JURA Bio, Inc.

Why not sample and physically test 200M designs against 100 targets with @jura_bio instead? That scale is beginning to achieve generalization. More at jurabio.com/blog/scalingin…

We have made the world's largest screening platform for AI-designed proteins and peptides. To celebrate it, & inspired by @ChordomaFDN great tbxtchallenge.org, we're soliciting hard targets. At scale, undruggable doesn't take any longer: ✉️ challenge@jurabio.com

At @jura_bio we take a lot of pride in solving the next big infrastructure problems, at scale -- and often before they're fully recognized as the next big problem.

@peterottsjo Peter, same day, 200M AI designed proteins were tested in the lab -- at once! Check it out: jurabio.com/blog/scalingin… and the Nature Biotech paper released the same day: rdcu.be/e8zUb

Build the lab into your generative process

Massive lab-to-AI loops. Wondering how exactly it is done? This "manufacturing-aware" model generates novel protein designs that can also be synthesized at scale. Using this they synthesised billions of antibodies, feeding that right back to train next-gen binding models.

sometimes there's a paper that is going to change the way you plan your research and what questions you get to ask and hope to answer. Still a little shaken with this before/after.

We recently used a model like the one described here to design and manufacture 209M candidate antibodies, which we screened against 100 targets. You can read more details: jurabio.com/blog/scalingin…

Manufacturing-aware generative models enable petascale synthesis of designed DNA go.nature.com/3NxXt1I

Grateful for the comparison, and I do suppose that if we were included in the benchmark and you squinted, it might happen to fall our way this time around, but I think this interpretation misses a big difference in goals. We design our ultra-large data sets to enable model training. This has some unintuitive consequences. For example, often the most informative data points are the ones thought to be very unlikely to bind or interact at all. When they end up working, they're maximally informative. Benchmarking models works differently: it wants to take the best guesses and confirm that they work. Very nice work from the @nvidia et al. team, and if you're interested in learning more about our (@jura_bio) AI systems that design and capture the data they want to learn from, you can read our newest technical post: jurabio.com/blog/screening… Or to learn more about the underlying technology, that allows us to scale to petascale in our DNA designs, our paper on scaling variational synthesis is out today in @NatureBiotech rdcu.be/e8zUb

Our paper on variational synthesis is out now in Nature Biotechnology. Manufacturing-aware generative models — AI architectures that know how to physically build their own designs — enabling synthesis of DNA encoding ~10^16 AI-designed proteins at a cost that would be roughly a quadrillion dollars using conventional methods.

Amazing work🦾🦾

209M AI designed proteins tested in the wetlab against 100 hard-to-drug targets, simultaneously. Highly informative readouts in human cell lines. And a model of antibody-intracellular target interactions trained on that data, that generalizes.

@CalvinMccarter Everyone come over to @jura_bio and finally start building something different 🐎

We built Sovereign AI to go from generative design to real, manufacturable therapeutics as fast as the problem demands: and in this case, even faster than the target can escape. Today we're announcing a partnership with @KoaBio to do rapid, function-based discovery of biologics against emerging multi-drug-resistant bacteria in-lab and in-theater. Collaborators at WRAIR and AFRIMS surveil threats, @jura_bio rapidly generates threat-neutralizing candidates, and Koa Bio scales and deploys what works. AMR is projected to kill 10 million people a year by 2050 and the bacteria aren't slowing down. It's for problems like these -- where speed is critical and success mandatory -- that I'm proudest that JURA is able to offer a solution unlike any other.

I get to announce a partnership today -- @jura_bio and Annogen will collaborate to build and train regulatory element design models using #VariationalSynthesis driven sovereign data 🚀

RT @lizbwood: Excited to share that our paper on LeaVS (Learning from Variational Synthesis) has been accepted at @AISTATS #AISTATS2026! L…

Try them too! Just for the love of god don’t build your own lab

Nah, come over to @jura_bio and have variational synthesis go brrr. 100M+ of your AI designs tested for you in vitro. One example: jurabio.com/mesa

New causal design blogpost up at @jura_bio: In 1929, a Johns Hopkins pathologist noticed something strange in his autopsy data: patients with active tuberculosis had lower rates of cancer, and he couldn't explain why. Three decades later, researchers showed that mice infected with Bacillus Calmette-Guérin (BCG) had increased tumor resistance, and in 1976, a urologist named Alvaro Morales tried BCG on nine bladder cancer patients and saw a twelve-fold reduction in recurrence. When he sought funding, he was told his concept was "a throwback to the stone age of tumor immunology," but he persisted, and by 1990 BCG was FDA-approved. Fifty years later, it remains the gold standard for first-line intervention for bladder cancer, while its mechanism is still debated. This is an example of what you might call causal medicine: interventions validated by observing their effect on patients, with mechanism understood later, or even never. They suggest something important about how we should be building AI for drug design. Most AI in this space inherited a target-based structure oriented around predicting structure, predicting binding, and optimizing affinity. This happened because structure and binding data was clean, relatively abundant, and well-labeled, while clinical outcomes were messy, confounded, and sparse. The PDB was carefully built over decades, and the field naturally followed the data toward what was measurable. For many (but not all) targets, binding is relatively easy to satisfy since lots of sequences will bind a given target, while clinical efficacy is hard because the sequences that actually help patients are rare and we don't fully understand the constraints that make them rare. Optimizing for binding and then trying to put into place the screens that will ensure that efficacy follows is likely working the problem backwards. The human population is already running massive parallel experiments on what works. Trillions of immune receptors encounter disease every day, with patient immune systems generating candidates, testing them against emergent diseases, and selecting for effectiveness. The results are recorded in clinical systems and observable through repertoire sequencing. The challenge is confounding, but the biology of immune receptor generation provides a way out: somatic recombination is essentially randomized at the sequence level, which means we can estimate causal effects from observational data if we're careful. Instead of starting with targets and hoping to reach clinical efficacy, you start with clinical outcomes and leverage natural experiments to train models capable of causal generation, training on patient data directly and using binding, stability, and manufacturability as filters at the end rather than constraints at the beginning. I wrote up the case for causal design & generation: jurabio.com/blog/causaldes….