Leon Hetzel

Thrilled to announce that we just presented „MAGNet: Motif-Agnostic Generation of Molecules from Scaffolds“ at #ICLR2025 🧲 @j_m_sommer @Pseudomanifold @fabian_theis @guennemann For those who couldn’t make it to our spotlight: openreview.net/forum?id=5FXKg…

Finding good resources on efficient AI is harder than it should be. We're fixing that! 🚀 Check out our new github.com/PrunaAI/awesom… repo —a curated hub of the best tools, papers, and techniques to make AI faster, smaller, and cheaper.

Congrats to my amazing PhD students: We have 9 papers accepted at #ICLR2025. Reliability, AI4Science, graphs, LLMs, and more (go.tum.de/936150). And if you follow the recent discussions about AI efficiency, you might like our blog and webinars (pruna.ai).

It’s rainy in Vancouver, poster hall is closed but we are ready 🙌 👉 Come and talk to us and learn about UniGuide at Poster#2600 (East) UniGuide is a new framework for molecular diffusion models that enables flexible geometric conditioning across tasks—no retraining required

Unified Guidance for Geometry-Conditioned Molecular Generation • UniGuide introduces a unified framework for geometry-conditioned molecular generation using diffusion models. It provides a flexible method for controlled molecular design, eliminating the need for additional training or specialized networks. • Key innovation: UniGuide employs “condition maps” to translate geometric constraints into actionable guidance, enabling applications in structure-based, fragment-based, and ligand-based drug design with a single model. • Unlike traditional conditional models, UniGuide excels in handling diverse geometric constraints like surfaces, densities, and volumes. This versatility allows it to generalize across a range of tasks without the limitations of task-specific models. • Performance benchmarks show UniGuide’s superior or comparable results to specialized approaches in drug discovery tasks, demonstrating its adaptability and efficiency. • Applications include ligand-based drug design (LBDD) for generating novel molecules that match predefined shapes, structure-based drug design (SBDD) for ligand-pocket interactions, and fragment-based drug design (FBDD) for linking and growing molecular fragments. • UniGuide’s method decouples training from conditioning, making it robust in data-scarce scenarios. It demonstrates state-of-the-art performance across ligand similarity, binding affinity, and drug-likeness metrics. @guennemann @fabian_theis @j_m_sommer @leon_het @sirine_ayadi1 📜Paper: openreview.net/pdf/704d9702fd… #MolecularDesign #DrugDiscovery #GeometryConditioning #AIinBiology #GenerativeModels

11 exciting news: Our group has 10 papers at #NeurIPS2024 (incl. 1 oral + 2 spotlights) 📃🎓. And as of October 1st, I am on entrepreneurial leave 🚀. Re papers: Congrats to all co-authors. Amazing work! go.tum.de/689644 Re startup: We are hiring! pruna.ai

Three papers @NeurIPSConf 2024 🎉 An efficient adv. training algorithm for LLMs arxiv.org/abs/2405.15589 Unlearned LLMs are not safe against adv. attacks arxiv.org/abs/2402.09063 Scaling robustness of Lipschitz-1 networks arxiv.org/abs/2305.10388 Happy to chat in Vancouver!

Presenting at #ICML2024: Fragment-Biases for Molecular GNNs 🧪 Tue, 23.07: Oral session 1F @ 11:00 Poster #105 @ 11:30 🔑 Fragment-Biased GNNs outperform others, match Transformers with better generalization & linear cost! 🤝 With N. Kemper, @leon_het , @j_m_sommer , @guennemann

@y0b1byte @magdalenoxford @KyriacosShiarli 5/7 perfect score. What a gem.

Oxford trip update: The city is as beautiful as it was. The covered market is even better and works till 11pm. @magdalenoxford is as splendid as it used to be. The students are drunk, excited, and stressed at the same time as usual. @KyriacosShiarli is the best. The life goes on, the Dude abides. Perfect 5 out of 7, miss it a lot, will come here again.

Mixed Models with Multiple Instance Learning (MixMIL) received an Oral & Outstanding Student Paper award at @aistats_conf last week! 🏆 MixMIL enables accurate & interpretable patient label prediction from single-cell data by adding attention to GLMMs.#singlecell #MachineLearning

Get to know #mcml junior member @leon_het 💡 He developed an algorithm called chemCPA designed to make drug discovery more effective. Read more (in German): mcml.ai/news/2024-03-2… This article was written by @leonie__fischer from @DJSde!

@AlieeHana @aistats_conf Same here

Thrilled to present our paper "(Provable) Adversarial Robustness for Group Equivariant Tasks" at #NeurIPS2023! Joint work with great collaborators @YanScholten @guennemann! Paper: arxiv.org/abs/2312.02708 Poster #800 Thursday 11 am - 13 pm #NeurIPS

@dagarfield @lauradmartens @gagneurlab @fabian_theis It works for me on mobile, so I assume it is firewall related?

@lauradmartens @gagneurlab @fabian_theis Glad to see this out! Can you double check the posted link? Don't know if it is a weird firewall thing or not, but I'm not seeing it.

Is binarization of scATAC-seq data really necessary? The conclusion from our analysis is that a quantitative treatment is in fact beneficial. Now out in Nature Methods! @gagneurlab @fabian_theis nature.com/articles/s4159… Many additions since the preprint 👇(1/n)

Do you care about uncertainty quantification? Do you like guarantees? Check out our #ICML2023 paper where we bring conformal prediction to Graph Neural Networks. tldr: Instead of a single label return a *set* that is guaranteed to contain the true label. Set size = uncertainty.

If you are interested in ML potentials and/or uncertainty estimation come and visit our presentation of "Uncertainty Estimation for Molecules: Desiderata and Methods" this week at #ICML! Tue, 11am, #415 Joint work with @n_gao96 @Bertrand_Charp @amine_ketata @guennemann 1/2

1/9 CPAis finally published. It can predict single-cell responses to combinatorial perturbation (drugs, CRISPR). This is a joint collab between @meta and @fabian_theis @HelmholtzMunich. Read the thread to understand the LEGO analogy! embopress.org/doi/full/10.15…

Happy to announce our new paper on Deterministic Uncertainty Methods @TMLunLimited #ICLR2023 ! We dissect how the design of training schemes, architecture, and prior can significantly impact feature collapse and uncertainty performance! w/ C. Zhang and @guennemann

Great talk Bastian, thanks for uploading!