Why do most AI-designed proteins fail in the lab?
The biggest evidence gap is that AI models are much better at dreaming up proteins than at making ones that actually work. State-of-the-art design protocols achieve experimental success rates nearing 20% [1]. That means for every 100 proteins a top AI model designs, only about 20 will fold correctly and function as intended when synthesized and tested in a lab. The other 80 fail, wasting time and resources. This low success rate shows that the computer models are missing something fundamental about how proteins behave in the real world.
A key reason for this gap is that the metrics used to evaluate designs on a computer don't reliably predict real-world performance. The field has a clear challenge in determining the best in silico (computer-based) metrics to prioritize designs for experimental testing [1]. Without trustworthy computer predictions, researchers have to test many more designs than they should, slowing the entire pipeline from design to application.
What kinds of proteins can AI not design yet?
AI models are particularly bad at designing proteins that need to move or be controlled. Current generative AI struggles to design proteins that can undergo large conformational changes (dramatic shape-shifting) or be regulated by post-translational modifications (chemical switches that turn proteins on and off) [1]. These abilities are essential for many natural proteins that act as sensors, motors, or switches in cells. Until AI can handle this complexity, it will be limited to designing static proteins, which represent only a fraction of what biology uses.
The reverse problem—taking a desired protein shape and predicting the amino acid sequence that will fold into it—has proven trickier than predicting structure from sequence [5]. While tools like AlphaFold2 have revolutionized structure prediction, the design problem is fundamentally harder because many different sequences can theoretically produce the same shape, and the model must pick one that will actually work in a living cell [3][5].
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2022 to 2024, 2 from 2024 or later, 4 in Q1 journals, collectively cited 155 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 40 papers retrieved from a database of over 500 million.
Sources used in this answer
Generative artificial intelligence for de novo protein design
This 2024 review finds that state-of-the-art AI design protocols achieve experimental success rates nearing 20%, and identifies key gaps: lack of reliable in silico metrics to prioritize designs, and inability to design proteins that undergo large conformational changes or are regulated by post-translational modifications.
Protein design meets biosecurity
This 2024 perspective by Baker and Church warns that the rapid increase in AI protein design accuracy creates biosecurity risks, and argues that all synthetic gene sequence and synthesis data should be stored in secure repositories queried only in emergencies.
AI-enhanced protein design makes proteins that have never existed
This 2023 news article reports that protein engineers are using machine learning tools and AlphaFold2 to pursue more sophisticated de novo protein designs, but notes the design problem (predicting sequence from shape) is harder than structure prediction.
AI models for protein design are driving antibody engineering
This 2023 review ties advances in deep learning-based protein structure prediction and design to antibody engineering, noting that structure-based generative models are emerging but do not report specific success rates or evidence gaps.
Dreaming up new protein designs with AI
This 2022 article reports that a new machine learning algorithm from the University of Washington can design protein molecules faster and more accurately than before, but notes the reverse problem (sequence from shape) has proven trickier than structure prediction.