Michael Jendrusch
LightNews LightNews
banner
mjendrusch.bsky.social
Michael Jendrusch
@mjendrusch.bsky.social
810 followers 310 following 33 posts
(he/him) Former PhD student / Postdoc @ Korbel group, EMBL.
Protein ML person, mathematics & science enthusiast.

developer of salad
preprint: https://www.biorxiv.org/content/10.1101/2025.01.31.635780v1
code: https://github.com/mjendrusch/salad
Posts Replies Media Videos
mjendrusch.bsky.social
Thank you!
September 24, 2025 at 4:36 PM
mjendrusch.bsky.social
Also, shoutout to @blender.org and @bradyajohnston.bsky.social for making it so easy to turn protein structures into nice images!
5/5
Protein structures colored teal and purple floating above a reflective surface on a dark background. The protein structures spell out the word "SALAD" in all-capital letters.
September 24, 2025 at 12:27 PM
mjendrusch.bsky.social
While I was working on salad as part of my PhD at Korbel group at EMBL, it was very much a passion project. So it makes me very happy to see @embl.org feature this paper!
Thanks again to Jan and the team supporting me on my more non-standard projects!
4/🧵
September 24, 2025 at 12:27 PM
mjendrusch.bsky.social
Compared to the preprint, we have added comparisons to Proteína (openreview.net/forum?id=TVQLu34…) in terms of unconditional design. Here, salad compares favorably in terms of both runtime per design and designability of generated structures.
3/🧵
Graph of runtime comparisons of diffusion models (per design: top; per iteration: botton). This panel compares the runtimes of previous protein diffusion models (RFdiffusion, Genie, Proteína, Chroma) to two versions of the salad model. Salad outperforms all previous models in terms of runtime across protein lengths between 50 and 1,000 amino acid residues. Boxplot of self-consistent RMSDs (scRMSDs) for salad models with different noise schedules (variance preserving, variance expanding, shaped) and the previous state of the art (RSO, Proteína), for proteins of length between 50 and 1,000 residues. Variance expanding and shaped noise salad models consistently exhibit lower (better) scRMSDs compared to previous approaches across all sizes. While Proteína produces comparable results to salad all the way up to 800 residues, it surprisingly fails for 1,000 residue proteins.
September 24, 2025 at 12:27 PM
mjendrusch.bsky.social
The preprint and code have been out for a while - if you want a quick overview, have a look at the thread we did back in February:
bsky.app/profile/mjendrus…
If you want to try salad in your browser, you can load up this Colab notebook:
colab.research.google.com/github/mjendrusc…
2/🧵
mjendrusch.bsky.social Michael Jendrusch @mjendrusch.bsky.social · Feb 6
New protein ML preprint from my PhD project.
We describe salad (sparse all-atom denoising), a family of efficient protein structure diffusion models and show that it works well on a bunch of protein design task previously described in the literature.

Preprint: www.biorxiv.org/content/10.1...

(1/N)
Five generated protein structures coloured with a gradient from purple to teal. The structures are shaped and positioned to spell out the word "salad" – the name of the software described in the mentioned preprint – in all-capital letters.
September 24, 2025 at 12:27 PM