iiinova — where your digital self keeps evolving

Reposted by Dimitry Tegunov

Przemek Dutka @przemekdut.bsky.social · 08/10/2026

Our latest development is now on bioRxiv: cryoSHIELDs — amphipathic peptide additives that make cryo-EM grid prep far more reliable. Tested across 24 protein systems, they enabled the first ABCC6 structure and GPCR–G protein structures at just 0.1–0.5 mg/mL! www.biorxiv.org/content/10.6...

14626

Dimitry Tegunov @dtegunov.bsky.social · 17/09/2026

There are countless houseboats in Sausalito, but this is the first time I see one on the move

030

Dimitry Tegunov @dtegunov.bsky.social · 25/08/2026

This Relay thing sure is getting popular in Europe now

080

Dimitry Tegunov @dtegunov.bsky.social · 24/08/2026

I never realized MiG-17 had an afterburner! Looks good on it

120

Reposted by Dimitry Tegunov

Przemek Dutka @przemekdut.bsky.social · 11/08/2026

Interested in cutting-edge industrial cryo-EM/ET? Join us!

066

Reposted by Dimitry Tegunov

Matt Johnson @drmatthewcjohnson.bsky.social · 11/08/2026

Interested in a career in industry cryoEM? We're hiring! We seek a cryoEM/cryoET expert to join our team at Genentech advance our mission to provide high-resolution structural insights into biological macromolecules involved in health and disease. Apply at: careers.gene.com/us/en/job/20...

careers.gene.com

Systems Specialist II in South San Francisco, California, United States of America | Manufacturing & Engineering at Genentech

Apply for Systems Specialist II job with Genentech in South San Francisco, California, United States of America. Manufacturing & Engineering at Genentech

01713

Dimitry Tegunov @dtegunov.bsky.social · 11/08/2026

Historically accurate: GDR relied on Bavaria disliking the Prussians so much they'd never let them in anyway – no wall needed

020

Dimitry Tegunov @dtegunov.bsky.social · 08/08/2026

I always assumed he meant it as a joke. Might as well call it in situ senso baumeisteriano

180

Dimitry Tegunov @dtegunov.bsky.social · 07/08/2026

It’s either that, or In Situ Plus, In Situ Pro, In Situ Ultra, and In Situ Max. With Max reserved strictly for when you put an entire frozen human in the TEM (is anyone working on that yet?)

1232

Dimitry Tegunov @dtegunov.bsky.social · 07/08/2026

We could even embrace the phonetics and take it to mean “involuntarily cellular,” i. e. you considered more efficient approaches first, but gave up and booked the FIB eventually.

1131

Dimitry Tegunov @dtegunov.bsky.social · 07/08/2026

As much as I find “in-cell” phonetically cringe for obvious internet culture reasons, it might be time to raise the bar to that and leave in situ for whenever there is any useful context around the protein.

1121

Dimitry Tegunov @dtegunov.bsky.social · 07/08/2026

Personally, I file those under in cheatu. Which isn't a bad thing: whenever you can get enough native context without having to FIB or even do tomography, you should do it.

160

Dimitry Tegunov @dtegunov.bsky.social · 18/07/2026

Maybe I misremember, but shaders went through similar rejection/acceptance phases. 1st gen hardware was too slow, shader model 1.x too limiting, and support in games lacking. SM 3.0 was roughly when they became mainstream. Then it flared up again when games began to drop fixed-function support.

111

Dimitry Tegunov @dtegunov.bsky.social · 13/07/2026

Strict C2D can help get the right initial reference from a very dirty particle set, but after that it's back to the original set to get junk attractors, and only C3D from there.

140

Dimitry Tegunov @dtegunov.bsky.social · 19/06/2026

These helices look eerily similar to the “artisanal proteins” from yesterday's GRC skit, including the colors

030

Dimitry Tegunov @dtegunov.bsky.social · 19/06/2026

If you do, a publicly available Warp- and MA-processed version of the chlamy dataset would be much appreciated 😘

070

Dimitry Tegunov @dtegunov.bsky.social · 13/06/2026

This is annoying even with lower-tier models. I've been using Sonnet to assign expertise tags to poster judges at the upcoming 3DEM GRC based on their paper titles, and one virologist would reliably trigger the safeguards

140

Dimitry Tegunov @dtegunov.bsky.social · 13/06/2026

So many people waking up today wishing they had prompted Fable a bit longer last night 😱 The government's reaction is completely overblown, of course. But it is a bit better than Opus 4.8 on average – enough that I wouldn't go back voluntarily if cost wasn't a factor

020

Reposted by Dimitry Tegunov

cryoEM papers @cryoempapers.bsky.social · 27/05/2026

CryoARC: Atomic-resolution conformational landscapes of protein assemblies from cryo-EM single particles with evolutionary priors www.biorxiv.org/content/10.64898/2026.05.25.727696v1 #cryoEM

0115

Dimitry Tegunov @dtegunov.bsky.social · 04/05/2026

Unless you ask PyTorch for determinism (rarely done in production because it's very slow), ops involving scattering won't be deterministic. Do you have a gif comparing 2 outputs with identical postprocessing? The high-pass filter in your example makes it hard to compare denoising specifically.

120

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

You've done an amazing job! I really had a blast working with you, Alister, and Miguel.

120

Reposted by Dimitry Tegunov

Marten Chaillet @martenchaillet.bsky.social · 03/05/2026

MissAlignment is out!!! Our new software that improves reference-free tilt series alignment by mimicking human intuition. Check out Dimitry's beautiful thread for a quick summary! And the preprint for more details: www.biorxiv.org/content/10.6... The program: github.com/warpem/miss-...

biorxiv.org

67518

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Patches, projections, L2 maximization: I've done them all. The birds and bananas wouldn't go away. Then @martenchaillet.bsky.social came along, right after finishing his PhD w/ @fridof.bsky.social. I don't want to jinx it, but I think I'm finally done with bananas. The future belongs to pineapples!

