Inside one protein, five different strings of building blocks can fold into the exact same shape. Nature picked one of them — but that doesn't mean it's the best one. A team at MIT has now built an AI that can dream up sequences evolution never chose, and that could one day lead to new medicines that grab onto disease-causing molecules.
The building blocks here are called amino acids, and the order they line up in decides how a protein folds. And how a protein folds decides what it does. For years, scientists designing new proteins worked backward: first they drew the structure they wanted, then they used AI to find amino acid sequences that could take that shape. The usual test of success was simple — could the AI reproduce the sequence that evolution happened to pick?
That's the wrong question, according to Amy E. Keating, the head of MIT's Department of Biology and senior author of the new study published in the journal PNAS. "For years, the field has measured success by asking whether a model can reproduce the protein sequence that evolution happened to select — our work shows that this isn't the best metric for protein design," she says.
The problem is that nature's choice isn't the only choice. Many sequences fold the same way, and even a single sequence can shift shape when it's flexible or triggered. Grad student Foster Birnbaum, the paper's lead author, realized that when designing a completely novel protein, there's no natural sequence to copy at all. What matters is whether the AI's ideas actually fold correctly and stay stable.
So the team built PottsMPNN, a new machine-learning framework that folds physical rules into its thinking. It captures the invisible "sequence-energy landscape" — the relationship between each amino acid and how stable the protein stays. Two tricks make it special. It tracks how every pair of positions interacts with the other 20 possible building blocks, something older models miss. And Birnbaum added a sprinkle of "noise" during training — small variations that stop the AI from merely mimicking natural proteins and push it to explore wilder, more varied designs.
The results are striking. As the model leans less on natural sequences, its predictions of stability actually improve, even for proteins that look like nothing in the wild. That's exactly what Keating means: the old metric measured copying, not creating.
The stakes reach far beyond the lab. Once scientists can design a stable protein to order, they gain a powerful tool to build molecules that bind to disease-causing targets inside our cells. PottsMPNN is one more step toward a future where "design any protein we want" stops being a slogan and becomes a routine lab procedure.