How AI Learned to Design Proteins from Scratch

For decades, protein engineers worked like architects with incomplete blueprints. They knew that a protein's three-dimensional shape determined what it could do—bind a drug, catalyse a reaction, recognise a virus. But predicting that shape from the amino acid sequence alone remained stubbornly difficult. Most proteins were understood only after they had been crystallised and imaged in the laboratory, a process that could take months or years.

Then, in the early 2020s, the pace changed.

From prediction to design

The breakthrough came from machine learning models trained on vast databases of known protein structures. These models learned statistical patterns in how amino acids fold into three dimensions. By 2020, AlphaFold—developed by DeepMind—could predict the structure of an unknown protein with remarkable accuracy, often matching experimental results. This was significant because it meant researchers could now model a protein's shape before making it.

But prediction and design are not the same thing. Prediction answers the question: given this sequence, what shape will it fold into? Design asks the harder question: given a desired shape or function, what sequence will produce it?

The leap came when researchers realised they could reverse the logic. If a neural network could learn the relationship between sequence and structure, it could be used to search for sequences that would fold into a target shape. This is not trivial—the space of possible protein sequences is astronomically large. But guided by learned patterns, the search became tractable.

What changed in practice

Before these models, protein design relied on rational engineering: researchers would study a natural protein, make educated guesses about which amino acids to swap, and test the variants experimentally. Progress was incremental. A successful design might improve a protein's stability or activity by a factor of two or three.

With AI-guided design, researchers could now generate dozens or hundreds of candidate sequences predicted to fold into a desired shape, test them in parallel, and iterate rapidly. More importantly, they could design proteins that had no natural counterpart—entirely novel molecular machines.

Examples emerged quickly. Researchers designed proteins that bound specific small molecules with high affinity. Others created enzymes that catalysed reactions not found in nature. Some designed proteins that self-assembled into geometric structures. These were not incremental improvements to existing proteins; they were new molecular tools, created computationally and validated experimentally.

Why this matters

The significance lies in speed and scope. Before, protein engineering was a bottleneck. If you wanted a protein with a particular property, you might spend years optimising it. Now, the computational phase—generating and ranking candidate sequences—takes hours or days. The experimental phase remains necessary—predictions are not perfect—but the ratio of computation to wet-lab work has shifted dramatically in favour of computation.

This has practical consequences. Therapeutic proteins can be designed faster. Industrial enzymes for manufacturing can be optimised for specific conditions. Biosensors can be engineered to detect novel targets. The field has moved from asking "can we design a protein to do this?" to asking "how quickly can we do it?"

The 2024 Nobel Prize

The Nobel Prize in Chemistry for 2024 recognised two contributions to this transformation. One was the development of AlphaFold and related structure prediction models. The other was the creation of generative models for protein design—systems that can propose novel sequences likely to fold into desired shapes.

The prize acknowledged that these were not incremental improvements but fundamental shifts in how the field works. They represent a change in the nature of the question being asked: from "how do proteins fold?" to "how do we build proteins that do what we want?"

This is not to say that prediction and design are now solved problems. Predictions remain imperfect, especially for large or multi-chain proteins. Designed proteins sometimes fail to behave as predicted when tested. The models work best when guided by experimental feedback. And many biological questions—how a protein interacts with a crowded cellular environment, how it evolves under selection—still require wet-lab work.

But the direction is clear. Computational models have become the first step, not the last. They have compressed the design cycle and expanded what is possible to attempt. The field has moved from studying nature's proteins to engineering new ones, and AI models provided the tools to do it at scale.

As protein design becomes routine, the quality of the underlying data—the structures used to train models, the sequences tested experimentally, the results reported clearly—becomes more important. Careful curation and analysis of biological data will remain essential to keeping pace with what the models can propose.