alignment and instead make better use of higher-resolution structural information.

Therefore, I expected that it might help prevent the strong detergent micelle signal from dominating the alignment.

However, in my case, the micelle still seems to have a strong influence on the reconstruction.

I would therefore appreciate any advice regarding the following questions:

1. Has anyone successfully applied HR-HAIR to a small GPCR or a similar detergent-solubilized membrane protein where the micelle signal strongly dominates the particle?

2. Is there any recommended way to reduce the contribution of the micelle during alignment? For example, would modifying the frequency range used for alignment, using a high-pass strategy, changing the initial/maximum resolution in HR-HAIR, or using a protein-focused mask be helpful?

3. If HR-HAIR still produces a micelle-dominated reconstruction, what would you recommend as the next processing step? Would it be better to continue with Local Refinement using a GPCR/TM-focused mask, perform additional particle cleaning, or return to an earlier stage and try a different alignment strategy?

The reason I am still hopeful about this dataset is that the 2D classifications look very clear, with recognizable protein features, and I also have a reasonably large number of particles. However, these features seem to disappear or become much weaker during 3D reconstruction/refinement.

Because of this, I suspect that I may be missing an important processing step or parameter rather than the dataset itself being completely unsuitable for high-resolution reconstruction.

I would greatly appreciate any suggestions, especially from anyone who has processed small GPCRs or other membrane proteins with strong micelle density.

Thank you very much for your help.

[PeterK]

I have processed a dataset of a GPCR in LMNG detergents, and I saw the same issue with the HR-HAIR approach. With the final resolution set too high, the map always ended up overfitting towards the detergent micelle.

Unfortunately I was not able to solve this issue through processing, so my solution was to add a Fab combined with an anti-fab nanobody that produced a rigid complex which was fairly simple to process.

Is it possible to add an anti-BRIL Fab to your sample? I believe this was the approach used in the inactive cone pigment structures by Kato https://www.science.org/doi/10.1126/science.adz3996

[Ishan]

I have successfully processed a dataset of GPCR in detergent using HR-HAIR protocol. I was able to get good alignment off the ab initio but not the ab initio refine (new job in CS >5.0). The micelle density was visible in my processing as well, but looked a bit better than what you observe. In my experience the density looks at bit low res generally after HR-HAIR. But a local refinement of good set of particles from ab initio yields a good res map for the receptor.

My approach was to get very high res particles with visible secondary structures form 2D classification, and use only those particles for HR-HAIR with 2 classes. I think low res particles would prevent good alignment and result in a poorer quality map, which subsequently wont be useful for local refinement.

Just based on your ab initio density, I am unable to see BRIL, which suggest to me that the particles may not have enough signal in the alignment marker, which might also prevent good alignment in HR-HAIR ab-initio.

[Beom95]

Thank you very much for sharing your experience. This is very helpful, especially since you observed a very similar issue with a GPCR in LMNG.

I also noticed that when I set the final resolution relatively high in HR-HAIR, the reconstruction tended to converge more strongly toward the detergent micelle rather than the TM region. Based on your experience, I will also test whether using a lower final resolution can reduce this behavior.

Thank you also for sharing the cone pigment paper. I’ll read it carfully and consider this approach for our next sample preparation.

Thank you again for your very helpful advice!

[Beom95]

Thank you very much for sharing your experience.

Your point about particle selection is particularly interesting. In my current workflow, I included particles from several good 2D classes, but I didn’t restrict the dataset only to classes in which secondary-structure features were clearly visible. Based on your experience, I think it would be worthwhile to perform a much more stringent 2D selection and repeat HR-HAIR using only the highest-quality particles.

I also appreciate your comment regarding the relatively low-resolution appearance of the HR-HAIR ab-initio map. I had been concerned that the poor TM density itself indicated that the alignment had failed.
It is encouraging to hear that, in your case, a good subset of particles from the ab-initio reconstruction could still produce a high-resolution receptor map after local refinement.

I will first try a more stringent 2D classification and repeat HR-HAIR with only particles showing clear secondary-structure features, using two classes as you suggested. I’ll also try local refinement on the best particle subset from the current ab-initio result to see whether the TM density improves.

If you do not mind, could you also share the approximate HR-HAIR parameters you used, particularly the initial/final resolution settings and the number of particles you used for the ab initio step? It would be very helpful for comparison with my dataset.

Thank you again for the very detailed advice!

[Beom95]

I have a few additional questions regarding the workflow after HR-HAIR.

After obtaining a good particle subset from the HR-HAIR ab initio reconstruction, did you first run Homogeneous Reconstruction Only and then use the resulting particles/volume for Local Refinement?

For the mask used in Local Refine., did you generate it from the volume obtained from Homogeneous Reconstruction Only, excluding the detergent micelle density and including only the TM region of the receptor?

I am also wondering whether the parameters I am planning to use for each job are appropriate. I have listed them in detail below, and I would greatly appreciate if you could take a look and let me know whether these setting are similar to what you need.

Homogeneous Reconstruction Only

• Force re-do half-set split: ON

Local Refinement

• Rotation search extent (deg): 3
• Shift search extent (Å): 1
• Re-center rotations each iteration: ON
• Re-center shifts each iteration: ON
• Initial lowpass resolution (Å): 6
• GSFSC split resolution (Å): 10

If possible, I would also be very interested to know whether you used similar settings, or whether you found any of these parameters particularly important for obtaining a good receptor map.

Thank you again for your help. Your advice has been extremely useful for troubleshooting my dataset.

[Ishan]

These are the parameters I have used for HR-HAIR. These were taken from @olibclarke paper.

You can do homogenous reconstruction only and then go to local refinement or go strait to it. I have tested a bunch of masks before finalizing one which gave good resolution. I did have have to play a lot with the settings of local refinement. But in my opinion tighter restrains on the movement of the particles work good.