Thanks! To be honest, we weren't sure if this would work when we first started with PLK1. We just had the luck of having 24/7 access to the instruments. I've since moved on to a new role, but I'm so glad the rest of the team finished it off so well. Never expected it to go this far.
Thanks.
For your question, as far as I know, they didn’t use it.
Thanks, Oli. What a relief :)
The small protein cryo-EM paper I helped start has been accepted in Nature Communications - what a relief. Thanks to @landerlab.bsky.social and @olibclarke.bsky.social for their constructive review. Congratulations to all co-authors, and wishing everyone good luck.
www.nature.com/articles/s41...
After my first post, @sjorsscheres.bsky.social asked what would happen if the AI were asked to optimise for the expected 2.9 Å result. I didn’t try that here, but it raises the question of how one might properly test AI-driven optimisation in cryo-EM processing. Curious how others would approach it.
4. From that point on the particle counts were already different from the tutorial, so a perfect reproduction was not possible. Even after fixing the skipped steps, the workflow still didn’t fully match the tutorial. Some parameters differed and a few steps were skipped or run with defaults.
3. There was also one parameter I had to change intentionally. The tutorial uses a first class_ranker threshold of 0.21, but in this run the class average scores were much lower (~0.01–0.04), so 0.21 selected zero particles. I lowered it to 0.01 to get enough particles for Topaz training.
2. I compared what Claude had done with the tutorial and several things were different. The full-dataset 2D classification step had been skipped. The class_ranker threshold was wrong (0.21 instead of 0.1). Bayesian polishing was run with default parameters instead of the tutorial settings.
1. After going back through the RELION tutorial run with Claude and correcting several issues, the reconstruction reached 2.9 Å, matching the tutorial result.
The first run had given 3.22 Å, which led me to look more carefully at what had happened in the processing. So I reviewed the workflow again.
At this stage Claude clearly doesn’t fully understand the processing logic yet, so it will be interesting to see how it performs on simpler datasets that are not structured like tutorials. For now, we still need to check the results at some key steps in the workflow. but we’ll see how it evolves(2)
Thanks for the comment. This wasn’t fully automated — I mainly tested whether Claude could follow the RELION tutorial workflow. I did step in to check the particle picking result and the maps after class3D. It also skipped the reference-free 2D classification. (1)
What an exciting result! Looking forward to seeing how it improves with higher magnification.
initial model - class 3d - post-processing
it’s been a while. Since I have no local workstation for data processing in my new role yet, I’ve been playing with AWS.
In my downtime, I asked Claude Code to follow a RELION tutorial from scratch. With a few fixes, I got a 3.22 Å result.
Not bad for AI handling routine tasks! :P
Attending the Scipion workshop at KAIST. Looking forward to learning a lot over the next 4 days
Congrats!!!
Talented researchers from Rho lab at Seoul National University just published impressive results.
Glad I could help with some of the cryo-EM data collection for this work.
www.science.org/doi/10.1126/...
Congratulations, Oli!
Truly fantastic work. The more small protein research, the better!
Also, please say hello to Kookju as well.
I carried out over the past two years. It’s very disappointing that my contribution leading to this achievement wasn’t recognized.
That said, I wish the authors and everyone involved the best of luck. If you’re interested in the paper, please reach out to the corresponding authors listed there.
I think this will be my last post on this topic. I don’t want to keep talking about the authorship issue. Even though I wasn’t directly involved in the experiments/analyses for the paper, I believe this achievement wouldn’t have been possible without the discussions I led on SNS and the work
The structures and maps in the paper have been newly deposited.
I was also told that the author list consists of people who directly participated in the actual paper writing, and they believe this is the right approach.
I contacted the first author and received specific reasons:
The data was newly processed after I left the company, and there is additional data as well, with ongoing data collection planned for better results.
At this point, it seems like including my name in the author list wouldn’t have much meaning. I just want to hear from the paper authors why I had to be excluded from this result. 3/3
After receiving this response, I'm so bewildered that I can't think straight. Looking at the paper's author contributions, 'H.K.' is included. Of course, there's no name corresponding to H.K. in the author list. 2/3
Just received this response about why I was excluded from authorship on the small protein cryo-EM paper "We have to respect the opinions of the members who actually conducted the research. What they confirmed is that the research was conducted after Dr. Kim left the company." 1/3
PNCC Cryo-EM modeling and validation workshop on Sep 3-5, 2025. Register now! tinyurl.com/Cryo-EMModel...
We tried to run RELION but encountered problems during data processing.
I’m not sure if some of the authors have much experience with RELION, but I can ask about that.
As you said, PLK1 is probably too small, making particle alignment very difficult.
Anyway, I will discuss this with the authors about how to address the issue. I will keep you updated.
Once again, thank you.
Thanks for the kind words, Oli.
The last time I processed the PLK1 map, it looked like the images I attached.
We also added 30° tilted micrographs for processing. It improved the orientation diagnostics results. However, the map is still anisotropic.
but I could not finish any of them and it looks like they wrote the paper with the same data I was working on.
Anyway, I will contact the authors to bring up some issues discussed here.
BTW, MBP structure is available.
www.rcsb.org/structure/8YBE
@olibclarke.bsky.social
@jhschaef.bsky.social
When I was at BaobabAiBIO, I was hesitant to publish the results because the PLK1 map was incomplete (I wrote about my concerns on Twitter).
Before quitting my job, I tried to finish at least these 2 things:
- getting a complete PLK1 map
- uploading MBP/PLK1 data to EMPIAR