A hundred quadrillion dollars. That's roughly $10^15 - about a thousand times the entire US GDP. It's also what it would cost to individually synthesize the DNA library that a team from JURA Bio and Harvard just produced for around a thousand bucks.
The trick? They stopped fighting chemistry and started collaborating with it.
When Bugs Become Features
Here's the dirty secret of DNA synthesis: it's inherently sloppy. The standard phosphoramidite method, unchanged since the 1980s, builds DNA one nucleotide at a time. At each step, the "wrong" base occasionally slips in. For decades, chemists have treated this randomness as the enemy - an error rate to minimize, a yield curve to optimize, a quality control nightmare.
Eli Weinstein, George Church, and colleagues at JURA Bio looked at that same randomness and thought: what if the noise IS the signal?
Their method, called variational synthesis, deliberately feeds mixtures of nucleotides at each synthesis step. Instead of trying to build one perfect sequence, every single molecule on the synthesis chip becomes a unique sample from a generative model. The DNA synthesis machine isn't just a printer anymore - it's a physical random number generator that happens to output designed proteins.
It's like the difference between hand-writing a million letters versus teaching a printing press to improvise. Same ink, same paper, wildly different throughput (Weinstein et al., 2026).
Training on 300 Million Antibodies (As One Does)
The team trained their manufacturing-aware generative model on 300 million observed human antibodies. The model learned the patterns and statistical structure of real antibody sequences - which amino acids tend to appear together, which regions can tolerate variation, which positions are essentially locked in by evolution.
But here's where it gets clever. Unlike conventional protein language models that dream up sequences in silicon and then hand them off to expensive synthesis pipelines, variational synthesis models are designed with parameters that map directly to the DNA synthesis process. The model doesn't just know what good antibodies look like - it knows how to physically build them using controlled nucleotide mixtures.
The result: approximately 10^17 unique antibody designs synthesized in a single experiment. That's a hundred million billion sequences. For context, the estimated number of grains of sand on Earth is around 7.5 x 10^18. They're within spitting distance of one sand grain per antibody, and all it cost was about a thousand dollars and some very precisely calibrated reagent bottles.
The Trillion-Fold Cost Collapse
The economics here are genuinely staggering. Traditional targeted gene synthesis runs about $0.07 per base pair, and producing individual sequences gets expensive fast at scale. Array-based synthesis is cheaper per base but still nowhere close to petascale. Previous methods would price an equivalent library at roughly $10^15 - a number so absurd it barely registers as real money.
Variational synthesis collapses that cost by recognizing that you don't need to synthesize each sequence individually. Since every molecule on the chip is independently sampling from the learned distribution, parallelism comes for free. The chemistry does the sampling. The algorithm just tells the chemistry what probability distribution to sample from (Weinstein et al., AISTATS 2022).
The approach also doesn't require new hardware. It runs on existing oligosynthesis platforms - you just change the software controlling the nucleotide mixture ratios. Same machines, same chemistry, radically different output.
Not Just Antibodies
To prove this wasn't a one-hit wonder, the team ran variational synthesis on two additional protein families: Taq polymerase (the workhorse enzyme of PCR, which every biology undergrad has cursed at least once) and the HLA-presented peptidome (the collection of peptide fragments displayed on cell surfaces for immune recognition). Both produced libraries of approximately 10^16 designs with quality comparable to state-of-the-art protein language models like ESM3 (Hayes et al., Science, 2025). The designed sequences were verified by DNA sequencing.
This breadth matters. Antibodies, polymerases, and immune peptides are structurally and functionally diverse families. If variational synthesis works across all three, it likely generalizes broadly - to enzymes, receptors, regulatory elements, and beyond.
The Bigger Picture: Closing the Gene Writing Gap
Our ability to read DNA has raced ahead of our ability to write it. Sequencing costs have cratered from $3 billion for the first human genome to roughly $600 today. Synthesis? Still expensive, still slow, still limited to about 200 nucleotides per run before yield falls off a cliff. This asymmetry - often called the "gene writing gap" - has been a persistent bottleneck in synthetic biology (Hughes & Ellington, Nature Reviews Chemistry, 2023).
Variational synthesis doesn't close this gap in the traditional sense. It doesn't make individual gene synthesis cheaper. What it does is something arguably more useful: it makes libraries of designed sequences absurdly cheap to produce, as long as you're willing to let a generative model choose what goes in them. For drug discovery, directed evolution, and therapeutic antibody screening, that's exactly the right tradeoff.
If you're the kind of person who likes to map out how different research threads connect - from protein language models to diffusion-based design to physical synthesis - this paper sits at a genuinely novel intersection. It's not just better AI for biology or better chemistry for synthesis. It's both, fused at the architectural level.
What Could Go Wrong?
The authors themselves acknowledge dual-use concerns. A technology that can produce 10^17 designed biological sequences for pocket change raises legitimate biosecurity questions. The paper's predecessor (Weinstein et al., AISTATS 2022) explicitly flagged synthetic biology as "a dual use technology, with a range of potentially dangerous applications." The field is grappling with these risks in real time, and the ease of this approach will only accelerate those conversations.
There's also the question of functional validation. Synthesizing 10^17 sequences is spectacular, but which of those sequences actually fold correctly, bind their targets, or do anything useful? The paper verified quality by sequencing, but the wet-lab validation of function at this scale remains the next frontier.
The Bottom Line
Variational synthesis is one of those ideas that seems obvious in retrospect - of course you should use the randomness of chemistry as your random number generator - but required deep expertise in both machine learning and synthetic biology to actually pull off. By making generative models manufacturing-aware, the JURA Bio team has turned DNA synthesis from an expensive bottleneck into a massively parallel sampling engine.
The age of computationally designing a few proteins and painfully synthesizing them one at a time may be ending. The age of designing a quadrillion and letting chemistry do the heavy lifting has apparently arrived - for about the cost of a nice espresso machine.
References
-
Weinstein, E.N., Gollub, M.G., Slabodkin, A., et al. (2026). Manufacturing-aware generative models enable petascale synthesis of designed DNA. Nature Biotechnology. DOI: 10.1038/s41587-026-03020-8. PMID: 41844990.
-
Weinstein, E.N., Amin, A.N., Grathwohl, W., et al. (2022). Optimal Design of Stochastic DNA Synthesis Protocols based on Generative Sequence Models. AISTATS 2022, PMLR 151:7450-7482.
-
Hayes, T., et al. (2025). Simulating 500 million years of evolution with a language model. Science. DOI: 10.1126/science.ads0018.
-
Hughes, R.A. & Ellington, A.D. (2023). DNA synthesis technologies to close the gene writing gap. Nature Reviews Chemistry. DOI: 10.1038/s41570-022-00456-9. PMID: 36714378.
-
Watson, J.L., Juergens, D., Bennett, N.R., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature. DOI: 10.1038/s41586-023-06415-8. PMID: 37433327.
-
Hie, B.L., et al. (2024). Unsupervised evolution of protein and antibody complexes with a structure-informed language model. Science. DOI: 10.1126/science.adk8946. PMID: 38963838.
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.