AI is starting to “design” proteins, not just predict them

For a long time, protein AI meant one headline: predicting what shape a protein will fold into. That story changed in 2021, when AlphaFold2 showed that deep learning could predict many protein structures with near-experimental accuracy in difficult cases, as reported in a Nature paper by Jumper and colleagues. But in the last two years, the focus has started to shift again. The newest wave is not only asking “What does this protein look like?” Instead, it is asking, “Can we design a protein that has the shape and function we want?”

Protein prediction starts with a sequence and tries to infer the 3D structure. Protein design flips that direction. Instead of explaining what nature already built, design tools try to generate new candidates—new structures and new sequences—that could work in the real world. It changes the workflow from interpreting biological molecules that already exist to proposing new molecular candidates that can be synthesized and tested.

One of the clearest examples of this shift is RFdiffusion. In a 2023 paper published in Nature, Watson and colleagues introduced it as a diffusion-based approach that can generate protein backbones (the protein’s 3D structural framework) for tasks such as monomer design, binder design, symmetric oligomers, and even enzyme active-site scaffolding. The key point is that this is not only a simulation story. In a July 11, 2023 post on the Baker Lab’s official website, the team explains that they experimentally tested hundreds of AI-generated proteins, connecting generative design to real wet-lab validation. That is why “AI designs proteins” is becoming less of a metaphor and more of a workflow.

Even when a plausible backbone is available, design requires a sequence that will fold into that structure reliably. In a 2022 Science paper, Dauparas and colleagues introduced ProteinMPNN, a deep learning method for protein sequence design, reporting strong performance in both computational and experimental tests. In practical terms, sequence-design methods such as ProteinMPNN help convert structural concepts into synthesizable candidates by proposing sequences that are more likely to adopt the target fold.

Protein AI is no longer only about predicting what already exists. AlphaFold2 helped make structure prediction dramatically more powerful, as described in the 2021 Nature paper by Jumper and colleagues. Now, models such as RFdiffusion and ProteinMPNN are pushing the field toward design, where AI generates new candidates and scientists validate them experimentally. If this trajectory continues, AI could potentially expand not just the speed of biology research, but the range of proteins humans can create on purpose.