0260

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

This project had an interesting personal angle: over 10 years ago, I joined @fridof.bsky.social's lab as an undergrad to work on reference-free tilt series alignment. I failed. I've been giving it another try with a new idea ca. once every year since then.

1190

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

We hope you're as excited about your future alignments as we are! Now finding and refining ribosomes before you can do anything else with your data is optional. Just kidding, because why else would you be doing cryo-ET if not for the beautiful ribosomes?

1160

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

What if you're not interested in ribosomes? First, HOW DARE YOU. Second, MissAlignment's got you covered. We took a tiny subset of EMPIAR-11830 containing parts of the nucleus and looked for nucleosomes. Here, further reference-based alignment with M made no difference.

1333

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Finally, we integrate it with Warp and M's geometry model to account for local deformations. And now look what it's done with the 70S ribosomes from EMPIAR-10499. Most importantly, there isn't much left for M to improve in terms of alignment (it can still optimize other things like CTF though).

1190

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Crucially, this trained loss function can extrapolate beyond the accuracy of the best alignments shown in training. Not perfectly at first, but we run multiple cycles of training and alignment to provide it with increasingly better alignments in training.

180

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Then we freeze the trained model and show it reconstructions with the current alignments. The entire forward model is implemented in PyTorch and thus fully differentiable, so we can just do backpropagation to ask it how to change the alignments to lower the final misalignment score.

1120

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

But we're starting with bad alignments! Where do we get the good ones? We don't. We just make the existing alignments even worse by adding different types of noise to the parameters. You know: to outrun a bear, run faster than the other person.

1120

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Instead, @martenchaillet.bsky.social developed an unsupervised training procedure where the model is asked to score a reconstruction with good alignment lower than one with bad alignment. No pre-defined values – just a contrastive loss with triplets of examples.

190

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

But some features are rather elusive.

1130

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

We also could have had @martenchaillet.bsky.social characterize the features and distill them into a traditional image processing algorithm.

Marten investigates bananas and pineapples.

1130

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

It's tempting to just hook up a human to a computer and have their intuition nudge the alignment in the right direction. But this doesn't scale well.

An imaginary experiment where human brain activity is measured to guide tilt series alignment.

1140

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

Here is what these patterns look like in real data. Very easy to recognize for humans, but a machine maximizing the cross-correlation between tilt images will happily go bananas. If only machines could learn.

Examples of good (1) and bad (2, 3) tilt series alignments in real data.

1150

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

You all know the situation: etomo or AreTomo tell you the alignment is really accurate, but you take a quick look at the reconstruction and immediately know it isn't. How? Point features resemble double-ended pineapples with good alignment, but turn into bananas and birds with bad alignment.

Patterns to look for when judging tilt series alignment accuracy: double-ended pineapples = good, bananas and birds = bad.

1180

Dimitry Tegunov @dtegunov.bsky.social · 03/05/2026

We're super excited to share MissAlignment: a new ML-based approach to reference-free tilt series alignment, spearheaded by @martenchaillet.bsky.social. We think it's going to make your cryo-ET life a lot better. Preprint: www.biorxiv.org/content/10.6... Code: github.com/warpem/miss-... 🧶 Thread:

6313118

Dimitry Tegunov @dtegunov.bsky.social · 30/04/2026

Now my mailing list reply bot is getting spam targeted specifically at agents: containing an API key for the advertised service, and instructions to try it out and to copy part of the message as a skill .md to Claude's config. I wonder how often Claws running without any human oversight bite on that

130

Dimitry Tegunov @dtegunov.bsky.social · 09/04/2026

Looks like electron cryo-microscopy/tomography has finally lost the battle

3130

Dimitry Tegunov @dtegunov.bsky.social · 01/03/2026

Thorsten, I am so sorry to read this. Thank you for sharing – it takes real courage. I wish you and your family great strength. Your work has meant a lot to our field and to me personally.

010

Dimitry Tegunov @dtegunov.bsky.social · 16/02/2026

Somewhat related: I was looking for a JS library to read MRC files a while ago, but couldn't find one (sorry if I missed it). So you can now “npm install @warpem/mrc-parser” github.com/warpem/mrc-p...

github.com

140

Dimitry Tegunov @dtegunov.bsky.social · 16/02/2026

I had Claude make this as a mini-game for Relay's disconnect screen (like Chrome's dinosaur), but then pivoted to stand-alone. In this game, you have to find Euler angles matching the target projection, i.e. cosplay RELION. It can be pretty hard. Works on mobile too! warpem.github.io/euler-hunt/

warpem.github.io

Euler Hunt

Can you find the Euler angles of a cryo-EM projection?

04615

Dimitry Tegunov @dtegunov.bsky.social · 13/02/2026

A cats lock key that lets you quickly disable the keyboard when a cat approaches it would be a lot more useful than the antiquated caps lock

Caps lock button on an Apple keyboard, activated by a cat stepping on it

3181

Dimitry Tegunov @dtegunov.bsky.social · 12/02/2026

Ahh, that must be the one I heard about!

100

Dimitry Tegunov @dtegunov.bsky.social · 12/02/2026

What is the current resolution record for in situ nucleosomes? The best unofficial result I remember seeing was around 7 Å, but I feel like I've heard a better figure mentioned somewhere already

210

Dimitry Tegunov @dtegunov.bsky.social · 06/02/2026

California is famous for its sunsets, but sunny early mornings are the real magic here

010