FutureHouse

14 rounds of directed evolution. That is what it previously took to engineer an enzyme for selective C(sp³)–H insertion—one of the most challenging transformations in organic chemistry. A single DISCO design matches that. No pre-specified catalytic residues, no theozymes, and no inverse folding. We are incredibly excited to introduce 🪩 DISCO (DIffusion for Sequence-structure CO-design), a multimodal generative model that simultaneously co-designs protein sequences and 3D structures from scratch, co-folding with any biomolecules. Evolution is an amazing chemist, but the reactions it has explored represent a remarkably narrow slice of what is possible. DISCO allows us to explore the chemistry nature never imagined. Here is a look at the key advances: 🧪 Unlocking New-to-Nature Chemistry: We challenged DISCO to design enzymes for carbene-transfer reactions—chemistry not known to biology. In addition to exceeding the directed evolution campaign for C(sp³)–H insertion in one step, a single DISCO design achieved 5,170 TTN for B–H insertion, outperforming three rounds of laboratory directed evolution by over 2x. 🧬 Unseen Active Sites: These active sites do not exist in nature. When searched against 200M+ structures in the AlphaFold Database, the majority of DISCO's binding motifs had no close natural homologs. Our top design, dCT-H11, completely repurposed a DNA transcription factor from a Dead Sea extremophile into a novel active-site geometry. 🧭 Multimodal Inference-Time Steering: In generating sequence and structure together, DISCO aligns the two modalities bidirectionally via cross-modal recycling, entropy-adaptive sequence temperature and sequence self-correction. This unlocks true on-the-fly steering. Using multimodal Feynman-Kac Correctors, we can enforce complex constraints during generation—such as engineering dense disulfide networks or designing specific binders that avoid decoy molecules. 📈 State-of-the-Art Benchmarks: DISCO achieves SOTA on 178 out of 179 molecular targets in our new Studio-179 benchmark, as well as DNA and RNA binders. It also shines in generating highly diverse, unconditional, and co-designable proteins. The chemistry nature never explored is now within our reach. 🌍🚀 We would love to hear your thoughts! Check out the details and the code below: 📝 Blog: https://lnkd.in/gqhZbugV 📄 Paper: https://lnkd.in/gi64bSeA 💻 Code: https://lnkd.in/gri67vJv So grateful for this incredible collaboration with Jarrid Rector-Brooks, Théophile Lambert, Marta Skreta, Daniel Roth, Yueming Long, Ziqi Li, Nicole Zhang, Miruna Cretu, Francesca-Zhoufan Li, Ph.D., Tanvi Ganapathy, Emily Jin, Joey Bose, Jason Yang, Kirill Neklyudov, Yoshua Bengio, Alexander Tong, and Frances Arnold Edit: The C-H insertion evolution campaign was first performed on another substrate (also ~2000 TTN). We wanted to convey C-H insertion is hard and many rounds of evolution was required to achieve reasonable activity.

FutureHouse is now accepting applications for our 2026 AI-for-Science Independent Postdoctoral Fellowship! We’re looking for exceptional early-career researchers who want to push the boundaries of discovery using cutting-edge AI tools. Fellows get a generous stipend, dedicated engineering support and access to compute, wet-lab capabilities, and the full Edison Scientific platform to pursue ambitious scientific questions. This year, we're excited to partner with The Kavli Foundation to offer one Fellowship to advance discoveries in the field of neuroscience. If you’re ready to chart your own path at the intersection of AI and biology, submit an application! 💡 Learn more: https://lnkd.in/eqPN4Xt6 🔗 Apply here: https://lnkd.in/gGP43pc8 📅 Applications are due February 13, 2026

Our Nature Biotechnology paper made the cover!! 🎉 Huge thanks to Xinyi Christine Zhang for the beautiful artwork. If you need someone who can magically turn awful sketches into exactly what you imagined, she’s the one. https://lnkd.in/epSxgPbT Thanks again to my amazing co-authors Delaney Sullivan, Maria Carilli, Kristjan Eldjarn, Alexander Viloria Winnett, Tara Chari & Lior Pachter, as well as the many mentors and sponsors who continue to make our work possible, including FutureHouse and Eric and Wendy Schmidt Center.

Today we announced the NVIDIA Nemotron™ 3 family of open models, data, and libraries, offering a transparent and efficient foundation for building specialized agentic AI across industries. Nemotron 3 features a hybrid mixture-of-experts (MoE) architecture and new open Nemotron pretraining and post-training datasets, paired with NeMo™ Gym, an open-source reinforcement learning library that enables scalable, verifiable agent training. Read more: https://nvda.ws/3YkDl54

I’ve been fascinated by the Crystal Structure Prediction (CSP) problem since I was an undergrad. We can perfectly define a 2D molecular graph - know every atom and bond - so why is it so expensive to predict how it packs in a solid? For decades, the standard answer has been brute force: burning thousands of CPU hours on physics-based simulations to hunt for intricate energy minima. It works, but it doesn't scale. Today, we’re sharing a different approach with OXtal. Instead of treating this as a search problem, we treated it as a generative modeling problem. We asked: Can a diffusion model learn the "chemical intuition" of how molecules pack directly from data? OXtal is an all-atom diffusion model that generates experimentally realizable crystal structures from a 2D graph in seconds. It jointly learns how a molecule bends (conformation), and how copies of that molecule pack in a periodic crystal. OXtal works with rigid molecules, flexible molecules, and even some co-crystals. What I find most interesting is how it learns. Based on AlphaFold3, we moved away from the complex symmetries of infinite periodic lattices. Instead, we focused on data augmentation and local chemical interactions. We teach the model how molecules prefer to sit relative to their neighbors. It turns out that if you get the local interactions right, the global packing emerges naturally at inference time. To be clear, CSP is not “solved”. OXtal is an efficient sampler, but we aren't yet ranking these structures. Think of this as a high-speed filter: we can now generate plausible candidates in seconds, allowing rigorous physics-based methods to be used only where they matter most. We’re excited to see how this accelerates discovery in pharma and materials science. Share your take with us! What an incredible scientific journey with Emily Jin, Andrei Nica, Michael Galkin, Jarrid Rector-Brooks, Kelvin Lee, PhD, Santiago Miret, Frances Arnold, Michael Bronstein, Joey Bose, Alexander Tong!  📄 Preprint: https://lnkd.in/g2v5-Qa7 🌐 Project page: https://oxtal.github.io/ 📰 Spotlight on FutureHouse: https://lnkd.in/gWMP6Fhx #ComputationalChemistry #AI4Science #MaterialScience #MachineLearning #Crystallography

Forbes Hungary listed me among the "25 Hungarians Behind the AI Revolution." It's an absolute honor to be mentioned alongside paradigm defining researchers like Ferenc Huszár and Christian Szegedy. I wouldn't be here without the help of my mentors and friends, and I am massively grateful to FutureHouse and Eric and Wendy Schmidt Center Broad Institute of MIT and Harvard for supporting me on this journey. See more at https://lnkd.in/e929FJPD