What is the single most convincing piece of clinical evidence for an AI-designed protein?
The most compelling evidence comes from a 2023 Nature study where researchers used a diffusion model called RFdiffusion to design hundreds of new proteins from scratch. They then experimentally tested these designs and found that the cryo-electron microscopy (cryo-EM) structure of one designed protein binder—built to grab onto influenza hemagglutinin—was nearly identical to the original computer model [3]. This is a big deal because cryo-EM gives you a 3D picture of the actual protein at near-atomic detail, so matching the design model means the AI's prediction was essentially correct. The study also validated many other designs: symmetric assemblies, metal-binding proteins, and additional binders, all of which worked as intended [3]. This isn't a single lucky hit—it's a systematic demonstration that the method reliably produces functional proteins.
Do other studies back up this finding, or is it just one flashy result?
The evidence is broader than a single study. A 2024 perspective in Cell reviewed the entire field of de novo protein design and concluded that new protein folds and higher-order assemblies can now be designed with 'considerable experimental success rates' [1]. That means AI methods are not just producing occasional hits—they are consistently making proteins that fold and function as intended. Another 2023 article in Nature Biotechnology described how protein engineers are using machine learning tools like AlphaFold2 to pursue 'more sophisticated de novo protein designs,' confirming that the approach is being adopted widely and producing results [4]. Together, these sources show that the RFdiffusion result is part of a larger, growing body of evidence that AI-designed proteins work in the lab.
What are the caveats—are there any risks or gaps in the evidence?
The evidence is strong for structural and binding accuracy, but it does not yet cover long-term clinical outcomes like safety and efficacy in humans. None of the papers here report results from human clinical trials; the evidence is all from lab experiments and animal models [1][3][4]. Additionally, a 2026 paper in Frontiers in Microbiology raises a biosecurity concern: AI-generated proteins can be functionally equivalent to known toxins while sharing very little sequence similarity, which means current screening methods might miss them [2]. This doesn't undermine the clinical promise, but it highlights that the same power that makes these proteins useful for therapeutics also creates new risks that need to be managed. The 2025 Open Research Europe article also notes that while AI has revolutionized protein structure prediction, challenges remain in understanding protein dynamics and aggregation, which could affect how these proteins behave in the body over time [5].
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2026, 3 from 2024 or later, 4 in Q1 journals, collectively cited 1,863 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 50 papers retrieved from a database of over 500 million.
Sources used in this answer
De novo protein design—From new structures to programmable functions
This 2024 perspective in Cell reviews the field and states that new protein folds and higher-order assemblies can now be designed with 'considerable experimental success rates,' and that difficult problems like tunable conformational control are coming into reach.
Protein design, generative AI and biological security.
This 2026 review in Frontiers in Microbiology warns that AI-generated proteins can be functionally equivalent to known toxins while sharing little sequence similarity, potentially evading current homology-based biosecurity screening.
De novo design of protein structure and function with RFdiffusion
This 2023 Nature study shows that RFdiffusion designed hundreds of new proteins, and the cryo-EM structure of a designed binder for influenza hemagglutinin matched the design model nearly identically, confirming the method's accuracy.
AI-enhanced protein design makes proteins that have never existed
This 2023 article in Nature Biotechnology reports that protein engineers are using rapidly evolving machine learning tools and AlphaFold2 to pursue more sophisticated de novo protein designs.
When artificial intelligence meets protein research
This 2025 article in Open Research Europe notes that AI tools like AlphaFold have transformed protein science but that challenges remain in understanding protein dynamics and amyloid aggregation.