Frank Noe

Good to come home after a trip 😂. @cecclementi.bsky.social

Absolutely. It’s more than a philosophical discussion - knowing the interpretation context helps the brain to remember abstract math, especially when controversial stories are involved.

Crick researchers to collaborate with Microsoft Research AI for Science to map protein motion A new collaboration between researchers in the Crick’s Visual Biochemistry Lab and Microsoft Research AI for Science will combine cutting-edge laboratory experiments with artificial intelligence, to b...

Researchers in the Crick’s Visual Biochemistry Lab are teaming up with Microsoft Research AI for Science to combine cutting-edge lab experiments with AI to better understand and predict how proteins move and interact.
@franknoe.bsky.social

www.crick.ac.uk/news/2025-07...

🇪🇸🍷💪

It’s very good!

2 big papers in the family in a week ==> good wine. With @cecclementi.bsky.social

Congrats to all authors and thanks to @freieuniversitaet.bsky.social and @einsteinberlin.bsky.social for continued support!

So where are we? This Protein CG potential is the first of its kind, and there are many things that can be improved in the model and training data. However, CGSchNet demonstrates that systematically approximating all-atom distributions is possible with a CG force field that is ~100x faster.

And by simulating the folded an unfolded state and standard reweighting tricks, we can compute changes in the folding free energy of mutants. Trying to do this with an all-atom model would be crazy because so many conformations need to be sampled. CGSchNet does it easily.

Zero-shot test: simulate folding-upon-binding. Nothing like this has been used for training. An intrinsically disordered peptide, PUMA, binds and folds to its target protein, which does not happen to an off-target protein.

Comparing CGSchNet and other CG models shows the difference in predictiveness

This also works well for bigger proteins, where folding with all-atom MD requires true "hero runs" and is not economically reasonable.

Other CG protein models can do structure prediction (now AlphaFold does that better) or model protein-association with fixed protein structures, but modeling protein dynamics and intermediates, e.g., during folding, had not been achieved. CGSchNet shows that a deep learning CG potential can do that.

@cecclementi.bsky.social's team performed 1000s of all-atom simulations of protein fragments, developed a deep learning architecture and physical priors and trained the model to achieve CGSchNet, the first CG model that can approximate all-atom distributions across the protein universe.

Awesome to see this epic piece of work, led by @cecclementi.bsky.social, finally appear in @natchem.nature.com.

The development of a general coarse-grained protein forcefield to describe folding, binding and conformation changes without solvent and all-atom, has been long anticipated!

A high-resolution color image of Mars with the small, dark, potato shaped satellite Phobos, seen in sharp contrast as it passes over the massive volcano Olympus Mons. Phobos appears irregular and shadowed against the softly rust colored Martian surface. In the distance, two other volcanic features Ascraeus Mons (upper right) and the flatter Alba Mons (upper left) are visible. Wisps of light clouds and the blackness of space frame the planet’s curved horizon.

1/n 🧵
Just released: NEW images of Phobos over Mars by @esa.int Mars Express 🔭🧪

This view shows Phobos above Olympus Mons!

Full-size (130MP) image & details: flic.kr/p/2rggHKy
Credit: ESA/DLR/FU Berlin/Andrea Luck CC BY

Captured on May 14, 2025 | Image ID: HQ967
Raw data from: psa.esa.int

Scalable emulation of protein equilibrium ensembles with generative deep learning Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that...

BioEmu is now on @science.org ! The revised version includes an upgraded model and makes a lot of MD simulation data internally generated at MSR available to the public. This took a lot of firepower from us in the last two years.
www.science.org/doi/10.1126/...

Microsoft AI predicts protein conformations The open-source tool goes beyond AlphaFold by finding proteins’ multiple equilibrium states and free energies

@maxhenrybarnhart.bsky.social has written a really nice piece on BioEmu for Chemical & Engineering news.

cen.acs.org/biological-c...

Ein einfarbiger Hintergrund, darauf der Claim "Neues aus der Forschung"

🚀 Durchbruch in der Proteinmodellierung! Mit #BioEmu hat ein Team der #FUBerlin & @msftresearch.bsky.social unter der Leitung von @franknoe.bsky.social ein KI-System vorgestellt, das Proteinbewegungen so genau simuliert wie nie zuvor.

Mehr dazu ➡️ www.fu-berlin.de/presse/infor...

Thanks Andrew!

Yes!

Definitely re-read the protein stability prediction section. TLDR, model generalizes better in predicting protein stabilities, now also mutants work quite well.

This is the way

Protein stability funktioniert jetzt besser. Auch für Mutanten!

Thanks Oles!

Thanks Pierpaolo!

Thank you man!

Scalable emulation of protein equilibrium ensembles with generative deep learning Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that...

Ryota @ryotat.bsky.social and Cecilia @cecclementi.bsky.social.

Work done at @msftresearch.bsky.social AI for Science!

Thanks to @freieuniversitaet.bsky.social for support

Check out the full paper here:
www.science.org/doi/10.1126/...

Scalable emulation of protein equilibrium ensembles with generative deep learning Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that...

Adam @adamefoster.bsky.social, Arne @rne.bsky.social , Jigyasa @jnigam.bsky.social, Federico Barbero, Vincent Stimper, Andrew Campbell, Jason @jyim.bsky.social, Marten @martenlienen.bsky.social, Yu Shi, Shuxin Zheng, Hannes Schulz, Usman Munir, Roberto Sordillo, ...

Scalable emulation of protein equilibrium ensembles with generative deep learning Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that...

Huge shoutout to the outstanding team: Sarah Lewis, Tim Hempel, José @jjimenezluna.bsky.social, Michael Gastegger, Yu @yuxie.bsky.social, Andrew @andrewfoongyk.bsky.social, Victor Sartorras, Osama Abdin, Bas @basv.bsky.social, Iryna Zaporozhets, Yaoyi Chen, Soojung @soojungy.bsky.social ...