Excited to release BoltzGen which brings SOTA folding performance to binder design! The best part of this project has been collaborating with many leading biologists who tested BoltzGen at an unprecedented scale, showing success on many novel targets and pushing its limits! 🧵..
Oct 26, 2025 · 9:58 PM UTC
We go after targets that require generalization. E.g. we tested 15 nanobodies against each of 9 targets selected for their dissimilarity to any protein with an existing bound structure. For 6 of 9 targets we obtain nM binders. The same 67% success rate holds for miniproteins 🤗
For peptides, we e.g., succeed for designing linear and disulfide-bonded peptides against crucial metabolic pathway targets and disordered proteins. When testing 5 designs, we have one success, posing the first evidence of de-novo peptides binding disordered proteins.
We have more campaigns where the validation goes beyond binding: we test 6 BoltzGen designs against each of 3 diversely structured peptides. We obtain nM binders against two and uM against the third. For every target, at least on design neutralizes its antimicrobial activity.
BoltzGen’s success stems from its unification of design and structure prediction. A purely geometry-based representation of designed residues enables scalable training on both tasks. As a result, unlike any previous design model, BoltzGen matches the performance of SOTA folding
Due to a purely geometry-based encoding of designed residues this unified boils down to supervising with the same diffusion loss for structure prediction and design.
The second point contributing to performance is its generality: As models learn to emulate physics primarily through examples provided, we believe expanding the generality of the method further improves its design capabilities for specific classes as well. One contributor toward that is our diverse set of training tasks that exercises the model in all contexts.
With this we train a model with the standard AF3 / Boltz-2 scalable architecture that has proven state-of-the-art for folding. Injecting conditioning inputs allows us to control the designed binder in various ways
This results in a design specification language for various constraints – including covalent bonds, structure groups, binding sites, secondary structures and design masks – that steer the diffusion process towards specific design objectives during inference.
Everything is integrated in one easy to use end-to-end pipeline! Just type up your design specification and try it out!
🚀 Model & code: github.com/HannesStark/boltz…
🤗 Join our fast-growing Slack community: boltz.bio/join-slack
🧠 Blog post: boltz.bio/boltzgen
📄 Full manuscript: hannes-stark.com/assets/bolt…
And join us for live presentations, demos, and discussions:
1. MIT (Cambridge) – Thursday, October 30th luma.com/7474iho2
2. London – Thursday, November 6th luma.com/l2zgvfwt
Huge thanks to @felix_faltings, MinGuy Choi, Yuxin Xie, Eunsu Hur, @timodonnell, @AntonBushuiev, Talip Uçar, @pass_saro, Weian Mao, @meteos_97, @roman_bushuiev, @tomas_pluskal, Josef Sivic, @karsten_kreis, @ArashVahdat, @shamayeetaray, Jonathan T. Goldstein, @bioSavinov, Jacob A. Hambalek, Anshika Gupta, Diego A. Taquiri-Diaz, @YZhang1997, A. Katherine Hatstat, @AngelikaArada, Nam Hyeong Kim, Ethel Tackie-Yarboi, Dylan Boselli, @LeeSchnaider, @chang_c_liu, @GeneWeiLiLab, @dhnisz, @DMSabatini, William F. DeGrado, @jeremyWohlwend, @GabriCorso, @BarzilayRegina, Tommi Jaakkola!
Also, massive thanks to @adaptyvbio for their fast turnaround in the experimental validation!
