
Thrilled to announce Zatom-2, an atomistic generative model spanning molecular, material, and protein science! Special thanks to @MirunaCretu2 for leading this effort. (1/11) 🌐 Project page: zatom-ai.github.io/nucleus

Hopper Postdoctoral Fellow @BerkeleyLab. Prev: #MachineLearning & #CompBio PhD @Mizzou; Research Intern @ Profluent & Absci. #DeepLearning & #GenerativeModels.

Thrilled to announce Zatom-2, an atomistic generative model spanning molecular, material, and protein science! Special thanks to @MirunaCretu2 for leading this effort. (1/11) 🌐 Project page: zatom-ai.github.io/nucleus

🕸️ LoG 2026 lands in Boston, Nov 20–22 @ Northeastern! 🎤 Keynotes: @PetarV_93, @YizhouSun, @WengongJin, Anna Scaglione & Rajmonda Caceres 📚 Hands-on tutorials, orals, posters & socials Only 10 days to go! Grab your spot 👇 logconference.org/register

Interested in keeping up with advancements in AI for Science? Me too. That's why I started a monthly blog covering new papers in the field. Feel free to check it out! Substack: amorehead.substack.com/

💻 Open-source code for Nucleus, our new monorepo featuring Zatom-2 and Zevals, a standardized evaluation suite for atomistic data: github.com/Zatom-AI/nucleus

🧬 If you’d like to contribute to the next generation of open-source atomistic generative models in chemistry and biology, please reach out! (11/11)

After finetuning, Zatom-2 can generate all-atom proteins by way of SCOPe. (10/11)

Zatom-2 also supports generative sampling of periodic materials via OMat24. (9/11)

Zatom-2 enables generative sampling of diverse molecular conformations thanks to OMol25. (8/11)

Analyzing representations, scaling behavior, and generation quality, our results show that joint pretraining across molecular+materials data improves generative modeling and representation transfer, pointing toward general-purpose generative models for atomistic systems. (7/11)

Furthermore, we demonstrate that conditioning on atomic forces enables control over the degree of geometric relaxation in generated structures. (6/11)

Zatom-2 establishes strong performance on QM9, GEOM-Drugs, MP20, OMat24, and OMol25. When transferring to as few as 2,000 protein training examples, Zatom-2 generates designable backbones and extrapolates strongly to sequence lengths beyond the training distribution. (5/11)

This multitask pretraining leverages the rich geometric information in quantum-chemical calculations to shape the model's internal representations. (4/11)

Uniquely, Zatom-2 is pretrained on the large-scale OMol25 and OMat24 datasets using a multitask objective that combines flow matching generation and structure prediction with auxiliary machine learning interatomic potential (MLIP) tasks predicting energies and forces. (3/11)

Atomistic machine learning models typically target specific domains, such as small molecules, crystals, or proteins. Here, we introduce a unified atomistic generative model across these domains thanks to Zatom-2’s new atom1+atom14 tokenization scheme for chem+bio data. (2/11)

📢 Only two days left to submit to LoG 2026! We welcome submissions on all areas broadly related to learning on graphs and geometry. 🔗 logconference.org/cfp/ 📅 Key dates (AoE): • July 29 – Abstract deadline • August 1 – Full paper deadline

🧬 MegaFold: Efficient AlphaFold-3 Training Across NVIDIA and AMD GPUs MegaFold will be presented at ISC High Performance 2026! 🗺️Research Paper Session: ML Systems and Frameworks 🗓️ Wednesday, June 24 · 1:00–2:00 PM (CEST) 📍Hall E, 2nd Floor, Hamburg, Germany Come find out how we train AlphaFold-3 on 3.36× longer sequences, up to 1.73× faster, on both NVIDIA and AMD GPUs. Using MegaFold, we can successfully train on H200s, MI250 and MI325X. 🚀 Think training AlphaFold-3 (AF3) is just like training an LLM? Think again. Despite having only ~500M parameters, AF3 is up to two orders of magnitude slower to train than a size-equivalent LLM and OOMs on sequences as short as 590 residues. Why? One operator: 3D attention over 2D pairwise representations. It makes compute and memory scale cubically with sequence length, blowing up activation memory, fragmenting execution into tens of thousans of tiny kernels per iteration, and starving the GPU while the CPU grinds through evolutionary-context preprocessing. In our latest work, MegaFold, we diagnose exactly why AF3 training is slow and build an end-to-end system that fixes it across the entire stack. The results: ✅ 3.36× longer trainable sequences on 32 GPUs ✅ Up to 1.73× faster per-iteration training (1.62× on AMD) ✅ 23% lower peak memory ✅ Runs on both NVIDIA H200, AMD MI250 & AMD MI325x, fully cross-platform 🔧 Four key innovations: - EvoFlash-3D: a cross-platform Triton kernel that computes 3D attention without ever materializing the cubic attention map (cutting EvoAttention memory by 4.24×). - EvoSP-3D: a communication-efficient sharding strategy built for alternating-axis 2D pairwise representations - EvoFusion: a fused operator stack that tames AF3's kernel fragmentation - EvoPipe: a determinism-aware host-device pipeline that eliminates preprocessing stalls And it's verified: MegaFold matches baseline training loss step-for-step, so you get all the speed and memory savings with zero impact on convergence. Huge thanks to our students Hoa La and Ahan Gupta for working hard on this project, and to our collaborators Jianlin Cheng and Alex Morehead for their valuable contributions, feedback, and support throughout the work! MegaFold is fully open source. 🎉 💻 Code: github.com/Supercomputing-Sy… 🔗 Project page: supercomputing-system-ai-lab… 📄 Paper: arxiv.org/pdf/2506.20686 📝 Blog: supercomputing-system-ai-lab… See you at ISC, Hall E! 👋