Such an incredible journey building DISCO 🪩. Love working with this team. DISCO is a co-design model with functional, de novo, new-to-nature enzymes. Huge shoutout to my co-authors for making this a reality! 🚀👇
Come join us in Vienna! 🇦🇹 I started via this call 6 months ago and couldn't ask for a better place to do AI/ML research. It's a highly collaborative and dynamic institute. Don't hesitate to reach out if you're thinking of applying and have questions!
AITHYRA is recruiting a LIFE SCIENCE PI
We are interested in research questions related to one of our four core disease areas: cancer research, immunological disorders, infectious diseases, neurodevelopment disorders.
Job Ad: aithyra.at/fileadmin/do...
Apply by 30.4.2026
#StartingPI #LS #AITHYRA
6 Papers (2 Oral 4 poster) at #ICLR from my group amoung the many more from #AITHYRA. See you in Rio!
Great work led by @marcostock.bsky.social and Florin Ratajczak , thanks to all coauthors Paul Bertin
@evahoermanseder.bsky.social @yoshuabengio.bsky.social Jason Hartford @interactome.bsky.social Matthias Heinig @alextong.bsky.social
AITHYRA Global Adjunct Principal Investigator Program:
Advancing AI-Driven Life Science Through Global Collaborations
Learn more about the Program: aithyra.at/fileadmin/do...
Application Deadline, 30 January 2026: application@aithyra.at
#AITHYRA #GlobalAdjunctPI
Danyal Rehman, Tara Akhound-Sadegh, Artem Gazizov, Yoshua Bengio, Alexander Tong: FALCON: Few-step Accurate Likelihoods for Continuous Flows https://arxiv.org/abs/2512.09914 https://arxiv.org/pdf/2512.09914 https://arxiv.org/html/2512.09914
Interesting idea! We didn't try this but this seems quite feasible as a conditional generation / inpainting problem. Would love to try it out.
Also thanks for the typo catch. Will be reflected in the updated version thanks :)
“One of the continuing scandals in the physical sciences is that it remains in general impossible to predict the structure of even the simplest crystalline solids” (John Maddox)
OXtal is a new all-atom generative diffusion model addressing this holy grail problem
Welcome to a new chapter in molecular materials design 🚀.
Read the full deep dive here: oxtal.github.io
🧵8/8
As well as institutional support:
@ox.ac.uk
@aithyra.bsky.social
@mila-quebec.bsky.social
@caltech.edu
FutureHouse
This work wouldn’t have been possible without a rockstar team who provided immeasurable support:
@mgalkin.bsky.social,
@jarridrb.bsky.social,
@inequivariant.bsky.social,
Santiago Miret,
@francesarnold.bsky.social,
@mmbronstein.bsky.social
@joeybose.bsky.social
Transparency: CSP isn’t solved, but OXtal enables high-throughput CSP screening upstream of physics-based ranking.
Future work will address ranking, sample efficiency, and most importantly, the direct design of molecular crystals ⚛️
🧵6/8
To solve the "infinite lattice" problem on a GPU, we developed S4.
Instead of predicting the lattice directly, we train on local neighborhoods ("shells"). We teach the model how molecules sit next to each other.
At inference, global periodicity naturally emerges from these local rules.
🧵5/8
Organic crystals are chaotic. Unlike proteins (structured by backbones) or inorganic materials (strong bonds), organic crystals are held together by weak, finicky forces.
Simulating usually requires expensive search that takes days.
We wanted to do it in seconds at a fraction of the cost. 💰
🧵4/8
OXtal generates crystal structures conditioned on 2D molecular graphs.
Trained on 600K data points, it works with:
- Rigid molecules
- Flexible molecules
- Co-crystals
Spanning applications in drug discovery 💊 and organic electronics 💡.
🧵3/8
How do we turn a 2D graph into a 3D crystal? OXtal makes 3 bold choices to enable scalability:
📈 Ditches explicit equivariance for scaling & soft symmetries
✨ Models growth via "Stoichiometric Stochastic Shell Sampling" (S4)
🔮Skips unit-cells to generate full supercells
🧵2/8
🔮Introducing OXtal – a new all-atom diffusion model for crystal structure prediction!
We tackle a grand challenge in computational chemistry: predicting the structure of crystalline solids directly from their chemical composition.
Paper: arxiv.org/abs/2512.06987
Blog Post: oxtal.github.io
1/8
🆕 “Foundations of Diffusion Models in General State Spaces: A Self-Contained Introduction”
Huge thanks to Tobias Hoppe, @k-neklyudov.bsky.social,
@alextong.bsky.social, Stefan Bauer and @andreadittadi.bsky.social for their supervision! 🙌
arxiv : arxiv.org/abs/2512.05092 🧵👇
#AITHYRA, Vienna's new Biomedical AI institute, is hiring Postdocs!
Come work with us. Openings in: 🔹 Generative AI 🔹 Multimodal ML 🔹 Virology 🔹 Enzyme Function
Apply by Nov 20: oeaw.ac.at/aithyra/post... #PostDoc #AI #ML #Vienna #ScienceJobs
Registration for this years CHAIR Structured Learning Workshop is open. Speakers include: Klaus Robert Müller, Jens Sjölund, @alextong.bsky.social ,
@janstuehmer.bsky.social, @arnauddoucet.bsky.social, @marcocuturi.bsky.social , Marta Betcke,
Elena Agliari, Beatriz Seoane, Alessandro Ingrosso
SuperDiff goes super big!
- Spotlight at #ICLR2025!🥳
- Stable Diffusion XL pipeline on HuggingFace huggingface.co/superdiff/su... made by Viktor Ohanesian
- New results for molecules in the camera-ready arxiv.org/abs/2412.17762
Let's celebrate with a prompt guessing game in the thread👇
I've spent more compute time than I care to admit calculating the likelihood under flow models. Now 5x-10x faster likelihoods for stochastic flows🤯. Check out our work on SuperDiff for a great use case of model mixing.
🔊 Super excited to announce the first ever Frontiers of Probabilistic Inference: Learning meets Sampling workshop at #ICLR2025 @iclr-conf.bsky.social!
🔗 website: sites.google.com/view/fpiwork...
🔥 Call for papers: sites.google.com/view/fpiwork...
more details in thread below👇 🧵
Amazing blog post on flow matching, stunning visuals! It also makes the connection with normalising flows crystal clear. Incredible effort!
Attending the Learning on Graphs conference (logconference.bsky.social) this year? Come check our introductory tutorial to building Geometric Generative Models co-delivered with Heli Ben-Hamu and
Alex Tong (alextong.bsky.social)
More details and forthcoming code: sites.google.com/view/ggm-log...
Huge shoutout to @alextong.bsky.social for making the awesome #TorchCFM library. Repos like this really unlocks the field and reduces entry barriers for others 😇
Thrilled to announce Boltz-1, the first open-source and commercially available model to achieve AlphaFold3-level accuracy on biomolecular structure prediction! An exciting collaboration with Jeremy, Saro, and an amazing team at MIT and Genesis Therapeutics. A thread!
