This piece from Works in Progress doesn't just report on a scientific breakthrough; it reframes the very nature of biological discovery. The most startling claim isn't that AI can predict protein structures, but that it can now hallucinate entirely new ones from scratch—creating functional molecules that evolution never conceived. For a busy reader, this represents a fundamental shift from observing nature to programming it, with implications that stretch far beyond the lab bench.
Beyond Evolution's Blueprint
The editors of Works in Progress anchor their argument in a provocative premise: nature is not the ultimate optimizer. While evolution has solved many problems, it has left vast gaps, particularly regarding human-made pollutants and novel pathogens. The piece notes that "plastic famously is not biodegradable," pointing out that "nature's not doing, in the sense of bio, ain't doing much degrading." This framing is crucial because it justifies the leap from tweaking existing biology to inventing new biology. The argument lands effectively by contrasting the slow, trial-and-error pace of natural selection with the rapid, targeted design of artificial intelligence.
The coverage highlights specific tools like RFDiffusion and AlphaFold, describing how researchers are using them to "hallucinate entirely novel proteins: designing them from scratch to solve problems evolution hasn't tackled." This is not merely a technical update; it is a philosophical pivot. The piece suggests that we are moving from a world where we discover what works to one where we define what works and then build it. Jacob Trefethen, a contributor to the podcast transcript featured in the article, illustrates this by discussing the design of a protein to block hepatitis B. He explains the goal is to create a binder that interrupts the virus's life cycle, stating, "I want to bind a part of hepatitis B to interrupt its life cycle, so it stops damn replicating in my liver cells." The raw, direct language here underscores the urgency and practicality of the technology.
"If you could create a protein that operated as a therapeutic, as a drug, and you just created it out of thin air and you never saw anything like it before, then could be really useful."
Critics might note that the term "hallucinate"—borrowed from generative AI image models—could be misleading when applied to molecular biology, where precision is paramount and errors can be toxic. However, the piece clarifies that these are not random guesses but structured outputs validated through rigorous loops. The editors argue that these models are part of a "design, create and validation" loop, ensuring that the hallucinated structures are not just mathematically probable but biologically functional.
From Digital Models to Physical Reality
The commentary shifts from the theoretical to the tangible, exploring how these digital designs translate into real-world applications. The piece argues that the potential extends well beyond medicine into agriculture and materials science. Saloni Dattani, another contributor, raises the possibility of creating proteins that can "purify water" or "get rid of dirt," while Trefethen envisions improving photosynthesis to be more efficient than the current 1% biomass conversion rate. The editors use these examples to demonstrate the breadth of the opportunity, noting that "nature's so good when there's a problem that evolution's really taken a swing at," but falters when faced with entirely new challenges like plastic waste or synthetic pathogens.
The article details the mechanics of this process, describing how diffusion models work similarly to image generators like Midjourney. "Instead of hallucinating cat pictures, you started hallucinating a protein structure," the piece explains. This analogy makes the complex technology accessible, stripping away the jargon to reveal the core mechanism: generating novel structures based on learned patterns. The editors emphasize that this is a collaborative future, predicting that "structural biologists who know how to use these computational tools" will pair with domain experts to solve specific problems. This reframing of scientific labor—from solitary discovery to interdisciplinary engineering—is a vital insight for readers tracking the future of innovation.
"The starting gun is basically 2022. What we're gonna see... is a lot of structural biologists who know how to use these computational tools, matching up with experts in particular fields... and in combination, I think those teams of people are gonna do really incredible things."
A counterargument worth considering is the timeline. While the technology is advancing rapidly, the path from a computer-generated model to a safe, approved drug or material is long and fraught with regulatory hurdles. The piece acknowledges that "AI still can't predict everything about proteins," hinting at the remaining gaps between simulation and reality. However, the editors maintain that the trajectory is clear: the ability to design proteins de novo is no longer science fiction but an emerging engineering discipline.
Bottom Line
The strongest part of this coverage is its refusal to treat AI protein design as a mere incremental improvement; instead, it correctly identifies it as a paradigm shift that allows humanity to bypass evolutionary constraints. The piece's biggest vulnerability lies in its optimism regarding the speed of adoption, as the gap between digital design and clinical or industrial deployment remains significant. Readers should watch for the first major commercial success of a de novo designed protein, which will serve as the definitive proof that this "hallucination" is the future of medicine and materials.