AI is designing proteins nature never made. The first one might save snakebite victims.

Snakebite still kills more than 100,000 people a year, mostly in low-income regions, and the standard treatment is older than the electric light. Antivenom is still made roughly the way it was in the 1890s: inject a horse with venom, harvest the plasma, hope the antibodies neutralize the poison. The approach works for some bites and fails against others, especially against the three-finger toxins that dominate cobra and mamba venom.

In January 2025, a team led by Sofía Vázquez Torres, David Baker at the University of Washington, and Timothy Jenkins at the Technical University of Denmark published a Nature paper showing that AI-designed miniproteins can neutralize those three-finger toxins in mice. Survival ran from 60 to 100 percent depending on the toxin, dose, and timing. Full protection came when the designed protein was pre-mixed with the toxin or given within fifteen minutes. Treatment delayed to thirty minutes dropped one binder to roughly 60 percent. The new molecules are heat-stable and can be brewed in bacteria, which matters more than the survival numbers for the parts of the world where the antivenom shortage is worst.

That result is the first concrete payoff. The more interesting shift is the one underneath it.

Evolution had roughly four billion years to build proteins and only ever sampled a sliver of what chemistry actually allows, because it can only tinker with what already exists, one mutation at a time. Generative protein design is not bound by that constraint. A model can walk into unexplored sequence space, propose a backbone with no natural ancestor, and check whether AlphaFold thinks it will fold. The standard open-source pipeline now runs in three named steps: RFdiffusion generates a backbone, ProteinMPNN picks the amino acids, and AlphaFold checks the fold. Each step is a separate neural network with a separate job, and the pipeline has been good enough for a few years that the bottleneck has moved.

As Wei Yang and colleagues wrote in an April 2026 Nature review, "the long-standing structural problems are close to solved, and the live question is no longer how to design a protein but what to design." Capability is no longer the binding constraint on generative biology. Judgment is.

That pivot changes what to watch. The next decade of protein design is a target-selection, validation, manufacturing, and governance problem, not an algorithm problem. The Baker lab's snakebite result is one example of a target worth chasing, and a deliberately easy one: the toxin structure was already mapped, the unmet medical need was obvious, and the manufacturing story (bacterial brewing, heat-stable molecules) lined up with the parts of the world where the problem is worst. Not every target is that clean.

MIT engineers reported in March 2026 that the same approach now extends to dynamic motion, not just static shape, which broadens what counts as a "designable function." A protein can be specified by the way it moves as well as the way it sits. Adjacent work in Nature Materials 2025 adds wins in functional design, from new enzymes to materials with programmable properties. Each new paper widens the menu.

That menu is now bigger than the field's ability to test what is on it. Validation still means wet-lab confirmation, animal testing, and a manufacturing and regulatory pathway that has not caught up to the design speed. "Designed" does not mean "ready." The snakebite miniproteins solve one access point, the breweable, heat-stable manufacturing, but the clinical timeline is a different question.

The governance question is the one the field has not yet answered. Microsoft Research has shown that generative protein design tools can produce sequences that slip past the nucleic-acid biosecurity screens that gate gene synthesis. The same pipelines that protect against natural pathogens can now be aimed at the synthesis chokepoint. The chokepoint has not yet been widened, and the governance is now the rate-limiter.

AI has been technically able to design proteins from scratch for years, and "AI helps make antivenom" buries the lead. The cleaner story is the shift Yang's review names: the structural problem is close to solved, and the open questions are what to ask for, how to test it, how to make it at scale, and how to keep the synthesis supply chain from becoming the easiest place to weaponize the same tools.